Novel search method for the discovery of functional relationships.

MOTIVATION: Numerous annotations are available that functionally characterize genes and proteins with regard to molecular process, cellular localization, tissue expression, protein domain composition, protein interaction, disease association and other properties. Searching this steadily growing amount of information can lead to the discovery of new biological relationships between genes and proteins. To facilitate the searches, methods are required that measure the annotation similarity of genes and proteins. However, most current similarity methods are focused only on annotations from the Gene Ontology (GO) and do not take other annotation sources into account.
RESULTS: We introduce the new method BioSim that incorporates multiple sources of annotations to quantify the functional similarity of genes and proteins. We compared the performance of our method with four other well-known methods adapted to use multiple annotation sources. We evaluated the methods by searching for known functional relationships using annotations based only on GO or on our large data warehouse BioMyn. This warehouse integrates many diverse annotation sources of human genes and proteins. We observed that the search performance improved substantially for almost all methods when multiple annotation sources were included. In particular, our method outperformed the other methods in terms of recall and average precision.

1 INTRODUCTION

Similarity search plays an important role in biological, pharmaceutical and medical investigations. For instance, the introduction of the BLAST algorithm by Altschul ) to search for similar sequences has been considered a milestone in genomics (Bahcall, 2007). Other similarity search methods to mine databases of 3D molecule conformations have been important for drug discovery (Willett ). In addition, the growing availability of annotations that characterize genes and proteins (Reeves ) opens the new possibility to find biological relationships by similarity searches based on function, domain composition, disease association, tissue expression, etc. For example, the identification of similarly annotated genes and proteins can reveal new gene–disease associations (Aerts ), suggest novel protein functions (Friedberg, 2006) and indicate new drug targets (Chan ).

In general, similarity searches compute pairwise similarities of a query with the entities in a database to obtain a ranked list of high-scoring similarities. In particular, a number of methods have been proposed for the quantification of pairwise similarities of gene and protein annotations. Most of those functional similarity methods are based on Gene Ontology (GO) annotations (Benabderrahmane ; Chabalier ; del Pozo ; Lerman and Shakhnovich, 2007; Lord ; Mistry and Pavlidis, 2008; Pesquita ; Popescu ; Schlicker ; Sevilla ; Speer ). However, the last years have shown a dramatic growth in datasets that result from high-throughput experiments and computational work and yield annotation sources that provide manifold information about, for instance, protein interactions, signaling circuits, metabolic pathways, cellular localization, tissue expression, disease associations and protein domain architecture. Currently, only one similarity search method explicitly takes multiple annotation sources into account, namely, the kappa coefficient used by the DAVID Gene Functional Classification Tool (Huang ). In contrast, the integration of multiple annotation sources into a network structure is often applied in the context of gene function prediction (Huttenhower ; Wang and Marcotte, 2010; Warde-Farley ).

When developing efficient methods for searching through gene and protein annotation data, a particular task is the construction of data structures that represent the annotations. Most methods rely on the graph structure of GO to estimate quantitative semantic relationships among the gene/protein annotations (Pesquita ). However, the GO structure limits the inclusions of non-ontological (i.e. non-GO) annotations into methods. A flattened representation of the GO hierarchy solves this problem and stores the annotations as Boolean arrays in which the presence and absence of annotations is recorded (Huang ). This representation implicitly contains the ontological relations and allows the inclusion of non-ontological annotations as part of the array. This avoids the inference of relationships through the hierarchical structure of GO. GO-based similarity methods that use this data structure are COS (Chabalier ), simGIC (Pesquita ) and TO (Mistry and Pavlidis, 2008). Although these methods do not consider annotation sources other than GO, they achieve better performance than methods such as those by Resnik (1999) and Lin (1998) that depend on the GO graph structure.

In the following, we will introduce the new method BioSim for similarity searches based on diverse annotation sources of gene and protein function and extend the existing methods cosine similarity, kappa coefficient, simGIC and TO to utilize annotations not only from GO, but also from 22 major biological databases for human genes and proteins. We will also compare the performance of BioSim with the other methods in different benchmarks.

2 MATERIALS AND METHODS

2.1 Annotation sources

Twenty-two publicly available annotation sources for human genes and proteins were integrated into our data warehouse BioMyn. These include functional annotations from all three GO categories (MF, molecular function; BP, biological process; CC, cellular component) (Camon ) and from the UniProtKB controlled vocabulary of keywords (Consortium, 2010). The data warehouse also contains clusters of similar sequences from Ensembl protein families (Flicek ) and from UniRef90 (Suzek ); protein domain architectures from Pfam (Finn ) and InterPro (Hunter ); metabolic and signaling pathways from HumanCyc (Romero ), KEGG (Kanehisa ), and Reactome (Matthews ); protein–protein interactions and protein complexes from CORUM (Ruepp ), DIP (Salwinski ), HiMAP (Rhodes ), HPRD (Prasad ), IntAct (Kerrien ), MINT (Chatr-Aryamontri ), PDB (Berman ; Velankar ) and STRING (Jensen ); disease associations from OMIM (Amberger ); enzyme classifications from the Enzyme nomenclature database (Bairoch, 2000); gene expression data for different tissues and cell lines from the Novartis Gene Atlas (Su ); Mammalian Phenotype Ontology annotations of human genes as provided by the Mouse Genome Database (Blake ); and orthologs of protein sequences from OrthoMCL (Chen ).

From the annotation sources, the functionally relevant features associated with individual genes and proteins were extracted. In the following, we refer to these features as annotation terms, which correspond, for example, to the specific molecular function (e.g. oxidoreductase activity) or domain (e.g. SH2) or pathway (e.g. glycolysis) annotated to genes and proteins. The different gene and protein identifiers used in the annotation sources were unified by mapping them to Entrez Gene ID and UniProtKB accession numbers. In total, our data warehouse contains 24 586 human Entrez Gene entries and 70 767 human UniProtKB protein entries (including 20 177 manually reviewed proteins in UniProtKB release 15.5, see Supplementary Material for further details). To enable comparisons between functional similarity methods using multiple annotation sources and those using only GO annotations, proteins with no available GO annotation were excluded. This resulted in a list of 18 076 protein entries out of 20 177 manually reviewed proteins in UniProtKB release 15.5.

2.2 Functional similarity methods

In the following, A and A denote the sets of annotation terms associated with the gene products X and Y, respectively. Annotations available for genes are transferred to the encoded proteins.

BioSim: Our novel functional similarity method BioSim is defined as follows: Here, t∈{A∩A} is the set of annotation terms shared by X and Y, and p(t) is the probability that both A and A contain the same term t by chance. Since BioSim is the product of the probabilities p(t), a score of zero represents the highest similarity and a score of one the lowest. This is in contrast to other methods described below, except TO. The probability p(t) is estimated using the cumulative hypergeometric distribution: N is the number of proteins in our database, and N is the number of proteins annotated with term t. D is the sum of |A| and |A|, that is, the total number of annotation terms for X and Y. Therefore, the resulting probability p(t) depends not only on the frequency N, and thus on the specificity, of the annotation term t, but also on D. This is an important property of BioSim and accounts for the annotation bias of intensively studied genes and proteins. A pair of proteins associated with many annotations terms (large D) has an increased probability p(t) to share the annotation term t (i.e. a decreased functional similarity) in comparison to a pair of proteins associated with few annotations terms (small D).

Term overlap length (TO): TO represents the number of annotations terms shared by two proteins X and Y (Mistry and Pavlidis, 2008):

Kappa coefficient (KC): This method is used in the well-known DAVID Gene Functional Classification Tool (Huang ). It computes a normalized difference of the observed number of annotation terms O(X, Y) shared by two proteins X and Y, and the expected number E(X, Y) of shared annotation terms that are randomly chosen (Huang ). It is defined as follows: In the following, we describe the two methods simGIC and COS. Unlike the previous methods, both methods incorporate term weights based on the information content (IC) of a term t (Resnik, 1995): Here, N is the number of proteins annotated with term t and N the total number of proteins in our study.

simGIC: This method introduced in Pesquita ) includes the summed information contents of shared versus all annotated terms for two proteins X and Y:

Cosine similarity (COS): This classical method is defined as follows (Salton ): Here, and are the annotation vectors of two proteins X and Y, respectively. In each vector, the absence of an annotation term is represented by 0 and the presence by IC(t). This method was first used in the context of functional similarity by Chabalier ).

2.3 Evaluation methods

Gold standard: To evaluate the performance of the functional similarity methods, we collected a gold standard dataset composed of groups of proteins that are assumed to be functionally related (to a certain extent) and contained in the list of 18 076 proteins with at least one available GO annotation (as described above). The protein groups in the dataset were obtained from four benchmark categories that we limited to at most 400 groups per category: (i) 400 groups from protein complexes (selected randomly from a total of 2030 complexes in CORUM); (ii) 88 groups from sequence clusters of related protein sequences based on UniRef90 clusters (sequences of at least 90% identity) and thus with putatively similar functions; (iii) 355 groups from reliable protein–protein interactions (here, an interaction is regarded as reliable if it is reported in at least three different publications); and (iv) 400 groups from metabolic and signaling pathways (selected randomly from a total of 424 pathways in KEGG and Reactome). Groups of more than 20 proteins were excluded as being too general. The average group size was 6.7 proteins and the overall standard deviation 4.3. In total, the gold standard consisted of 1243 groups that covered 8150 proteins overall (some proteins were shared by different groups). In the following, we will refer to those groups as validation groups.

Benchmarking procedures: From each validation group, a query protein was randomly selected and the remaining group members were regarded as gold standard positives. To obtain ranked lists, pairwise functional similarity scores were computed between the query protein and all other 18 076 protein entries used in our study. The same evaluations were carried out using either only GO annotations or all aforementioned annotation sources (excluding the respective annotation source of the benchmark category). For baseline comparison, a dataset of 10 000 protein pairs was randomly created. To compute a background distribution of BLAST bit scores (NCBI blastp version 2.2.22), 100 000 protein pairs were randomly drawn from the list of studied proteins. Since the bit score of a protein pair is not symmetric, the average bit score of the pair was used (Pesquita ).

Performance measures: The recall at a rank k is the number of positives in the k top ranks of the computed ranking list divided by the total number of positives, i.e. the members of the respective validation group. The average precision is the mean of the precisions obtained for the ranks of all positives in the ranking list (Buckley and Voorhees, 2000). For example, in case of three positives found at ranks 2, 5 and 10, the average precision would be (1/2+2/5+3/10)/3=0.4. The Precision at a rank k is the number of positives in the k top ranks divided by k. The first relevant rank (FRR) is the best rank of a positive in some ranking list.

Score cut-offs for the functional similarity methods: Using the ranking lists obtained for each validation group, we identified the functional similarity score that yielded 50 false positives. This number is a reasonable threshold suggested by Gribskov and Robinson (1996) for their ROC50 method. By averaging these functional similarity scores, suitable score cut-offs were obtained for every similarity method. We refer to these score cut-offs as SC50. The performance curves were generated using the ROCR package (Sing ).

3 RESULTS AND DISCUSSION

3.1 Evaluating the performance of functional similarity methods

The performance of BioSim in identifying known functional similarities was compared with that of four other methods: TO, KC, simGIC and COS. The results were averaged over all validation groups. While all methods showed similar performance when using only GO annotations, the performance was improved when considering multiple annotation sources (Fig. 1). Notably, BioSim achieved better performance than the other methods. For instance, the top 20 hits of BioSim had an average recall of 0.58. The second best method, COS, had an average recall of 0.44 (Fig. 1A). The average precision of BioSim was 0.39, which was significantly higher than that of the other methods (P<0.01, Wilcoxon signed-rank test). Likewise, BioSim had a median value of 2 for the FRR, surpassing the other methods (Table 1).

Fig. 1.

Performance of functional similarity methods. Average recall is plotted for different top ranks k using either multiple annotations sources (A) or only GO annotations (B). The average values were obtained from benchmarking with 1243 validation groups. See Supplementary Fig. S3 for details on the performance of the methods in each of the four benchmark categories.

Table 1.

Performance comparison of functional similarity methods using multiple annotation sources versus using only GO annotations, over all 1243 validation groups

Method	Multiple sources		Only GO
	Avg. precision	FRR	Avg. precision	FRR
BioSim	0.39	2	0.22	7
COS	0.28	3	0.22	7
KC	0.21	5	0.20	5
simGIC	0.28	3	0.22	5
TO	0.24	3	0.17	11

See Supplementary Table S1 for details on the performance of the methods in each of the four benchmark categories. avg. precision: average precision.

The overall performance of the methods varied for each benchmark category. It was lower for all methods when using the protein–protein interaction category and higher when using the sequence cluster category (Supplementary Fig. S3A and Table S1). The combined average recall for all methods was more than one-third lower in the protein–protein interactions category than in the sequence cluster category (the respective recalls were 0.29 and 0.75). The observed high performance when using the sequence clusters category is due to the tendency of the methods to rank similar sequences at the top. This can be explained, to some extent, by annotation transfer between homologous protein sequences, by gene annotations that are transferred to all encoded proteins and by domain annotations that are almost identical for similar sequences. Therefore, the tendency to rank similar sequences at the top reduces the performance of the methods when using benchmark categories different from sequence clusters because gold standard positives are displaced to lower ranks.

3.2 Including multiple annotation sources improves performance

The use of multiple annotation sources improved the performance of four of the five methods although they were not originally developed to handle multiple annotations (in contrast to BioSim). Much of this increase seems to be attributable to the availability of more annotation terms per protein. The number of terms annotated to each protein increased from a median of 7.5 GO terms to a median of 15.0 annotation terms when all annotation sources were included (Supplementary Figs S4A and S5A). The TO method, which counts the number of common terms, but does not account for term specificity, improved its average precision from 0.17 to 0.24 when all annotations were used.

Notably, the use of multiple annotation sources does not only increase the number of annotation terms per protein, but also improves the specificity of the annotations. While GO terms annotated to at most four proteins were available for 8096 proteins, this number doubled to 16 649 proteins in case of multiple annotation sources when not only using GO (Supplementary Figs S4B and S5B). The positive effect of the increased annotation specificity on the performance can be observed with the three functional similarity methods COS, simGIC and BioSim. All three methods weight annotation terms and showed the strongest performance improvement when multiple annotation sources were included.

In particular, BioSim was best able to take advantage of the increased number and improved specificity of annotations terms, as shown by the near doubling of its average precision (Table 1). In the case of BioSim, as explained in Section 2, the functional similarity between two proteins increases if both are annotated with specific terms (terms that are annotated to few proteins) because the corresponding probabilities of the terms are low. Additionally, since BioSim computes the product of the probabilities of all terms shared by two proteins, a certain number of even less specific terms still increases the overall functional similarity. Annotations from protein–protein interactions, sequence clusters, pathways and disease associations are normally the most specific and least abundant ones, annotated to no more than a hundred proteins. In contrast, annotations as from cellular localization and tissue expression frequently cover thousands of proteins; and annotations from GO, UniProtKB keywords and protein domains span the whole range from just a few proteins to thousands (Supplementary Fig. S2).

As an example, we looked in detail at one known SNARE protein complex formed by the proteins VAMP2, SNAP25, STX1a and CPLX1. These four proteins are involved in the fusion of neurotransmitter-containing vesicles with the pre-synaptic membrane (McMahon ). When BioSim was applied using multiple annotation sources to compute the functional similarity of VAMP2 with each of the 18 076 human proteins in our study, SNAP25 achieved the top rank 1 with the strongest functional similarity. The other two complex members STX1a and CPLX1 were found at ranks 3 and 5, respectively. At rank 2 we found PRKD3, a protein that interacts directly with VAMP2, and at rank 4 we found VAMP1 who shares the Synaptobrevin domain with VAMP2. In contrast, when BioSim made use of only GO annotations, the rankings of SNAP25, STX1a and CPLX1 decreased to 25, 187 and 805, respectively. Specific annotations, which led to the identification of SNAP25 as functionally similar to VAMP2, included both four experimental results that reported the interaction between VAMP2 and STXa1 and several shared pathways in Reactome such as the proteolytic cleavage of SNARE complex proteins. Less specific annotations were a shared coiled-coil domain and a similar tissue expression profile. When only GO annotations were taken into account, ICA69 was the protein functionally most similar to VAMP2, primarily, because both proteins are annotated with the term secretory granule membrane. This term covers only 25 other proteins, none of which is SNAP25, STX1a or CPLX1. The current knowledge about ICA69 is very limited. It might play a functional role in the transport regulation of insulin secretory granule proteins (Buffa ) as well as in neurotransmitter transport as inferred by sequence similarity in UniProtKB. However, ICA69 has not been associated with the fusion of pre-synaptic vesicles.

In general, although GO annotations are expected to improve over time as more information is added, the use of other annotation sources helps to bridge the time until new data is incorporated. Furthermore, useful annotations to derive functional similarities such as protein–protein interactions and disease associations are not part of GO. Moreover, the use of multiple annotation sources can also reduce the impact of incorrect annotations found in biological databases (Schnoes ).

3.3 BioSim scoring versus other methods

BioSim distinguished functional relationships of gold standard positives from those of randomly paired proteins better than the other methods. Gold standard positives consistently received a BioSim score close to 0, while random pairs obtained a score close to 1 (Fig. 2A). In particular, we plotted precision and recall averages from our benchmark results for every method at different score cut-offs (Fig. 2B). We also computed a score cut-off (SC50) that resulted in 50 false negatives on average. The obtained SC50 score cut-offs, along with the score range of each method from lowest to highest functional similarity, were as follows: BioSim: ≤1.18×10−9 (range [1; 0]), TO: ≥115 (range [0; ∞)), KC: ≥0.360 (range [0; 1]), simGIC: ≥0.096 (range [0; 1]), and COS: ≥0.101 (range [0; 1]). For COS and simGIC, the second and third best methods, the SC50 score cut-offs were very close to zero, their non-similarity score; the recall at the respective SC50 cut-off had a median of 0.50 and a distribution covering the whole range (Fig. 2B). In other words, for both methods, the same SC50 cut-off resulted in a different recall. The KC and TO methods had a recall median below 0.5 for their respective SC50 score cut-offs. In comparison, the recall for BioSim at the SC50 score cut-off had the highest median (0.82) and the corresponding distribution concentrated around high values.

Fig. 2.

Comparison of functional similarity methods. (A) Histograms of the functional similarity scores that were obtained for 6907 pairs of gold standard positives and for 10 000 random pairs. (B) Precision (straight lines) and recalls (dashed lines) are averaged at different cut-offs. The vertical red lines highlight the SC50 score cut-offs that yield, on average, 50 false positives. The box plot to the left of the y-axis shows the distribution of recalls at this cut-off. BioSim scores are in logarithmic scale for better visualization. (C) Functional similarity and sequence similarity scores are compared based on 100 000 random pairs of proteins. Sequence similarity is measured as ln(bit score). Green lines depict the average functional similarity. Red lines illustrate the standard deviation. In each plot, the background contains a scatter plot where darker colors indicate a higher density of dots.

The limited consistency of the scores of COS, KC, simGIC and TO is probably caused by annotation bias toward better studied molecules (Rhee ) as these methods appear to be best suited for unbiased data (Wang ). In our data warehouse, a handful of proteins have over thousand annotations, while the majority has less than 10 annotations. A similar pattern can be observed when considering only GO annotations (Supplementary Figs S4 and S5). About 16% of all proteins are annotated only with less specific terms such as the UniProtKB keyword ‘Receptor’ or the GO term ‘protein binding’. The functional similarity of any two proteins sharing such terms is overestimated by the COS, KC and simGIC methods, which yield the highest score of 1. This misleading result is undistinguishable from a genuine functional similarity based on several shared annotation terms.

Furthermore, the same methods tend to underestimate the genuine similarity of any two proteins that are annotated with numerous terms and do not share a large proportion of their annotation terms. For example, the cellular tumor antigen TP53 (with 1642 annotation terms including 332 literature-curated protein interactions) shares about 19% of its annotation terms with the closely related E3 ubiquitin-protein ligase MDM2, which is known to bind and inhibit TP53 (Vassilev ). Relevant terms indicate common metabolic and signaling pathways, disease associations and protein interactions. However, the remaining 81% of TP53 annotation terms that are not shared with MDM2 lead to the following low functional similarity scores: COS=0.097, KC=0.120 and simGIC=0.056. These functional similarity scores are even below the SC50 cut-offs for the respective methods. This means that low functional similarity scores are often obtained for truly functionally related proteins. Such low similarity scores are also obtained when only GO annotations are considered: COS=0.206, KC=0.379, simGIC=0.142.

The TO method, which is simply the count of annotation terms shared by two proteins, avoids some of the described shortcomings by focusing only on the shared annotations. However, it cannot distinguish those annotations that occur by accident because it judges an event of two proteins sharing a rather unspecific, frequent annotation term (e.g. ‘protein binding’) as likely as an event of two proteins sharing a very specific, rare annotation term (e.g. ‘actin filament binding’).

3.4 Comparing functional similarity with sequence similarity

The correlation between the functional similarity of two proteins and their sequence similarity is often used to evaluate functional similarity methods (Lord ; Pesquita ). In our results, rank correlations for all methods were close to 0.1 when comparing BLAST bit scores and functional similarity scores for 100 000 random pairs of proteins. This low correlation is likely due to many protein pairs with almost no sequence similarity, but some functional similarity (Fig. 2C). To filter out protein pairs with low sequence similarity, we discarded all pairs having a ln(bit score) below 3.3. This threshold was chosen after observing that, for all methods, the averaged functional similarity scores increases above this value. In total, 631 (0.63%) of the random pairs had a ln(bit score) of at least 3.3. The rank correlations for these pairs were COS: 0.77, KC: 0.67, BioSim: 0.69, simGIC: 0.73, TO: 0.48.

Since BioSim showed a slightly lower correlation than COS and simGIC, we additionally analyzed some interesting cases manually. Supplementary Table S2 summarizes the manual inspection of annotations shared by the 15 pairs of proteins with the highest sequence similarity bit score. Seven protein pairs do not share specific annotation terms to infer a clear functional relationship. Accordingly, the low BioSim scores of those pairs are above the previously determined SC50 score cut-off of 1.18×10−9, which indicates a weak functional similarity. In contrast, a true functional relationship between the remaining eight protein pairs is more evident due to several shared specific annotations terms. This agrees well with BioSim scores below or very close to the SC50 cut-off, which suggests a considerable certainty of a real functional similarity. However, in contrast to BioSim, the scores from the other methods do not allow a clear-cut distinction in those cases as explained in the preceding Section 3.3. For example, the 2nd and 15th rows in Supplementary Table S2 are cases of low functional similarity scores for COS, KC and simGIC in contrast to BioSim although the respective proteins share numerous annotations. This suggests that a meaningful comparison of scoring methods based on the correlation of functional similarity and sequence similarity is limited by the available annotation datasets and their overall characteristics and quality, which can also be affected by annotation bias and incompleteness. Since BioSim is particularly designed to be more sensitive to the number and specificity of annotation terms in contrast to the other methods, its overall performance depends more on the annotation datasets and the individual annotation terms.

3.5 Finding disease-associated genes

Genes associated with the same disease phenotype tend to be functionally related (Schlicker ; Vidal ). Using BioSim, we ranked genes based on their functional similarity to genes known to be associated with a particular OMIM disease phenotype (Amberger ). To this end, for each gene not associated with a disease phenotype, we averaged the computed scores of its functional similarity to the previously known disease genes. The functional similarity scores were computed using a snapshot of our data warehouse that contained only gene annotations from before January 1, 2009. We then compared our results with an updated version of OMIM from October 31, 2009. This update contained 54 new gene associations for 46 diseases. In our results, 11 of the new genes were found at the top four ranks and 12 others between ranks 6 and 54 (Table 2 and Supplementary Tables S3–S26). The median rank of the new genes was 9.5. This is a drastic improvement due to the use of multiple annotation sources in contrast to the ranks obtained when using only GO annotations with a resultant median of 133.5.

Table 2.

Disease genes recently added to OMIM and identified by the BioSim method

Phenotype	# genes	gene New	Gene description	Rank	GO rank	Shared annotations
Familial glioma of brain	7	BRCA2	Breast cancer 2, early onset	1	102	Direct and indirect PPIs; same disease, GO and pathway annotation
Epidermolytic palmoplantar keratoderma	2	KRT1	Keratin 1	2	26	Direct and indirect PPIs; same disease, domain and GO annotation
Antley–Bixler syndrome	1	FGFR1	Fibroblast growth factor receptor 1	2	1	Indirect PPI; same disease, domain, GO and pathway annotation
Cardiofaciocutaneous syndrome	3	MAP2K1	Mitogen-activated protein kinase kinase 1	2	16	Direct and indirect PPIs; same pathway annotation
Folate-sensitive neural tube defects	3	MTHFR	5,10-methylenetetrahydrofolate reductase (NADPH)	2	3	Indirect PPI; same GO and pathway annotation.
Obesity	17	POMC	Proopiomelanocortin	3	83	Direct and indirect PPIs; same GO, pathway and UniProtKB keyword annotation
Autosomal recessive deafness-1A	1	GJB6	Gap junction protein, beta 6, 30 kD	3	6	Same disease, domain and GO annotation
Autosomal idiopathic short stature	3	GHR	Growth hormone receptor	3	182	Direct PPI; same GO annotation
Hypogonadotropic hypogonadism	3	FGFR1	Fibroblast growth factor receptor 1	3	1183	Direct PPI; same GO and UniProtKB keyword annotation
Non-insulin-dependent diabetes mellitus	25	PPARG	Peroxisome proliferator-activated receptor gamma	4	31	Direct and indirect PPIs; same disease, domain and GO annotation
Susceptibility to atypical hemolytic uremic syndrome-1	2	CFI	Complement factor I	4	14	Indirect PPI; same GO, pathway and UniProtKB keyword annotation
Non-insulin-dependent diabetes mellitus	25	SLC2A4	Solute carrier family 2 (facilitated glucose transporter), member 4	6	424	Indirect PPI; same GO, pathway and UniProtKB keyword annotation

The table lists 12 new disease gene associations found between ranks 1 and 6. The table column ‘# genes’ gives the number of known genes associated with the disease phenotype before January 1, 2009. The column ‘New gene’ contains the symbol of the gene that was added to the phenotype between January and October 2009 and correctly identified by BioSim. The columns ‘Rank’ and ‘GO rank’ give the position of the new gene in the ranking list if all annotations were used or only GO, respectively. The column ‘Shared annotations’ contains a summary of the most specific annotation terms shared by the known genes and the new gene. The detailed list of shared annotations can be found in Supplementary Tables S3–S26. Gene symbols and descriptions correspond to the official nomenclature from HGNC (Seal ). Indirect PPI refer to all direct interaction partners of the same protein.

Figure 3 highlights two disease phenotypes: obesity, which had 17 associated genes known before January 2009, and familial glioma of brain, which had seven associated genes. The new gene POMC, which was added to the obesity phenotype in the updated version of OMIM, was found on the third rank. Annotations shared by POMC and the other known disease genes included protein–protein interactions (with AGRP, ENPP1, GHRL, MC3R and MC4R) and the annotation term ‘obesity’ from UniProtKB keywords, which covers POMC and 10 other obesity genes (Supplementary Table S9). The genes ranked first and second, LEP (leptin) and INS (insulin), are also related to obesity (Spiegelman and Flier, 2001) even if they are not among the genes of the specific obesity phenotype in OMIM.

Fig. 3.

Disease-associated genes and their 10 most functionally similar genes. Our BioSim method was used to identify related genes for obesity (left) and the familial glioma of brain (right). The black frames highlight the new genes POMC and BRCA2 found by using BioSim. The vertical axis alphabetically lists the previously known disease genes. The horizontal axis ranks the most similar genes from left (most similar) to right. The colors indicate the strength of the functional similarity scores between the respective genes as computed by BioSim; lower scores indicate stronger similarity, see depicted color bar.

BRCA2, the new gene included into the updated version of OMIM for the glioma of brain phenotype, achieved the first rank of genes functionally related to the disease. BRCA2 showed strong BioSim functional similarity to five of the seven previously known genes for glioma of brain. Some of the annotations shared by BRCA2 and the five disease genes are protein–protein interactions (with ERBB2, MSH2 and PTEN), the joint disease association of BRCA2 and DMBT1 to medulloblastoma as well as of BRCA2 and PTEN to prostate cancer in OMIM and a number of GO and pathway annotations (Supplementary Table S3).

4 CONCLUSION

We presented the novel method BioSim to compute and search for functional similarities of genes and proteins based on diverse annotations such as protein interactions, domain architectures, biological pathways and disease associations. BioSim was evaluated together with four other published methods. All methods are fast to compute and just depend on the number of available annotation terms; thus they can scale well to larger datasets.

In our benchmarks, the use of multiple annotation sources resulted in improved performance of most methods than the use of solely GO annotations. BioSim achieved the best performance by consistently ranking functionally related proteins among the top two out of over 18 000 human gene products. BioSim in contrast to other scoring methods might be particular useful for applications based on functional similarity when consistent scores are especially desirable, for example, for the quality assessment of protein–protein interactions (Ramírez ) and for the clustering of genes or proteins by function (Huang ). We also showed how BioSim can be applied to discover potential disease genes.