Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves.

Despite the central role that antibodies play in the adaptive immune system and in biotechnology, much remains unknown about the quantitative relationship between an antibody's amino acid sequence and its antigen binding affinity. Here we describe a new experimental approach, called Tite-Seq, that is capable of measuring binding titration curves and corresponding affinities for thousands of variant antibodies in parallel. The measurement of titration curves eliminates the confounding effects of antibody expression and stability that arise in standard deep mutational scanning assays. We demonstrate Tite-Seq on the CDR1H and CDR3H regions of a well-studied scFv antibody. Our data shed light on the structural basis for antigen binding affinity and suggests a role for secondary CDR loops in establishing antibody stability. Tite-Seq fills a large gap in the ability to measure critical aspects of the adaptive immune system, and can be readily used for studying sequence-affinity landscapes in other protein systems.

Introduction

During an infection, the immune system must recognize and neutralize invading pathogens. B-cells contribute to immune defense by producing antibodies, proteins that bind specifically to foreign antigens. The astonishing capability of antibodies to recognize virtually any foreign molecule has been repurposed by scientists in a wide variety of experimental techniques (immunofluorescence, western blots, ELISA, ChIP-Seq, etc.). Antibody-based therapeutic drugs have also been developed for treating many different diseases, including cancer (Chan and Carter, 2010).

Much is known about the qualitative mechanisms of antibody generation and function (Murphy et al., 2008). The antigenic specificity of antibodies in humans, mice, and most jawed vertebrates is primarily governed by six complementarity determining regions (CDRs), each roughly 10 amino acids (aa) long. Three CDRs (denoted CDR1H, CDR2H, and CDR3H) are located on the antibody heavy chain, and three are on the light chain. During B-cell differentiation, these six sequences are randomized through V(D)J recombination, then selected for functionality as well as against the ability to recognize host antigens. Upon participation in an immune response, CDR regions can further undergo somatic hypermutation and selection, yielding higher-affinity antibodies for specific antigens. Among the CDRs, CDR3H is the most highly variable and typically contributes the most to antigen specificity; less clear are the functional roles of the other CDRs, which often do not interact with the target antigen directly.

Many high-throughput techniques, including phage display (Smith, 1985; Vaughan et al., 1996; Schirrmann et al., 2011), ribosome display (Fujino et al., 2012), yeast display (Boder and Wittrup, 1997; Gai and Wittrup, 2007), and mammalian cell display (Forsyth et al., 2013), have been developed for optimizing antibodies ex vivo. Advances in DNA sequencing technology have also made it possible to effectively monitor both antibody and T-cell receptor diversity within immune repertoires, e.g. in healthy individuals (Boyd et al., 2009; Weinstein et al., 2009; Robins et al., 2009, 2010; Mora et al., 2010; Venturi et al., 2011; Murugan et al., 2012; Zvyagin et al., 2014; Elhanati et al., 2014; Qi et al., 2014; Thomas et al., 2014; Elhanati et al., 2015), in specific tissues (Madi et al., 2014), in individuals with diseases (Parameswaran et al., 2013) or following vaccination (Jiang et al., 2013; Vollmers et al., 2013; Laserson et al., 2014; Galson et al., 2014; Wang et al., 2015). Yet many questions remain about basic aspects of the quantitative relationship between antibody sequence and antigen binding affinity. How many different antibodies will bind a given antigen with specified affinity? How large of a role do epistatic interactions between amino acid positions within the CDRs have on antigen binding affinity? How is this sequence-affinity landscape navigated by the V(D)J recombination process, or by somatic hypermutation? Answering these and related questions is likely to prove critical for developing a systems-level understanding of the adaptive immune system, as well as for using antibody repertoire sequencing to diagnose and monitor disease.

Recently developed ‘deep mutational scanning’ (DMS) assays (Fowler and Fields, 2014) provide one potential method for measuring binding affinities with high enough throughput to effectively explore antibody sequence-affinity landscapes. In DMS experiments, one begins with a library of variants of a specific protein. Proteins that have high levels of a particular activity of interest are then enriched via one or more rounds of selection, which can be carried out in a variety of ways. The set of enriched sequences is then compared to the initial library, and protein sequences (or mutations within these sequences) are scored according to how much this enrichment procedure increases their prevalence.

Multiple DMS assays have been described for investigating protein-ligand binding affinity. But no DMS assay has yet been shown to provide absolute quantitative binding affinity measurements, i.e., dissociation constants in molar units. For example, one of the first DMS experiments (Fowler et al., 2010) used phage display technology to measure how mutations in a WW domain affect the affinity of this domain for its peptide ligand. These data were sufficient to compute enrichment ratios and corresponding sequence logos, but they did not yield quantitative affinities. Analogous experiments have since been performed on antibodies using yeast display (Reich et al., 2015; Kowalsky et al., 2015) and mammalian cell display (Forsyth et al., 2013). Yeast-display-based DMS assays have also proven particularly useful for mapping protein epitopes that are targeted by specific antibodies of interest (Kowalsky et al., 2015; Doolan and Colby, 2015; Van Blarcom et al., 2015). Still, none of these approaches provides quantitative affinity values. SORTCERY (Reich et al., 2015, ), a DMS assay that combines yeast display and quantitative modeling, has been shown to provide approximate rank-order values for the affinity of a specific protein for short unstructured peptides of varying sequence. Determining quantitative affinities from SORTCERY data, however, requires separate low-throughput calibration measurements (Reich et al., 2014). Moreover, it is unclear how well SORTCERY, if applied to a library of folded proteins rather than unstructured peptides, can distinguish sequence-dependence effects on affinity from sequence-dependent effects on protein expression and stability. Other recent work has described a DMS assay, again based on yeast display, for measuring fold-changes in affinity relative to a reference protein (Kowalsky and Whitehead, 2016). This method, however, does not provide absolute values for dissociation constants, is vulnerable to the confounding effects of sequence-dependent expression and protein stability, and was observed to have only a 10-fold dynamic range.

To enable massively parallel measurements of absolute binding affinities for antibodies and other structured proteins, we have developed an assay called ‘Tite-Seq.’ Tite-Seq, like SORTCERY, builds on the capabilities of Sort-Seq, an experimental strategy that was first developed for studying transcriptional regulatory sequences in bacteria (Kinney et al., 2010). Sort-Seq combines fluorescence-activated cell sorting (FACS) with high-throughput sequencing to provide massively parallel measurements of cellular fluorescence. In the Tite-Seq assay, Sort-Seq is applied to antibodies displayed on the surface of yeast cells and incubated with antigen at a wide range of concentrations. From the resulting sequence data, thousands of antibody-antigen binding titration curves and their corresponding absolute dissociation constants (here denoted ) can be inferred. By assaying full binding curves, Tite-Seq is able to measure affinities over many orders of magnitude (We note that Kowalsky et al. (2015) have described yeast display DMS experiments performed at multiple concentrations. These data, however, were not used to reconstruct titration curves or infer quantitative values). Moreover, the resulting affinity values provided by Tite-Seq are not confounded by the (rather substantial) effect that sequence variation can have on either (a) the amount of protein expressed on the surface of cells or (b) the specific activity of displayed proteins (i.e., the fraction of protein molecules that are functional).

We demonstrated Tite-Seq on a protein library derived from a well-studied single-chain variable fragment (scFv) antibody specific to the small molecule fluorescein (Boder and Wittrup, 1997; Boder et al., 2000). Mutations were restricted to CDR1H and CDR3H regions, which are known to play an important role in the antigen recognition of this scFv (Boder et al., 2000; Midelfort et al., 2004). The resulting affinity measurements were validated with binding curves for a handful of clones measured using standard low-throughput flow cytometry. Our Tite-Seq measurements reveal both expected and unexpected differences between the effects of mutations in CDR1H and CDR3H. These data also shed light on structural aspects of antigen recognition that are independent of effects on antibody stability.

Results

Overview of Tite-Seq

Our general strategy is illustrated in Figure 1. First, a library of variant antibodies is displayed on the surface of yeast cells (Figure 1A). The composition of this library is such that each cell displays a single antibody variant, and each variant is expressed on the surface of multiple cells. Cells are then incubated with the antigen of interest, bound antigen is fluorescently labeled, and fluorescence-activated cell sorting (FACS) is used to sort cells one-by-one into multiple ‘bins’ based on this fluorescent readout (Figure 1B). Deep sequencing is then used to survey the antibody variants present in each bin. Because each variant antibody is sorted multiple times, it will be associated with a histogram of counts spread across one or more bins (Figure 1C). The spread in each histogram is due to cell-to-cell variability in antibody expression, and to the inherent noisiness of flow cytometry measurements. Finally, the histogram corresponding to each antibody variant is used to compute an ‘average bin number’ (Figure 1C, dots), which serves as a proxy measurement for the average amount of bound antigen per cell.

Figure 1.

Schematic illustration of Tite-Seq.

(A) A library of variant antibodies (various colors) are displayed on the surface of yeast cells (tan). (B) The library is exposed to antigen (green triangles) at a defined concentration, cell-bound antigen is fluorescently labeled, and FACS is used to sort cells into bins according to measured fluorescence. (C) The antibody variants in each bin are sequenced and the distribution of each variant across bins is computed (histograms; colors correspond to specific variants). The mean bin number (dot) is then used to quantify the typical amount of bound antigen per cell. (D) Binding titration curves (solid lines) and corresponding values (vertical lines) can be inferred for individual antibody sequences by using the mean fluorescence values (dots) obtained from flow cytometry experiments performed on clonal populations of antibody-displaying yeast. (E) Tite-Seq consists of performing the Sort-Seq experiment in panels A–C at multiple antigen concentrations, then inferring binding curves using mean bin number as a proxy for mean cellular fluorescence. This enables measurements for thousands of variant antibodies in parallel. We note that the Tite-Seq results illustrated in panel E were simulated using three bins under idealized experimental conditions, as described in Appendix 1. The inference of binding curves from real Tite-Seq data is more involved than this panel might suggest, due to the multiple sources of experimental noise that must be accounted for.

DOI: http://dx.doi.org/10.7554/eLife.23156.003

It has previously been shown that values can be accurately measured using yeast-displayed antibodies by taking binding titration curves, i.e., by measuring the average amount of bound antigen as a function of antigen concentration (VanAntwerp and Wittrup, 2000; Gai and Wittrup, 2007). The median fluorescence of labeled cells is expected to be related to antigen concentration via

where is proportional to the number of functional antibodies displayed on the cell surface, accounts for background fluorescence, and is the concentration of free antigen in solution. Figure 1D illustrates the shape of curves having this form. By using flow cytometry to measure on clonal populations of yeast at different antigen concentrations , one can infer curves having the sigmoidal form shown in Equation 1 and thereby learn . Such measurements, however, can only be performed in a low-throughput manner.

Tite-Seq allows thousands of binding titration curves to be measured in parallel. The Sort-Seq procedure illustrated in Figure 1A–C is performed at multiple antigen concentrations, and the resulting average bin number for each variant antibody is plotted against concentration. Sigmoidal curves are then fit to these proxy measurements, enabling values to be inferred for each variant.

We emphasize that values cannot, in general, be accurately inferred from Sort-Seq experiments performed at a single antigen concentration. Because the relationship between binding and is sigmoidal, the amount of bound antigen provides a quantitative readout of only when the concentration of antigen used in the labeling procedure is comparable in magnitude to . However, single mutations within a protein binding domain often change by multiple orders of magnitude. Sort-Seq experiments used to measure sequence-affinity landscapes must therefore be carried out over a range of concentrations large enough to encompass this variation.

Furthermore, as illustrated in Figure 1C and D, different antibody variants often lead to different levels of functional antibody expression on the yeast cell surface. If one performs Sort-Seq at a single antigen concentration, high affinity (low ) variants with low expression (blue variant) may bind less antigen than low affinity (high ) variants with high expression (orange variant). Only by measuring full titration curves can the effect that sequence has on affinity be deconvolved from sequence-dependent effects on functional protein expression.

Proof-of-principle Tite-Seq experiments

To test the feasibility of Tite-Seq, we used a well-characterized antibody-antigen system: the 4-4-20 single chain variable fragment (scFv) antibody (Boder and Wittrup, 1997), which binds the small molecule fluorescein with nM (Gai and Wittrup, 2007). This system was used in early work to establish the capabilities of yeast display (Boder and Wittrup, 1997), and a high resolution co-crystal structure of the 4-4-20 antibody bound to fluorescein, shown in Figure 2A, has been determined (Whitlow et al., 1995). An ultra-high-affinity ( fM) variant of this scFv, called 4m5.3, has also been found (Boder et al., 2000). In what follows, we refer to the 4-4-20 scFv from Boder and Wittrup (1997) as WT, and the 4m5.3 variant from Boder et al. (2000) as OPT.

Figure 2.

Yeast display construct and antibody libraries

(A) Co-crystal structure of the 4-4-20 (WT) antibody from Whitlow et al. (1995) (PDB code 1FLR). The CDR1H and CDR3H regions are colored blue and red, respectively. (B) The yeast display scFv construct from Boder and Wittrup (1997) that was used in this study. Antibody-bound antigen (fluorescein) was visualized using PE dye. The amount of surface-expressed protein was separately visualized using BV dye. Approximate location of the CDR1H (blue) and CDR3H (red) regions within the scFv are illustrated. (C) The gene coding for this scFv construct, with the six CDR regions indicated. The WT sequence of the two 10 aa variable regions are also shown. (D) The number of 1-, 2-, and 3-codon variants present in the 1H and 3H scFv libraries. Figure 2—figure supplement 1 shows the cloning vector used to construct the CDR1H and CDR3H libraries, as well as the form of the resulting expression plasmids.

DOI: http://dx.doi.org/10.7554/eLife.23156.004

(A) The iRA11 amplicon library, which was prepared from microarray-synthesized oligos containing variant CDR1H or variant CDR3H regions. This amplicon is flanked by inward-facing BsaI restriction sites. (B) The pRA10 cloning vector, which contains the ccdB selection gene within a cassette flanked by outward-facing BsmBI restriction sites. (C) The pRA11 plasmid library, which was cloned by ligating BsaI-digested iRA11 amplicons and BsmBI-digest pRA10 vector. (D) The sequencing amplicon that was amplified from sorted cells after Tite-Seq and Sort-Seq experiments and submitted for ultra-high-throughput DNA sequencing. Appendix 3 provides more details about iRA11 amplicons, the pRA10 vector, and the pRA11 plasmid library. Appendix 4 provides more information about the creation of sequencing amplicons.

DOI: http://dx.doi.org/10.7554/eLife.23156.005

Figure 2—figure supplement 1.

Cloning strategy.

(A) The iRA11 amplicon library, which was prepared from microarray-synthesized oligos containing variant CDR1H or variant CDR3H regions. This amplicon is flanked by inward-facing BsaI restriction sites. (B) The pRA10 cloning vector, which contains the ccdB selection gene within a cassette flanked by outward-facing BsmBI restriction sites. (C) The pRA11 plasmid library, which was cloned by ligating BsaI-digested iRA11 amplicons and BsmBI-digest pRA10 vector. (D) The sequencing amplicon that was amplified from sorted cells after Tite-Seq and Sort-Seq experiments and submitted for ultra-high-throughput DNA sequencing. Appendix 3 provides more details about iRA11 amplicons, the pRA10 vector, and the pRA11 plasmid library. Appendix 4 provides more information about the creation of sequencing amplicons.

DOI: http://dx.doi.org/10.7554/eLife.23156.005

The scFv was expressed on the surface of yeast as part of the multi-domain construct illustrated in Figure 2B and previously described in Boder and Wittrup (1997). Following (Boder et al., 2000), we used fluorescein-biotin as the antigen and labeled scFv-bound antigen with streptavidin-RPE (PE). The amount of surface-expressed protein was separately quantified by labeling the C-terminal c-Myc tag using anti-c-Myc primary antibodies, followed by secondary antibodies conjugated to Brilliant Violet 421 (BV). See Appendix 2 for details on this labeling procedure.

Two different scFv libraries were assayed simultaneously. In the ‘1H’ library, a 10 aa region encompasing the CDR1H region of the WT scFv (see Figure 2C) was mutagenized using microarray-synthesized oligos (see Appendix 3 for details). The resulting 1H library consisted of all 600 single-codon variants of this 10 aa region, 1100 randomly chosen 2-codon variants, and 150 random 3-codon variants (Figure 2D). An analogous ‘3H’ library was generated for a 10 aa region containing the CDR3H region of this scFv. In all of the Tite-Seq experiments described below, these two libraries were pooled together and supplemented with WT and OPT scFvs, as well with a nonfunctional scFv referred to as .

Tite-Seq was carried out as follows. Yeast cells expressing scFv from the mixed library were incubated with fluorescein-biotin at one of eleven concentrations: 0 M, M, M, M, M, M, M, M, M, M, and M. After subsequent PE labeling of bound antigen, cells were sorted into four bins using FACS (Figure 3A). Separately, BV-labeled cells were sorted according to measured scFv expression levels (Figure 3B). The number of cells sorted into each bin is shown in Figure 3C. Each bin of cells was regrown and bulk DNA was extracted. The 1H and 3H variable regions were then PCR amplified and sequenced using paired-end Illumina sequencing, as described in Appendix 4. The final data set consisted of an average of sequences per bin across all 48 bins (Figure 3D). Three independent replicates of this experiment were performed on three different days.

Figure 3.

Details of our Tite-Seq experiments.

(A) Gates used to sort cells based on PE fluorescence, which provides a readout of bound antigen. Cells were labeled at the eleven different antigen concentrations. Shades of red indicate the four fluorescence gates used to sort cells; these correspond to bins 0, 1, 2, and 3 (from left to right). (B) Gates, indicated in shades of purple, used to sort cells based on BV fluorescence, which provides a readout of antibody expression. (C) The number of cells sorted into each bin. (D) The number of Illumina reads obtained from each bin of sorted cells after quality control measures were applied. The data shown in this figure corresponds to a single Tite-Seq experiment. Figure 3—figure supplement 1 and Figure 3—figure supplement 2 show data for two independent replicates of this experiment.

DOI: http://dx.doi.org/10.7554/eLife.23156.006

Analog of Figure 3 in the main text, but for the replicate 2 Tite-Seq experiment.

DOI: http://dx.doi.org/10.7554/eLife.23156.007

Analog of Figure 3 in the main text, but for the replicate 3 Tite-Seq experiment.

DOI: http://dx.doi.org/10.7554/eLife.23156.008

Figure 3—figure supplement 1.

Tite-Seq experiment, replicate 2.

Analog of Figure 3 in the main text, but for the replicate 2 Tite-Seq experiment.

DOI: http://dx.doi.org/10.7554/eLife.23156.007

Figure 3—figure supplement 2.

Tite-Seq experiment, replicate 3.

Analog of Figure 3 in the main text, but for the replicate 3 Tite-Seq experiment.

DOI: http://dx.doi.org/10.7554/eLife.23156.008

For each variant scFv gene, a value was inferred by fitting a binding curve to the resulting Tite-Seq data, with separate curves independently fit to data from each Tite-Seq experiment (Figure 4A). As illustrated in Figure 1E, this fitting procedure uses the sigmoidal function in Equation 1 to model mean bin number as a function of antigen concentration. However, the need to account for multiple sources of noise in the Tite-Seq experiment necessitates a more complex procedure than Figure 1E might suggest; the details of this inference procedure are described in Appendix 5.

Figure 4.

Accuracy and precision of Tite-Seq.

(A) Binding curves and measurements inferred from Tite-Seq data. (B) Mean fluorescence values (dots) and corresponding inferred binding curves (lines) obtained by flow cytometry measurements for five selected scFvs (WT, OPT, C5, C45, and C107). In (A,B), values corresponding to 0 M fluorescein are plotted on the left-most edge of the plot, dotted lines show the upper ( M) and lower ( M) limits on sensitivity, vertical lines show inferred values, and different shades correspond to different replicate experiments. (C) Comparison of the Tite-Seq-measured and flow-cytometry-measured values for all clones tested. Colors indicate different scFv protein sequences as follows: WT (purple), OPT (green), (black), 1H clones (blue), and 3H clones (red). Each value indicates the mean value obtained across all replicates, with error bars indicating standard error. Clones with outside of the affinity range are drawn on the boundaries of this range, which are indicated with dotted lines. The coefficient of determination () between log Tite-Seq values and log flow values includes clones outside of the affinity range; in such cases, the corresponding boundary value ( M or M) has been used. The amino acid sequences and measured values for all clones tested are provided in Table 1. Figure 4—figure supplement 1 provides plots, analogous to panels A and B, for all of the assayed clones. Figure 4—figure supplement 2 compares and values obtained across all three Tite-Seq replicates. Figure 4—figure supplement 3 quantifies measurement error using synonymous mutants. Figure 4—figure supplement 4 provides information about library composition. Figure 4—figure supplement 5 illustrates the poor correlation between scFv enrichment and Tite-seq measured values. Figure 4—figure supplement 6 shows a 2-fold difference in the specific activities of OPT and WT scFvs. Figure 4—figure supplement 7 illustrates the simulations we used in Figure 4—figure supplement 8 to validate the ability of our analysis to infer correct values.

DOI: http://dx.doi.org/10.7554/eLife.23156.009

Binding curves, measured using (A) Tite-Seq or (B) flow cytometry, for all clones analyzed in this paper and described in Table 1. Plots are drawn as in Figure 4, panels A and B.

DOI: http://dx.doi.org/10.7554/eLife.23156.010

Density plots of (A) Tite-Seq-measured values and (B) Sort-Seq-measured values between all pairs of replicate experiments. Measurements for these quantities that were judged to be of low precision due to low sequence counts are not plotted. indicates the percentage of total assayed sequences plotted; is the Pearson correlation and includes clonal measurements outside the boundaries of our measurable ranges ( M for , 0–2 for expression). Clones outside of these ranges were given values at the closest boundary.

DOI: http://dx.doi.org/10.7554/eLife.23156.011

Density plots for (A) Tite-Seq-measured log standard deviation and average log and (B) Sort-Seq-measured standard deviation and average are shown for each scFv sequence with more than one synonymous mutant for each of the replicate experiments. The error peaked between M. The expression error peaked at or above WT expression (i.e. 1) levels.

DOI: http://dx.doi.org/10.7554/eLife.23156.012

(A) Comparison of library composition between all pairs of replicate experiments. (B) Zipf plots showing the library composition in each replicate experiment. In both panels, the prevalence of each scFv sequence in each replicate experiment was determined as part of the Tite-Seq curve fitting procedure, as described in Appendix 5.

DOI: http://dx.doi.org/10.7554/eLife.23156.013

To assess how well simple enrichment calculations might reproduce the values measured by Tite-Seq, we did the following calculation. For each of the two libraries (1 H and 3 H), we partitioned scFvs into seven groups based on their measured s (columns). For each group at each antigen concentration (rows), we then computed the enrichment of each scFv in the high PE bins (bins 2,3) relative to the low PE bins (bins 0,1). In these enrichment calculations, the number of counts in each bin was re-weighted to accurately reflect the fraction of library cells falling within the fluorescence range of that bin. This figure shows the resulting Spearman rank correlation between enrichment and log values computed for each scFv group at each antigen concentration. In both libraries, we see that correlation values above background (which can be assessed from the values in the 0 M fluorescein row) only occur close to the diagonal, i.e., when is close to the fluorescein concentration used.

DOI: http://dx.doi.org/10.7554/eLife.23156.014

2D flow cytometry histograms showing both OPT- and WT-expressing cells labeled with PE and BV after incubation at 2 M fluorescein. At this fluorescein concentration, nearly all functional WT and OPT scFvs are bound. Regression lines (fixed to have slope 1) were fit to data points with BV signal between and . The vertical shift of the OPT data relative to the WT data indicates a factor of difference (computed from four replicate experiments) in the amount labeled antigen. This difference is not due to a difference in the number of surface-displayed scFvs, as this would cause the OPT and WT clouds to lie along the same diagonal. Rather, this difference between WT and OPT is due to variation in specific activity.

DOI: http://dx.doi.org/10.7554/eLife.23156.015

Realistic Tite-Seq data were simulated separately for each distinct pair of affinity () and amplitude () values, as described in Appendix 7. This figure shows simulated data, akin to the data displayed in Figure 4—figure supplement 6, for WT values of and .

DOI: http://dx.doi.org/10.7554/eLife.23156.016

values were inferred for Tite-Seq data simulated using (green) the same number of cells, (light green) times as many cells, or (black) times as many sorted cells as in our experiments. Areas indicate approximately plus or minus one standard deviation in the fitted values obtained for each true value.

DOI: http://dx.doi.org/10.7554/eLife.23156.017

Table 1.

Clones measured using flow cytometry and Tite-Seq. List of scFv clones, ordered by their flow-cytometry-measured values. With the exception of OPT and , these clones differed from WT only in their 1H and 3H variable regions. WT amino acids within these regions are capitalized; variant amino acids are shown in lower case. No sequence is shown for because this clone contained a large deletion, making identification of the 1H and 3H variable regions meaningless. values saturating our lower detection limit of M or upper detection limit of M are written with a or sign to emphasize the uncertainty in these measurements. Tite-Seq values indicate mean and standard errors computed across the three replicate Tite-Seq experiments; they are not averaged across synonymous variants.

DOI: http://dx.doi.org/10.7554/eLife.23156.018

Name	1H variable region	3H variable region	No. replicates (flow)	K_D [M] (flow)	K_D [M] (Tite-Seq)
OPT	TFghYWMNWV	GasYGMeYlG	3	≲10−9.5	≲10−9.5
C107	TFSDYWMNWV	GaYYGMDYWG	3	10−9.28±0.04	10−9.18±0.11
C112	TFSDYWMNWV	GSYYGMDYcG	3	10−8.95±0.07	10−9.19±0.14
WT	TFSDYWMNWV	GSYYGMDYWG	10	10−8.61±0.07	10−8.92±0.10
C144	vFSDYWMNWV	GSYYGMDYWG	3	10−8.57±0.03	10−8.86±0.04
C133	aFSDYWMNWV	GSYYGMDYWG	3	10−8.55±0.06	10−8.62±0.09
C132	TFmDYWlNWV	GSYYGMDYWG	3	10−8.48±0.08	10−8.38±0.29
C94	TFSDYWMNWV	GSYYGMDsWG	3	10−8.46±0.06	10−8.50±0.04
C5	TFSDYWiNWV	GSYYGMDYWG	3	10−8.34±0.10	10−8.55±0.09
C93	TFSDYWMNWV	GSYrGMDYWG	3	10−7.35±0.08	10−7.60±0.70
C39	TFSDYWMNWV	GSYYGMDYWa	3	10−7.08±0.20	10−7.28±0.17
C102	TFSDYWMNWV	sSkYGMDYWG	3	10−5.76±0.16	10−7.25±0.60
C22	ssSDYWMNWV	GSYYGMDYWG	3	10−5.69±0.31	10−7.53±0.07
C7	hFSDYWMNWl	GSYYGMDYWG	3	10−5.53±0.18	10−5.39±0.18
C45	TFSDYWMNWV	GSYdGnDYWG	3	10−5.40±0.24	≳10−5.0
C103	TFSDYWMNWV	GSYYGMDlWG	3	10−5.15±0.47	10−5.44±0.55
C3	TFSDYWMsWV	GSYYGMDYWG	3	≳10−5.0	≳10−5.0
C18	TFSDYsMNWV	GSYYGMDYWG	3	≳10−5.0	≳10−5.0
Δ	–	–	12	≳10−5.0	≳10−5.0

Figure 4—figure supplement 1.

Binding curves for all clones.

Binding curves, measured using (A) Tite-Seq or (B) flow cytometry, for all clones analyzed in this paper and described in Table 1. Plots are drawn as in Figure 4, panels A and B.

DOI: http://dx.doi.org/10.7554/eLife.23156.010

Figure 4—figure supplement 2.

Concordance between replicate experiments.

Density plots of (A) Tite-Seq-measured values and (B) Sort-Seq-measured values between all pairs of replicate experiments. Measurements for these quantities that were judged to be of low precision due to low sequence counts are not plotted. indicates the percentage of total assayed sequences plotted; is the Pearson correlation and includes clonal measurements outside the boundaries of our measurable ranges ( M for , 0–2 for expression). Clones outside of these ranges were given values at the closest boundary.

DOI: http://dx.doi.org/10.7554/eLife.23156.011

Figure 4—figure supplement 3.

Error estimates from synonymous mutants.

Density plots for (A) Tite-Seq-measured log standard deviation and average log and (B) Sort-Seq-measured standard deviation and average are shown for each scFv sequence with more than one synonymous mutant for each of the replicate experiments. The error peaked between M. The expression error peaked at or above WT expression (i.e. 1) levels.

DOI: http://dx.doi.org/10.7554/eLife.23156.012

Figure 4—figure supplement 4.

Composition of scFv libraries.

(A) Comparison of library composition between all pairs of replicate experiments. (B) Zipf plots showing the library composition in each replicate experiment. In both panels, the prevalence of each scFv sequence in each replicate experiment was determined as part of the Tite-Seq curve fitting procedure, as described in Appendix 5.

DOI: http://dx.doi.org/10.7554/eLife.23156.013

Figure 4—figure supplement 5.

Sort-Seq enrichment correlates poorly with Tite-Seq-measured affinity.

To assess how well simple enrichment calculations might reproduce the values measured by Tite-Seq, we did the following calculation. For each of the two libraries (1 H and 3 H), we partitioned scFvs into seven groups based on their measured s (columns). For each group at each antigen concentration (rows), we then computed the enrichment of each scFv in the high PE bins (bins 2,3) relative to the low PE bins (bins 0,1). In these enrichment calculations, the number of counts in each bin was re-weighted to accurately reflect the fraction of library cells falling within the fluorescence range of that bin. This figure shows the resulting Spearman rank correlation between enrichment and log values computed for each scFv group at each antigen concentration. In both libraries, we see that correlation values above background (which can be assessed from the values in the 0 M fluorescein row) only occur close to the diagonal, i.e., when is close to the fluorescein concentration used.

DOI: http://dx.doi.org/10.7554/eLife.23156.014

Figure 4—figure supplement 6.

Differing specific activities of OPT and WT.

2D flow cytometry histograms showing both OPT- and WT-expressing cells labeled with PE and BV after incubation at 2 M fluorescein. At this fluorescein concentration, nearly all functional WT and OPT scFvs are bound. Regression lines (fixed to have slope 1) were fit to data points with BV signal between and . The vertical shift of the OPT data relative to the WT data indicates a factor of difference (computed from four replicate experiments) in the amount labeled antigen. This difference is not due to a difference in the number of surface-displayed scFvs, as this would cause the OPT and WT clouds to lie along the same diagonal. Rather, this difference between WT and OPT is due to variation in specific activity.

DOI: http://dx.doi.org/10.7554/eLife.23156.015

Figure 4—figure supplement 7.

Realistic Tite-Seq simulations.

Realistic Tite-Seq data were simulated separately for each distinct pair of affinity () and amplitude () values, as described in Appendix 7. This figure shows simulated data, akin to the data displayed in Figure 4—figure supplement 6, for WT values of and .

DOI: http://dx.doi.org/10.7554/eLife.23156.016

Figure 4—figure supplement 8.

Validation of analysis pipeline.

values were inferred for Tite-Seq data simulated using (green) the same number of cells, (light green) times as many cells, or (black) times as many sorted cells as in our experiments. Areas indicate approximately plus or minus one standard deviation in the fitted values obtained for each true value.

DOI: http://dx.doi.org/10.7554/eLife.23156.017

Separately, the Sort-Seq data obtained by sorting the BV-labeled libraries were used to determine the expression level of each scFv. Specifically, we use to denote (for each scFv in the library) the mean bin number that results from this expression-based sorting; this value provides a measurement of the surface expression level of that scFv. All values have been scaled so that the mean of such measurements for all synonymous WT scFv gene variants is 1.0.

Low-throughput validation experiments

To judge the accuracy of Tite-Seq, we separately measured binding curves for individual scFv clones as described for Figure 1D. In addition to the WT, OPT, and scFvs, we assayed eight clones from the 1H library (named C3, C5, C7, C18, C22, C132, C133 and C144) and eight clones from the 3H library (C39, C45, C93, C94, C102, C103, C107, C112). Each clone underwent the same labeling procedure as in the Tite-Seq experiment, after which median fluorescence values were measured using standard flow cytometry. values were then inferred by fitting binding curves of the form in Equation 1 using the procedure described in Appendix 6. These curves, which can be directly compared to Tite-Seq measurements (Figure 4A), are plotted in Figure 4B; at least three replicate binding curves were measured for each clone. See Figure 4—figure supplement 1 for the titration curves of all the tested clones.

Tite-Seq can measure dissociation constants

Figure 4C reveals a strong correspondence between the values measured by Tite-Seq and those measured using low-throughput flow cytometry. The robustness of Tite-Seq is further illustrated by the consistency of values measured for the WT scFv. Using Tite-Seq, and averaging the results from the 33 synonymous variants and over all three replicates, we determined M for the WT scFv. These measurements are largely consistent with the measurement of M obtained by averaging low-throughput flow cytometry measurements across 10 replicates, and coincides with the previously measured value of nM M reported in (Gai and Wittrup, 2007). The three independent replicate Tite-Seq experiments give reproducible results as measured by direct comparison (Figure 4—figure supplement 2), from synonymous mutant variation (Figure 4—figure supplement 3) and library composition Figure 4—figure supplement 4) with Pearson coefficients ranging from to for all the measured values between replicates; note that values outside of the sensitivity range are included in the calculation of these Pearson coefficients as described in the Figure 4 caption.

The error bars for values in Figure 4C calculated from the variability of the fits to different replicates therefore support the reproducibility of the experiment. The main discrepancy in these error bar calculations occurred for clones c22 and c102 (see also Figure 4—figure supplement 1). The reason for this discrepancy is currently unclear. We note that Tite-Seq-measured values for these two clones are close to M, and that the analysis of synonymous variants (Figure 4—figure supplement 3) found that Tite-Seq-measured s in this region exhibited the largest variations.

The necessity of performing measurements over a wide range of antigen concentrations is illustrated in Figure 4—figure supplement 5. At each antigen concentration used in our Tite-Seq experiments, the enrichment of scFvs in the high-PE bins correlated poorly with the values inferred from full titration curves. Moreover, at each antigen concentration used, a detectable correlation between and enrichment was found only for scFvs with values close to that concentration.

Figure 4—figure supplement 6 suggests a possible reason for the weak correlation between values and enrichment in high-PE bins. We found that, at saturating concentrations of fluorescein (M), cells expressing the OPT scFv bound twice as much fluorescein as cells expressing the WT scFv. This difference was not due to variation in the total amount of displayed scFv, which one might control for by labeling the c-Myc epitope as in Reich et al. (2015). Rather, this difference in binding reflects a difference in the specific activity of displayed scFvs. Yeast display experiments performed at a single antigen concentration cannot distinguish such differences in specific activity from differences in scFv affinity.

To further test the capability of Tite-Seq to infer dissociation constants from sequencing data over a wide range of values, as well as to validate our analysis procedures, we simulated Tite-Seq data in silico and analyzed the results using the same analysis pipeline that we used for our experiments. Details about the simulations are given in Appendix 7. The simulated data is illustrated in Figure 4—figure supplement 7. values inferred from these simulated data agreed to high accuracy with the used in the simulation (Figure 4—figure supplement 8), thus validating our analysis pipeline.

Properties of the affinity and expression landscapes

Figure 5 shows the effect that every single-amino-acid substitution mutation within the 1H and 3H variable regions has on affinity and on expression; histograms of these effects are provided in Figure 5—figure supplement 1. In both regions, the large majority of mutations weaken antigen binding (1H: 88%; 3H: 93%), with many mutations increasing above our detection threshold of M (1H: 36%; 3H: 52%). Far fewer mutations reduced (1H: 12%; 3H: 7%), and very few dropped below our detection limit of M (1H: 0%; 3H: 3%). Histograms of the effect of two or three amino acid changes relative to WT, shown in Figure 5—figure supplement 2A, reveal that multiple random mutations tend to further reduce affinity. We also observed that mutations within the 3H variable region have a larger effect on affinity than do mutations in the 1H variable region. Specifically, single amino acid mutations in 3H were seen to increased more than mutations in 1H (1H median ; 3H median ; , one-sided Mann-Whitney U test). This result suggests that binding affinity is more sensitive to variation in CDR3H than to variation in CDR1H, a finding that is consistent with the conventional understanding of these antibody CDR regions (Xu and Davis, 2000; Liberman et al., 2013).

Figure 5.

Effects of substitution mutations on affinity and expression.

Heatmaps show the measured effects on affinity (A,B) and expression (C,D) of all single amino acid substitutions within the variables regions of the 1H (A,C) and 3H (B,D) libraries. Purple dots indicate residues of the WT scFv. Green dots indicate non-WT residues in the OPT scFv. Figure 5—figure supplement 1 provides histograms of the non-WT values displayed in panels A–D. Figure 5—figure supplement 2 compares the effects on of both single-point and multi-point mutations.

DOI: http://dx.doi.org/10.7554/eLife.23156.020

(A,B) Histogram showing the values measured for all substitution mutations in the 1 H (A) and 3 H (B) libraries. Note that these are the values plotted in panels A and B of Figure 5, except that the WT value is not included. Dashed lines indicate the of the WT scFv; dotted lines indicate thresholds just within our detection boundaries, M and M, while the colored bars outside this interval indicate the number of substitution mutations with above (blue) and below (red) this range. (C,D) Histogram of values for all single-substitution variants in the 1 H (C) or 3 H (D) libraries. These values, save those of the WT scFv, are plotted in panels C and D of Figure 5. Dashed lines indicate the WT expression level of .

DOI: http://dx.doi.org/10.7554/eLife.23156.021

The effect of 1, 2, or three mutations on (A) Tite-Seq-measured values or (B) Sort-Seq-measured values. Plots show the relative probability density (over 30 bins along the or axes) observed for variants in each class.

DOI: http://dx.doi.org/10.7554/eLife.23156.022

Figure 5—figure supplement 1.

Histograms of substitution effects on affinity and expression.

(A,B) Histogram showing the values measured for all substitution mutations in the 1 H (A) and 3 H (B) libraries. Note that these are the values plotted in panels A and B of Figure 5, except that the WT value is not included. Dashed lines indicate the of the WT scFv; dotted lines indicate thresholds just within our detection boundaries, M and M, while the colored bars outside this interval indicate the number of substitution mutations with above (blue) and below (red) this range. (C,D) Histogram of values for all single-substitution variants in the 1 H (C) or 3 H (D) libraries. These values, save those of the WT scFv, are plotted in panels C and D of Figure 5. Dashed lines indicate the WT expression level of .

DOI: http://dx.doi.org/10.7554/eLife.23156.021

Figure 5—figure supplement 2.

Effects of multi-point mutations on affinity and expression.

The effect of 1, 2, or three mutations on (A) Tite-Seq-measured values or (B) Sort-Seq-measured values. Plots show the relative probability density (over 30 bins along the or axes) observed for variants in each class.

DOI: http://dx.doi.org/10.7554/eLife.23156.022

Our observations are thus fully consistent with the hypothesis that the amino acid sequences of the CDR1H and CDR3H regions of the WT scFv have been selected for high affinity binding to fluorescein. We know this to be true, of course; still, this result provides an important validation of our Tite-Seq measurements.

To further validate our Tite-Seq affinity measurements, we examined positions in the high affinity OPT scFv (from [Boder et al., 2000]) that differ from WT and that lie within the 1H and 3H variable regions. As illustrated in Figure 5A and B, five of the six OPT-specific mutations reduce or are nearly neutral. Previous structural analysis (Midelfort et al., 2004) has suggested that D106E, the only OPT mutation that we find significantly increases , may indeed disrupt antigen binding on its own while still increasing affinity in the presence of the S101A mutation.

Next, we used our measurements to build a ‘matrix model’ (also known as a ‘position-specific affinity matrix,’ or PSAM [Foat et al., 2006]) describing the sequence-affinity landscape of these two regions. Our model assumed that the value for an arbitrary amino acid sequence could be computed from the value of the WT scFv, plus the measured change in produced by each amino acid substitution away from WT. We evaluated our matrix models on the 1H and 3H variable regions of OPT, finding an affinity of M. Our simple model for the sequence affinity landscape of this scFv therefore correctly predicts that OPT has higher affinity than WT. The quantitative affinity predicted by our model does not match the known affinity of the OPT scFv ( M), but this is unsurprising for three reasons. First the OPT scFv differs from WT in 14 residues, only 6 of which are inside the 1H and 3H variable regions assayed here. Second, one of the OPT mutations (W108L) reduces below our detection threshold of M; in building our matrix model, we set this value equal to , knowing it would likely underestimate the affinity-increasing effect of the mutation. Third, our additive model ignores potential epistatic interactions. Still, we thought it worth asking how likely it it would be for six random mutations within the 1H and 3H variable regions to reduce affinity as much as our model predicts for OPT. We therefore simulated a large number () of variants having a total of 6 substitution mutations randomly scattered across the 1H and 3H variable regions. The fraction of these random sequences that had an affinity at or below our predicted affinity for OPT was . This finding is fully consistent with the fact that the mutations in OPT relative to WT were selected for increased affinity, an additional confirmation of the validity of our Tite-Seq measurements.

The sequence-expression landscape measured in our separate Sort-Seq experiment yielded qualitatively different results (Figure 5C and D). We observed no significant difference in the median effect that mutations in the variable regions of 1H (median ) versus 3H (median ) have on expression (, two-sided Mann-Whitney U test); see also Figure 5—figure supplement 1. The variance in these effects, however, was larger in 3H than in 1H (, Levene’s test). These results suggest two things. First, the 3H variable region appears to have a larger effect on scFv expression than the 1H variable region has. At the same time, since we observe fewer beneficial mutations in 1H (Figure 5C) than in 3H (Figure 5D), the WT sequence appears to be more highly optimized for expression in CDR1H than in CDR3H. The effect of double or triple mutations further reduced expression in both CDRs (Figure 5—figure supplement 2B), similar to what was observed for affinity.

Structural correlates of the sequence-affinity landscape

We asked if the sensitivity of the antibody to mutations could be understood from a structural perspective. To quantify sensitivity of affinity and expression at each position , we computed two quantities:

Here, and respectively denote the dissociation constant and expression level measured for the WT scFv, and denote analogous quantities for the scFv with a single substitution mutation of amino acid at position , and denotes an average computed over the 19 non-WT amino acids at that position.

Figure 6A shows the known structure (Whitlow et al., 1995) of the 1H and 3H variable regions of the WT scFv in complex with fluorescein. Each residue is colored according to the and values computed for its position. To get a better understanding of what aspects of the structure might govern affinity, we plotted values against two other quantities: the number of amino acid contacts made by the WT residue within the antibody structure (Figure 6B), and the distance between the WT residue and the antigen (Figure 6C). We found a strong correlation between and the number of contacts, but no significant correlation between and distance to antigen. By contrast, did not correlate significantly with either of these structural quantities (Figure 6D and E).

Figure 6.

Structural context of mutational effects.

(A) Crystal structure (Whitlow et al., 1995) of the CDR1H and CDR3H variable regions of the WT scFv in complex with fluorescein (green). Each residue (CDR1H: positions 28–37; CDR3H: positions 100–109) is colored according to the and values computed for that position. These variables, and , respectively quantify the sensitivity of and to amino acid substitutions at each position, with larger values corresponding to greater sensitivity; see Equations 2 and 3 for definitions of these quantities. (B,C) For each position in the CDR1H and CDR3H variable regions, is plotted against either (B) the number of contacts the WT residue makes within the protein structure, or (C) the distance of the WT residue to the fluorescein molecule. (D,E) Similarly, is plotted against either (D) the number of contacts or (E) the distance to the antigen. is the coefficient of determination.

DOI: http://dx.doi.org/10.7554/eLife.23156.023

Discussion

We have described a massively parallel assay, called Tite-Seq, for measuring the sequence-affinity landscape of antibodies. The range of affinities measured in our Tite-Seq experiments ( M to M) includes a large fraction of the physiological range relevant to affinity maturation ( M to ~10−6 M) (Batista and Neuberger, 1998; Foote and Eisen, 1995; Roost et al., 1995). Expanding the measured range of affinities below M might require larger volume labeling reactions, but would be straight-forward. Tite-Seq therefore provides a potentially powerful method for mapping the sequence-affinity trajectories of antibodies during the affinity maturation process, as well as for studying other aspects of the adaptive immune response.

The details of our Tite-Seq experiments (e.g., 11 antigen concentrations, four sorting bins per concentration, etc.) were chosen largely for experimental convenience. The effects of varying these parameters have not been systematically explored, and a future investigation of these effects might be valuable. Figure 4—figure supplement 8 does illustrate, via simulation, the effect of read depth on the precision of measured values. These simulations, along with an analysis of synonymous variants (Figure 4—figure supplement 3), suggest that the primary source of noise in our experiments came not from a lack of sorted cells or Illumina reads, but rather from the inefficient post-sort recovery of antibody sequences. We therefore suggest that improvements to our post-sort DNA recovery protocol might substantially improve the resolution of Tite-Seq.

Tite-Seq fundamentally differs from prior DMS experiments in that full binding titration curves, not two-bin enrichment statistics, are used to determine binding affinities. The measurement of binding curves provides three major advantages. First, binding curves provide absolute values in molar units, not just rank-order affinities, like those provided by SORTCERY (Reich et al., 2015), or relative affinity ratios, like those provided by the method of Kowalsky and Whitehead (2016). Second, because ligand binding is a sigmoidal function of affinity, DMS experiments performed at a single ligand concentration (e.g., [Kowalsky and Whitehead, 2016]) are insensitive to receptor s that differ substantially from this ligand concentration. Binding curves, by contrast, integrate measurements over a wide range of concentrations and are therefore sensitive to a wide range of s.

The third advantage of measuring binding curves pertains to the fact that protein sequence determines not just ligand-binding affinity, but also the quantity and specific activity of surface-displayed proteins. Our data (Figure 4—figure supplement 5 and Figure 4—figure supplement 6) suggest that these confounding effects can be large and that they can distort yeast display affinity measurements computed from enrichment statistics gathered at a single antigen concentration. Strong sequence-dependent effects on both the expression and specific activity of yeast-displayed proteins has been reported by other groups as well (e.g., [Burns et al., 2014]), although the absence of such effects has also been reported (e.g., [Kowalsky and Whitehead, 2016]). Ultimately, the magnitude of these effects is likely to vary substantially from protein to protein. It should also be noted that many DMS studies using yeast display (e.g., epitope mapping studies [Kowalsky et al., 2015; Doolan and Colby, 2015; Van Blarcom et al., 2015]) might not suffer from these potentially confounding effects, and in such cases it probably makes sense to employ a simpler experimental design than is required for Tite-Seq. Nevertheless, either Tite-Seq or other experimental methods that assay full binding curves are probably essential if one wants to quantitatively and reliably measure values in a massively parallel fashion.

We wish to emphasize, more generally, that changing a protein’s amino acid sequence can be expected to change multiple biochemical properties of that protein. Our work illustrates the importance of designing massively parallel assays that can disentangle these effects. Tite-Seq provides a general solution to this problem for massively parallel studies of protein-ligand binding. Indeed, the Tite-Seq procedure described here can be readily applied to any protein binding assay that is compatible with yeast display and FACS. Many such assays have been developed (Liu, 2015). We expect that Tite-Seq can also be readily adapted for use with other expression platforms, such as mammalian cell display (Forsyth et al., 2013).

Our Tite-Seq measurements reveal interesting distinctions between the effects of mutations in the CDR1H and CDR3H regions of the anti-fluorescein scFv antibody studied here. As expected, we found that variation in and around CDR3H had a larger effect on affinity than did variation in and around CDR1H. We also found that CDR1H is more optimized for protein expression than is CDR3H, an unexpected finding that appears to be novel. Yeast display expression levels are known to correlate with thermostability (Shusta et al., 1999). Our data is limited in scope, and we remain cautious about generalizing our observations to arbitrary antibody-antigen interactions. Still, this finding suggests the possibility that secondary CDR regions (such as CDR1H) might be evolutionarily optimized to help ensure antibody stability, thereby freeing up CDR3H to encode antigen specificity. If this hypothesis holds, it could provide a biochemical rationale for why CDR3H is more likely than CDR1H to be mutated in functioning receptors (Liberman et al., 2013) and why variation in CDR3H is often sufficient to establish antigen specificity (Xu and Davis, 2000).

Tite-Seq can also potentially shed light on the structural basis for antibody-antigen recognition. By comparing the effects of mutations with the known antibody-fluorescein co-crystal structure (Whitlow et al., 1995), we identified a strong correlation between the effect that a position has on affinity and the number of molecular contacts that the residue at that position makes within the antibody. By contrast, no such correlation of expression with this number of contacts is observed. Again, we are cautious about generalizing from observations made on a single antibody. If our observation were to hold for other antibodies, however, it would suggest that the functional geometry of paratopes might be governed by networks of residues whose positions and orientations are strongly interdependent.

Materials and methods

Tite-Seq was performed as follows. Variant 3H and 1H regions were generated using microarray-synthesized oligos (LC Biosciences, Houston TX. USA). These were inserted into the 4-4-20 scFv of (Boder and Wittrup, 1997) using cassette-replacement restriction cloning as in (Kinney et al., 2010); see Appendix 3. Yeast display experiments were performed as previously described (Boder et al., 2000) with modifications; see Appendix 2. Sorted cells were regrown and bulk DNA was extracted using standard techniques, and amplicons containing the 1H and 3H variable regions were amplified using PCR and sequenced using the Illumina NextSeq platform; see Appendix 4. Three replicate experiments were performed on different days. Raw sequencing data has been posted on the Sequence Read Archive under BioProject ID PRJNA344711. Low-throughput flow cytometry measurements were performed on clones randomly picked from the Tite-Seq library. Sequence data and flow cytometry data were analyzed using custom Python scripts, as described in Appendices 5 and 6. Processed data and analysis scripts are available at github.com/jbkinney/16_titeseq.

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

[Editors’ note: a previous version of this study was rejected after peer review, but the authors submitted for reconsideration. The first decision letter after peer review is shown below.]

Thank you for submitting your work entitled "Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Aviv Regev as the Senior Editor. The following individuals involved in the review of your submission have agreed to reveal their identity: Claudia Bank and Dmitriy Chudakov (peer reviewers).

I regret that after careful discussion with the reviewers, we have come to the decision to reject the manuscript in its current form.

All of us believe that your new approach is extremely exciting and potentially powerful. However, the reviewers identified a number of shortcomings that led us to decide against recommending acceptance of the manuscript. In particular, the manuscript as written clearly is fundamentally a methodology paper rather than a paper reporting a novel new biological result. However, many of the standards necessary to establish the robustness and reproducibility of a new method are not adequately met.

The policy of eLife is to not require revisions that would involve large amounts of new work, and as you will see below it would require substantial new work to address the concerns raised by the reviewers. We recognize that you may therefore either prefer to submit the paper elsewhere, or make extensive changes and re-submit to eLife. Below I outline what I consider to be the main points so that you are aware of them if you choose the latter route. In addition, the full reviewer comments are pasted below.

An experimental replicate is needed to assess reproducibility of the method. While we appreciate the effort to develop an error model using synonymous counts, for a new high-throughput methodology, the only robust ways to assess errors is to perform an independent replicate of the experiment and assess the variation between replicates. Replicates are a standard in recent deep mutational scanning studies (e.g. PMID 25723163, 25006036), and we would require one before accepting the paper. It is not essential that the replicates have perfect correlations (we don't expect that they will), but it is necessary to have some independent means of assessing the noise and validating which findings are robust to this noise.

Reviewer 2's point about what "experimental details" constrained the use of ligand values outside the K range must be addressed – is the method inapplicable to antibodies with higher affinities? If so, that would be a very important limitation. This same point applies to the low-throughput assays. In the original Wittrup paper (cited in this manuscript), low-throughput replicates yielded K estimates that varied by no more than 2-fold. Here, the K estimates vary by two orders of magnitude. The reason appears to be that the range of ligand concentrations does not contain the K, which is a basic requirement of a proper titration curve.

eLife's policy is to encourage making data and computer code available. We encourage you to do this regardless of whether you re-submit to eLife or elsewhere.

The reviews list a variety of other easy-to-fix points, such as better citations of the literature and clarification of unclear points.

Reviewer #1:

In this manuscript the authors present a new method called Tite-Seq to assess the effect of mutations in an antibody on antigen-binding affinity in a high-throughput fashion. They use their method on two regions of the scFv antibody and report a correlation between the number of contacts of a wt residue inside the antibody and its sensitivity to mutations.

Although my expertise in this area is limited, I am generally fascinated by novel high-throughput approaches, and I believe that the presented approach may prove useful to study antigen binding affinities on a large scale. However, I have several concerns regarding the validation of the approach and the appeal of the biological results.

Most importantly, I would like to see a true replicate experiment in order to get an idea of the correlation between measured values on a larger scale than just a couple of low-throughput comparisons. As far as I know, this is common practice when presenting a new approach like this, and it would (1) allow for a better idea of the error (which, in my opinion, is calculated in a highly optimistic way), and (2) allow for a much better quantification of the results. E.g., is the difference of 56% vs. 41% of mutations above the detection limit (in subsection “E. Differing effects of mutations in CDR1H and CDR3H”, first paragraph) a large one, or maybe not even distinguishable given the accuracy of the experiment? Even if the same library was used, a replicate of the subsequent steps would be highly informative, and, in my opinion, necessary for validation of the approach.

I do not see a striking biological result from the analysis. What occurs to me as the main result of the paper (the correlation between contacts of wt residue with sensitivity to mutations at that position) is not too surprising to me, given many studies that have shown protein stability to be an important determinant of its function (and, as I understand, the sensitivity is measured on an absolute scale). Other reported findings seem suspicious and vague, especially considering my main concern expressed above.

As I understand, 1850 mutations per region were surveyed. Such a high number introduces a lot of noise due to sampling by FACS and sequencing, even if the initial library is large and evenly distributed (and especially if the number of sequences is not large for some data points, cf. Figure 2—figure supplement 2B). I may have missed this, but is the distribution of the mutations in the initial library known, and what was the distribution of absolute reads per mutant at each data point?

Reviewer #2:

The authors present Tite-Seq, a method that uses high-throughput DNA sequencing of yeast-displayed antibody libraries to assess mutant K and surface expression at a large scale. In general, the experiments and analysis are well executed, and the paper is mostly well written. My enthusiasm is restrained for two main reasons. First, the authors fail to cite and discuss substantially similar work. Their elucidation of K values for a large library of variants does represent an advance, but it's incremental. The work must be discussed in the context of what has come before. Second, several of the analyses are poorly described and, if I understood them correctly, will inflate the reader's confidence in the method. In particular, the method for estimating error for each K value is insufficient. Both of these points are particularly important because the manuscript is focused on the method itself rather than an important/exciting application. I therefore do not recommend publication of the manuscript in its current form.

Specific comments:

Introduction – The authors should be more thorough in discussing the strengths and weaknesses of prior deep mutational scanning work. In particular, the fact that yeast display has been coupled to deep mutational scanning and that affinity ranking of variants has been achieved is something readers should know before they get to the end of the Discussion (PMID 25311858). Mammalian antibody display has also been coupled to deep mutationals canning (PMID 23765106). Ribosome antibody display and deep mutational scanning using varying ligand concentrations has also been done (PMID 23103372). Of course, the existence of a substantially similar paper (PMID 26296891) should also be acknowledged and the work discussed in that context.

Subsection “C. Low-throughput validation experiments” – The authors show low-throughput validation data for three clones and compare this data to Tite-Seq results. In the subsection “A. Overview of Tite-Seq”, the authors make the point that, if curve fitting is to be successful, the K value must fall within the range of ligand values used. However, for the validation data shown, the fit K values are generally at the very low end of the ligand concentration range. In Figure 3, most of the titration curves more closely resemble flat lines. This can be explained by the fact that the K of the WT is 0.7e-9 M whereas the lowest concentration of ligand used was 1e-8.5 M. I am puzzled as to why the authors chose to employ a ligand concentration range that did not include the known K of the WT sequence. In the Discussion they mention "experimental details," that constrained them, but I'd suggest a fuller and earlier disclosure of these limitations. Given that most antibody-antigen interactions have K at least as low as the one studied, I wonder whether the method could really be generally applied.

Subsection “C. Low-throughput validation experiments” – Pursuant to the previous comment, the authors' simulations (Figure 1) are done using two hypothetical interactions whose K are substantially higher than the experimental case considered. They should do simulations using the actual K and ligand concentrations employed.

Subsection “C. Low-throughput validation experiments” – The low-throughput validation experiments do not themselves look particularly robust. The K values inferred from these experiments range over two orders of magnitude. The authors state "Although individual data points can be noisy, fitting curves to multiple data points nevertheless provide reasonably accurate measurements of affinity. The accuracy of these measurements is increase by averaging over replicates." This is a vague statement. How noisy? How much does replication help? A key problem, I think, is that the low throughput flow validation data looks as noisy as the Tite-Seq data. These experiments should be improved and repeated. Even better, an orthogonal method of validation should be employed.

Figure 2D – This panel and accompanying figure legend is vague. The text makes it clear that the numbers show the diversity of each library, but it would be helpful to clarify the legend.

Figure 4A/subsection “D. Tite-Seq can measure dissociation constants” – In the figure legend the authors state "Error bars on flow K values are the same for all data points; they show the average mean squared error computed computed using three replicate measurements for each clone." In the text, they state "Error bars on flow cytometry K values were computed using the average variance observed in replicate measurements." I have a few problems here. First, neither statement makes totally clear what the authors did/what is shown in Figure 4A. Furthermore, MSE usually used when the parametric value of an estimator is known. For the flow data, we don't have parametric values, just three replicates. So, a confidence interval would be more appropriate. Finally, and most troublingly, the authors have elected to present the mean variance/MSE of all three clones rather than the variance (or CI, or whatever) of each clone independently. I can't think of a good reason to do this. Because the WT replicates happened to look great while the two clones didn't, using the mean of all three is misleading. The authors should improve this analysis and then better explain it. Also, note the repeated "computed."

Figure 4B/Subsection “D. Tite-Seq can measure dissociation constants” – The error estimation procedure for Tite-Seq, based on a single experiment, uses synonymous variants to estimate the standard deviation of each measurement. Variants are binned based on read depth and then a regression is performed to determine the error in each bin. If I understood the procedure, the error is assumed to be wholly dependent on read depth. The read depth vs. signal to noise/error plot (Figure 4B) is really noisy, which suggests that the read depth alone does not capture all of the error. Many factors are known to impact experiment-to-experiment variability (e.g. PCR bias, experimenter error, etc.) in these types of experiments. Replication would enable much more robust error estimation, and would considerably strengthen the work.

Subsection “D. Tite-Seq can measure dissociation constants” – The authors state: "We note that…measurements for the WT scFV are about a factor of 10 larger than the previously measured value of K". I wonder if this is due to the lack of data points at lower concentrations rather than any differences in buffer/etc. Again, a gold-standard, non flow-based assay would help here.

Figure 6A – Because of the two-dimensional heatmap it is somewhat difficult to appreciate the relationship between expression and binding. Two separate panels with individual color schemes might be easier to comprehend.

Appendix C – The authors say they used custom microarray nucleotides to generate the library, but the nature of the sequences on the array is unclear. A supplementary data file containing the sequences ordered should be included.

Appendix F – The authors used the 10% lowest fluorescence values to estimate autofluorescence/background. This is a curious choice, if the library contained mutations to stop codons (see comment above). Stop codons, particularly early ones, virtually guarantee loss of function.

Reviewer #3:

I could hardly evaluate the mathematical aspects of the work. However, it looks like the previous expertise of this team ensures the high mathematical quality.

Concerning the whole idea of the work – I believe it is brilliant.

The approach allows studying the aspects of dependence of antibody affinity and potentially cross-reactivity (with several antigens used) on the sequence landscape, which is beautiful.

[Editors’ note: what now follows is the decision letter after the authors submitted for further consideration.]

Thank you for submitting your heavily revised manuscript "Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves" for consideration by eLife. The manuscript was evaluated by three reviewers who are experts in the field: two were also reviewers for the original submission, and one is a new reviewer. The evaluation has been overseen by Jesse Bloom as the Reviewing Editor and Aviv Regev as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Timothy A Whitehead (Reviewer #1) and Claudia Bank (Reviewer #2).

We anticipate that we can accept your manuscript for publication provided that you make revisions that address the comments below.

First, we appreciate your careful attention to addressing the issues raised in the in the initial review, particularly adding replicates and extending the antigen range. We recognize that these were time-consuming changes, and the seriousness with which you took those critiques has greatly improved this paper. Your study now represents an impressive use of deep sequencing to obtain fairly rigorous K values.

Specific points:

1) Please make sure that your deep sequencing data is available on the SRA and relevant computer code is available as supporting file or on a publicly accessible repository. If this is already the case, please clearly indicate in the manuscript where these can be found.

2) There are a few additional papers that you should consider discussing in the context of prior work:

Doolan KM, Colby DW: Conformation-dependent epitopes recognized by prion protein antibodies probed using mutational scanning and deep sequencing. J. Mol. Biol. 2015, 427:328-340.

Van Blarcom T, Rossi A, Foletti D, Sundar P, Pitts S, Bee C, Melton Witt J, Melton Z, Hasa-Moreno A, Shaughnessy L, et al.: Precise and efficient antibody epitope determination through library design, yeast display and next-generation sequencing. J. Mol. Biol. 2015, 427:1513-1534.

3) Your approach estimates fairly accurate K at the cost of more experiments (multiple binds, multiple expression levels). For certain applications where precise K values are overkill, it may be more efficacious to use the simpler designs like those used in some of the references that you cite – perhaps worth mentioning. Also, do you have any comments on the trade-off between sequencing depth and the accuracy of the inferred K? Similarly for the number of sorting bins? Right now it isn't clear how these were chosen. Clearly your choices worked fine, but it would be nice to explain if there was rigorous rationale for choosing these (alternatively, you could just say that exploring the effects of these parameters is interesting for future work).

4) The relatively large unsigned error on two variants in Figure 4C (approximately 2 orders of magnitude in K) should be commented on. Why is there such a discrepancy (Subsection “D. Tite-Seq can measure dissociation constants”; Figure 4C)?

5) Are the mutations outside of the dynamic range (K > 10 μM) being used in determining correlation coefficients (Figure 4C; Figure 4—figure supplement 1; Figure 4—figure supplement 2)?

6) Please quantify the error in precision and accuracy of the method. In particular, the poorer correspondence between replicates in the binding range between 10-1000 nM K should be mentioned.

7) Can you quantify error within replicates by inferring K for synonymous mutations?

8) How big is the benefit of controlling for active surface-displayed protein? While it is clear that yeast surface expression and folded surface displayed protein varies between variants (Burns et al., 2014), it has been shown (yet still surprising!) that surface expression (and proper folding) effects are modest for yeast surface display experiments, particularly for residues that are reasonably surface-exposed (Kowalsky and Whitehead, 2016).

Burns, Michael L., et al. Directed evolution of brain-derived neurotrophic factor for improved folding and expression in Saccharomyces cerevisiae. Applied and environmental microbiology 2014, 80:5732-5742.

Kowalsky CA, Whitehead TA Determination of binding affinity upon mutation for type I dockerin-cohesin complexes from Clostridium thermocellum and Clostridium cellulolyticum using deep sequencing. PROTEINS 2016 84: 1914-1928

[Editors’ note: the author responses to the first round of peer review follow.]

All of us believe that your new approach is extremely exciting and potentially powerful. However, the reviewers identified a number of shortcomings that led us to decide against recommending acceptance of the manuscript. In particular, the manuscript as written clearly is fundamentally a methodology paper rather than a paper reporting a novel new biological result. However, many of the standards necessary to establish the robustness and reproducibility of a new method are not adequately met. The policy of eLife is to not require revisions that would involve large amounts of new work, and as you will see below it would require substantial new work to address the concerns raised by the reviewers. We recognize that you may therefore either prefer to submit the paper elsewhere, or make extensive changes and re-submit to eLife. Below I outline what I consider to be the main points so that you are aware of them if you choose the latter route. In addition, the full reviewer comments are pasted below. An experimental replicate is needed to assess reproducibility of the method. While we appreciate the effort to develop an error model using synonymous counts, for a new high-throughput methodology, the only robust ways to assess errors is to perform an independent replicate of the experiment and assess the variation between replicates. Replicates are a standard in recent deep mutational scanning studies (e.g. PMID 25723163, 25006036), and we would require one before accepting the paper. It is not essential that the replicates have perfect correlations (we don't expect that they will), but it is necessary to have some independent means of assessing the noise and validating which findings are robust to this noise. Reviewer 2's point about what "experimental details" constrained the use of ligand values outside the KD range must be addressed – is the method inapplicable to antibodies with higher affinities? If so, that would be a very important limitation. This same point applies to the low-throughput assays. In the original Wittrup paper (cited in this manuscript), low-throughput replicates yielded KD estimates that varied by no more than 2-fold. Here, the KD estimates vary by two orders of magnitude. The reason appears to be that the range of ligand concentrations does not contain the KD, which is a basic requirement of a proper titration curve. eLife's policy is to encourage making data and computer code available. We encourage you to do this regardless of whether you re-submit to eLife or elsewhere. The reviews list a variety of other easy-to-fix points, such as better citations of the literature and clarification of unclear points.

We thank the editors for considering our work and for providing this encouraging assessment. We also thank the referees for their careful reading of our work and for providing thoughtful criticism. This excellent feedback has spurred us to perform additional experiments and to modify our analysis methods. We believe that these changes have greatly improved our paper. In particular,

1) The revised manuscript now describes three replicate Tite-Seq experiments. These replicates are used to quantify the uncertainty in Tite-Seq measurements.

2) We have shifted the range of antigen concentrations lower by one decade, thereby bringing the K of the wild type antibody into experimental range. In doing so, we obtained a K for the wild type antibody that is consistent with previous studies.

3) We have increased the number of low-throughput K measurements used to test the validity of the affinity values found by Tite-Seq.

4) We have modified the yeast display protocol to substantially improve the fraction of antibody-displaying cells and thereby increase the precision of both our Tite-Seq and our low-throughput measurements.

5) We have implemented an improved computational method for extracting K values from raw Tite-Seq data. The code for implementing this new method is freely available at https://github.com/jbkinney/16_titeseq

6) We have validated our new analysis method using realistic simulations, which are described in the revised text.

We believe that these changes will address the concerns voiced by the editors and referees. Below we respond to specific referee comments in more detail.

Reviewer #1:

In this manuscript the authors present a new method called Tite-Seq to assess the effect of mutations in an antibody on antigen-binding affinity in a high-throughput fashion. They use their method on two regions of the scFv antibody and report a correlation between the number of contacts of a wt residue inside the antibody and its sensitivity to mutations.

Although my expertise in this area is limited, I am generally fascinated by novel high-throughput approaches, and I believe that the presented approach may prove useful to study antigen binding affinities on a large scale. However, I have several concerns regarding the validation of the approach and the appeal of the biological results.

Most importantly, I would like to see a true replicate experiment in order to get an idea of the correlation between measured values on a larger scale than just a couple of low-throughput comparisons. As far as I know, this is common practice when presenting a new approach like this, and it would (1) allow for a better idea of the error (which, in my opinion, is calculated in a highly optimistic way), and (2) allow for a much better quantification of the results. E.g., is the difference of 56% vs. 41% of mutations above the detection limit (in subsection “E. Differing effects of mutations in CDR1H and CDR3H”, first paragraph) a large one, or maybe not even distinguishable given the accuracy of the experiment? Even if the same library was used, a replicate of the subsequent steps would be highly informative, and, in my opinion, necessary for validation of the approach.

We thank the referee for this thoughtful suggestion. In retrospect, we completely agree that replicate experiments are needed to provide appropriate validation of assays such as this. Our revised manuscript describes the results of three independent replicate Tite-Seq experiments, which are used to estimate errors, and whose results are given in Figure 4—figure supplement 2. We expect that these replicate experiments, and the analysis thereof, will address the referee’s concern.

I do not see a striking biological result from the analysis. What occurs to me as the main result of the paper (the correlation between contacts of wt residue with sensitivity to mutations at that position) is not too surprising to me, given many studies that have shown protein stability to be an important determinant of its function (and, as I understand, the sensitivity is measured on an absolute scale). Other reported findings seem suspicious and vague, especially considering my main concern expressed above.

Our paper is primarily a methods paper directed at addressing a fundamental problem in deep mutational scanning (DMS) assays: how to separate out the sequence-dependence of ligand binding affinity from the other sequence-dependent effects, e.g. on expression level or on the fraction of expressed proteins that are properly folded. Our work is the first to solve this fundamental problem inherent to DMS assays. These results are important because DMS assays are being rapidly adopted in a wide variety of fields including protein science, immunology, virology, and evolution.

We have also clarified the biological importance of our specific findings. Namely, our results suggest (a) that secondary CDRs many serve to stabilize antibodies against destabilizing variation in CDR3H, and (b) that binding affinity and specificity might be controlled by “sectors” within the antibody structure. Because our paper is primarily about a method, we do not follow up on these hypotheses, and so they remain preliminary. Still, this illustrates the kinds of questions that Tite-Seq can help address.

As I understand, 1850 mutations per region were surveyed. Such a high number introduces a lot of noise due to sampling by FACS and sequencing, even if the initial library is large and evenly distributed (and especially if the number of sequences is not large for some data points, cf. Figure 2—figure supplement 2B). I may have missed this, but is the distribution of the mutations in the initial library known, and what was the distribution of absolute reads per mutant at each data point?

We apologize for the lack of clarity on this issue. We did indeed sequence the unsorted library. Zipf plots illustrating the prevalence of each sequence in the library used for each replicate are now shown in Figure 4—figure supplement 3. Moreover, Figure 3 shows the number of cells sorted into each bin, as well as the number of reads obtained from each bin.

There is indeed a large variation in the number for reads from bin to bin, but our improved analysis method explicitly accounts for this variability. The variation in K values measured by Tite-Seq that we expect to result from the finite sampling of each sequence was estimated using simulations. The results of these simulation tests are shown in Figure 4—figure supplement 7.

Reviewer #2: The authors present Tite-Seq, a method that uses high-throughput DNA sequencing of yeast-displayed antibody libraries to assess mutant KD and surface expression at a large scale. In general, the experiments and analysis are well executed, and the paper is mostly well written. My enthusiasm is restrained for two main reasons. First, the authors fail to cite and discuss substantially similar work. Their elucidation of KD values for a large library of variants does represent an advance, but it's incremental. The work must be discussed in the context of what has come before. Second, several of the analyses are poorly described and, if I understood them correctly, will inflate the reader's confidence in the method. In particular, the method for estimating error for each KD value is insufficient. Both of these points are particularly important because the manuscript is focused on the method itself rather than an important/exciting application. I therefore do not recommend publication of the manuscript in its current form.

We thank the reviewer for this critique. We believe the revised manuscript addresses all of these concerns.

Our revised manuscript provides an expanded discussion of the importance of our work in the context of the prior literature, including the works cited by the referee. We argue that our work represents an important conceptual and technological advance over previously described DMS experiments.

Specifically, all DMS experiments face an important challenge: how to separate the effect that protein sequence has on one specific biochemical property of interest from the effect that it has on other biochemical properties of the protein that can affect measurements in a DMS assay. In our case, the challenge is to separate the effect that protein sequence has on ligand binding energy from the effect that protein sequence has on protein expression and on the fraction of this expressed protein that is properly folded. Our paper solves this problem in the context of antibody-antigen interactions, and appears to be the first in the literature to do so for any protein-ligand binding energy. More generally, we provide a template for how such distinctions between sequence-function relationships can be made. The Discussion has been revised to emphasize this important point.

We agree with the referee’s criticism regarding the error bars on the Tite-Seq K measurements. In response, we have performed the Tite-Seq experiment in triplicate, and now use these triplicate measurements to assess error bars. Our Tite- Seq measurements are also validated by an increased number of K measurements on individual clones sampled from the library.

Specific comments:

Introduction – The authors should be more thorough in discussing the strengths and weaknesses of prior deep mutational scanning work. In particular, the fact that yeast display has been coupled to deep mutational scanning and that affinity ranking of variants has been achieved is something readers should know before they get to the end of the Discussion (PMID 25311858). Mammalian antibody display has also been coupled to deep mutationals canning (PMID 23765106). Ribosome antibody display and deep mutational scanning using varying ligand concentrations has also been done (PMID 23103372). Of course, the existence of a substantially similar paper (PMID 26296891) should also be acknowledged and the work discussed in that context.

We have expanded the Introduction to provide a more detailed discussion of the prior literature, particularly regarding prior DMS experiments using yeast display, prior work on antibody sequence-affinity landscapes, and prior work at varying ligand concentrations. The revised text now explicitly explains how our work differs substantially from these previous papers, including the papers cited by the referee above. We believe this more detailed discussion greatly clarifies why our work provides a major advance over the prior literature.

We have moved our discussion of Reich et al., 2015 (PMID 25311858), which describes the yeast-display DMS method SORTCERY, to the Introduction. We have also revised the Introduction to emphasize that both yeast display (Reich et al., 2015, PMID 25311858; Kowalsky et al., 2015, PMID 26296891) and the conceptually similar method of mammalian cell display (Forsyth et al., 2013, PMID 23765106) have indeed already been used to perform DMS experiments. We emphasize that our key advance in this regard is not the use of a cellular display system to do DMS experiments, but rather the ability to measure binding titration curves. None of these works measures binding titration curves. Therefore, all of the DMS-measured affinities reported in these three works are vulnerable to being convolved with sequence-dependent effects on expression or protein stability. Our paper shows how to overcome this common problem.

The referee is correct that Fujino et al., 2012, PMID 23103372 performed Ribosome-display-based DMS experiments at multiple concentrations. The revised text points this out. The multiple concentration measurements used by Fujino et al., however, were not used to infer binding titration curves nor to make estimates of K; they were used only to identify codons to randomize in a combinatorial library. The only K values reported in this paper were measured in a low-throughput manner, either by SPR or KinExA.

Subsection “C. Low-throughput validation experiments” – The authors show low-throughput validation data for three clones and compare this data to Tite-Seq results. In the subsection “A. Overview of Tite-Seq”, the authors make the point that, if curve fitting is to be successful, the K

We chose our initial range of concentrations to target the K of the bulk of sequences in our library. It does make sense, however, to include the WT K. The triplicate experiments reported in our revised manuscript have therefore been performed with a range of antigen concentrations (10-9.5 M to 10-5 M) that is one decade lower than the range used for the previous manuscript (10-8.5 M to 10-4 M). We also modified the yeast display protocol to improve the fraction of yeast that express antibody on their surface, and this has further improved the precision of both Tite-Seq and our low-throughput flow cytometry measurements.

From our new experiments we find a WT Kof 1.9 nM using Tite-Seq and 2.4 nM using low-throughput flow cytometry. This is now consistent with the measurements from Dane Wittrup’s laboratory. We note that, although Boder et al. (2000, PMID 10984501) report a WT Kof 0.4-0.7 nM, later work (Gai and Wittrup, 2007, PMID 17870469) revised this Kupwards to 1.2 nM, which is within a factor of 2 of both our low-throughput and high-throughput measurements. The difference in salt concentration used in our experiments may account for some of the remaining discrepancy.

Subsection “C. Low-throughput validation experiments” – Pursuant to the previous comment, the authors' simulations (

The revised version of Figure 1 now shows results simulated using Kvalues that were measured. We emphasize in the revised caption that this figure is only meant to provide a schematic illustration of the inference method, and that the method for inferring Kvalues from real read counts is substantially more involved. Figure 4—figure supplements 6 and 7 illustrate more realistic simulations as well as the results that our real analysis pipeline produces from these data.

Subsection “C. Low-throughput validation experiments” – The low-throughput validation experiments do not themselves look particularly robust. The K

We have modified the yeast display protocol, substantially increasing the fraction of yeast cells that display antibody. This, in turn, has substantially increased the precision of our low-throughput measurements. The low-throughput titration curves measured using this modified protocol are described in the revised manuscript.

We have also quantified the uncertainty in all high-throughput and low-throughput Kmeasurements. High-throughput measurements, in triplicate, with estimated error bars, are provided as the supplemental data files. A new Supplementary file I shows both high-throughput and low-throughput Kmeasurements for each of the clones plotted in Figure 4A.

The Figure 2 legend has been revised to clarify the meaning of panel 2D.

The revised text now provides clone-by-clone estimates of Kvalues and their associated uncertainties.

The revised text reports the uncertainty of each Tite-Seq-measured Kvalue using results from three replicate Tite-Seq experiments.

Subsection “D. Tite-Seq can measure dissociation constants” – The authors state: "We note that…measurements for the WT scFV are about a factor of 10 larger than the previously measured value of K

The reviewer’s intuition seems to have been correct. The improvements to our yeast display assay, as well as a shift downward in the antigen concentration range used, have substantially reduced the discrepancy between our measured WT Kvalues (1.9 nM from Tite-Seq and 2.4 nM from flow cytometry) and the previous results reported by Dane Wittrup’s group (1.2 nM, as reported by Gai and Wittrup, 2010).

The new Figure 5—figure supplement 1 show density estimates for the Kand expression levels of single-point CDR1 and CDR3 mutants. We believe that this more clearly illustrates the typical effect that mutations in these two regions have on affinity and expression.

Appendix C – The authors say they used custom microarray nucleotides to generate the library, but the nature of the sequences on the array is unclear. A supplementary data file containing the sequences ordered should be included.

A listing of the sequences in our microarray-synthesized CDR1H and CDR3H libraries is now provided online, along with our analysis code and preprocessed data.

Appendix F – The authors used the 10% lowest fluorescence values to estimate autofluorescence/background. This is a curious choice, if the library contained mutations to stop codons (see comment above). Stop codons, particularly early ones, virtually guarantee loss of function.

The revised manuscript uses a more principled analysis approach, which is detailed in Appendix F. In particular, the mean fluorescence of cells at 0M antigen is used to estimate autofluorescence.

[Editors' note: the author responses to the re-review follow.]

Specific points:

1) Please make sure that your deep sequencing data is available on the SRA and relevant computer code is available as supporting file or on a publicly accessible repository. If this is already the case, please clearly indicate in the manuscript where these can be found.

As described in the Methods, raw data has been deposited on the SRA (BioProject ID PRJNA344711), and both processed data and scripts have been posted at github.com/jbkinney/16_titeseq.

2) There are a few additional papers that you should consider discussing in the context of prior work:

Doolan KM, Colby DW: Conformation-dependent epitopes recognized by prion protein antibodies probed using mutational scanning and deep sequencing. J. Mol. Biol. 2015, 427:328-340.

Van Blarcom T, Rossi A, Foletti D, Sundar P, Pitts S, Bee C, Melton Witt J, Melton Z, Hasa-Moreno A, Shaughnessy L, et al.: Precise and efficient antibody epitope determination through library design, yeast display and next-generation sequencing. J. Mol. Biol. 2015, 427:1513-1534.

We thank the reviewers and editor for this suggestion. These two papers, along with Kowalsky et al. JBC (2015) are now cited in the Introduction as showing how yeast-display-based DMS experiments can be used for mapping the antibody binding epitopes of proteins. We have also cited these papers in the Discussion as examples of the type of experiment that probably does not require full titration curves.

3) Your approach estimates fairly accurate K

This is a good point. The revised Discussion section addresses this matter, i.e., that many experiments (such as the epitope mapping studies discussed above) do not require quantitative Kmeasurements and that, in such cases, it will often make sense to use a simpler experimental design.

Also, do you have any comments on the trade-off between sequencing depth and the accuracy of the inferred K

We have performed simulations that explore the effect of sequencing depth on the precision of Kestimates (Figure 4—figure supplement 8). We have not, however, experimentally tested variations in sequencing depth, the number of sorting bins, or the number of different antigen concentrations. The number of bins we used for sorting, as well as the number of antigen concentrations, was chosen in large part for experimental convenience (e.g., the FACS instrument can sort into 4 bins simultaneously, and the number of sorted cells was chosen to enable a full sort to be completed in about 5 hours). These matters are mentioned in the revised Discussion. We also discuss the fact that the precision of our measured Kvalues appears to have been limited by the efficiency with which antibody sequences were recovered from sorted cells, and that improvements in the post-sort recovery of such sequences would probably improve the accuracy of Tite-Seq.

4) The relatively large unsigned error on two variants in

The revised Results section now mentions this discrepancy. We do not know the cause of this. The revised text, along with the new Figure 4—figure supplement 3 (see below) notes that the analysis of uncertainty using synonymous mutants reveals a high-variability region of affinity (K~1E-7 M) that coincides with these outliers.

5) Are the mutations outside of the dynamic range (K

Yes. This point is clarified in the revised Figure 4C caption, in the caption for Figure 4—figure supplement 2, and in the main text.

6) Please quantify the error in precision and accuracy of the method. In particular, the poorer correspondence between replicates in the binding range between 10-1000 nM K

We have included an additional supplemental figure, Figure 4—figure supplement 3, which shows the error estimates for both Kand E using synonymous mutations. These plots indicate substantially higher noise at ~1E-7 M, and this point is mentioned in the revised main text and figure caption.

7) Can you quantify error within replicates by inferring K

Yes, see the new Figure 4—figure supplement 3.

8) How big is the benefit of controlling for active surface-displayed protein? While it is clear that yeast surface expression and folded surface displayed protein varies between variants (Burns et al., 2014), it has been shown (yet still surprising!) that surface expression (and proper folding) effects are modest for yeast surface display experiments, particularly for residues that are reasonably surface-exposed (Kowalsky and Whitehead, 2016).

Burns, Michael L., et al. Directed evolution of brain-derived neurotrophic factor for improved folding and expression in Saccharomyces cerevisiae. Applied and environmental microbiology 2014, 80:5732-5742.

Kowalsky CA, Whitehead TA Determination of binding affinity upon mutation for type I dockerin-cohesin complexes from Clostridium thermocellum and Clostridium cellulolyticum using deep sequencing. PROTEINS 2016 84: 1914-1928

This is an important point and is now addressed at greater length in the Discussion section. While there has been work (e.g. Kowalsky, 2016) suggesting that the effect of mutations on the expression and specific activity of displayed proteins is modest, other work (Burns et al., 2014) finds that these effects can be quite large. In fact, the magnitude of these effects is likely to vary substantially from protein to protein in a largely unpredictable manner. Our main point, which is now clarified, is that one does not know a priori if these contaminating effects will be present, and that one can guard against them only by assaying full binding curves.

Table 2.

Primers. Oligonucleotide sequences are written 5 to 3. Bold sequences indicate variable regions. The ‘1H library’ and ‘3H library’ primers respectively contained the 1H and 3H variable regions (bold) analyzed in this paper. These primer libraries were synthesized by LC Biosciences using microarray-based DNA synthesis. All other primers were ordered from Integrated DNA Technologies. The ‘[XX]’ portion of L1AF_XX and L1AR_XX indicates the location of each of 64 different barcodes (i.e., XX = 01, 02, , 64), which ranged in length from 7 bp to 10 bp and which differed from each other by at least two substitution mutations.

DOI: http://dx.doi.org/10.7554/eLife.23156.019

Name	Sequence
1H library	GTGTTGCCTCTGGATTCACTTTTAGTGACTACTGGATGAACTGGGTCCGCCAGTCTCCAGA
3H library	GTGACTGAGGTTCCTTGACCCCAGTAGTCCATACCATAGTAAGAACCCGTACAGTAATAGATACCCAT
oRAL10	TTCTGAGGAGACGGTGACTGAGGTTCCTTG
oRAR10	TGAAGACATGGGTATCTATTACTGTACG
oRAL11	CAGTCCTTTCTCTGGAGACTGGCG
oRAR11	ATGAAACTCTCCTGTGTTGCCTCTGGATTC
3H1F	TTCTGAGGAGACGGTGACT
3H2R	TGAAGACATGGGTATCTATTACTGTAC
1H2F	CAGTCCTTTCTCTGGAGACTG
1H1R	ATGAAACTCTCCTGTGTTGCCT
oRA10	GCATATCTAAGGTCTCGTTCTGAGGAGACGGTGAC
oRA11	GCCGATTGTTGGTCTCCATGAAACTCTCCTGTGTTGC
PE1v3ext	AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACG
PE2v3	AAGCAGAAGACGGCATACGAGATCGGTCTCGGCATTCCTGCT
L1AF_XX	ACACTCTTTCCCTACACGACGCTCTTCCGATCT[XX]AGTCTTCTTCAGAAATAAGC
L1AR_XX	CTCGGCATTCCTGCTGAACCGCTCTTCCGATCT[XX]GCTTGGTGCAACCTG