Deep Learning Promotes the Screening of Natural Products with Potential Microtubule Inhibition Activity.

Natural microtubule inhibitors, such as paclitaxel and ixabepilone, are key sources of novel medications, which have a considerable influence on anti-tumor chemotherapy. Natural product chemists have been encouraged to create novel methodologies for screening the new generation of microtubule inhibitors from the enormous natural product library. There have been major advancements in the use of artificial intelligence in medication discovery recently. Deep learning algorithms, in particular, have shown promise in terms of swiftly screening effective leads from huge compound libraries and producing novel compounds with desirable features. We used a deep neural network to search for potent β-microtubule inhibitors in natural goods. Eleutherobin, bruceine D (BD), and phorbol 12-myristate 13-acetate (PMA) are three highly effective natural compounds that have been found as β-microtubule inhibitors. In conclusion, this paper describes the use of deep learning to screen for effective β-microtubule inhibitors. This research also demonstrates the promising possibility of employing deep learning to develop drugs from natural products for a wider range of disorders.

Introduction

Natural products are always an important source of new drugs.[1] Many famous drugs have been developed from plants, microbial metabolites, and marine organisms. Natural products play a vital role in the discovery and development of drugs, which are particularly evident in anti-tumor drugs. At present, more than 60% of anti-tumor drugs are closely related to natural products.[2] However, precise and efficient characterization of their biological effects remains challenging as the number of newly discovered natural products exponentially increased.[3] Therefore, without incurring the unsustainable costs of simply scaling typical discovery processes in parallel, artificial intelligence-assisted natural drug design has also emerged.[4−6]

Recently, there has been a growing enthusiasm for using deep learning to advance drug discovery.[7−9] Deep learning has been successfully applied in compound property prediction,[10,11] de novo design,[12−14] lead discovery,[15] repurposing,[16−18] and synthetic design.[19] Deep learning models demonstrate significant improvements in rapidly screening potent leads from massive compounds in available compound libraries. More recently, deep learning models achieved encouraging results in identifying antibacterial compounds,[18] candidates against osteoclastogenesis,[20] repurposing candidates for COVID-19.[21]

As a major target for chemotherapy of solid tumors, β-tubulin is essential for the growth and metastasis of cancer cells.[22,23] Paclitaxel is a milestone natural drug that has been found to have a special therapeutic effect and action mechanism for breast cancer and ovarian cancer.[24] It stabilizes microtubule polymers and prevents their division, and chromosomes cannot achieve medium-term spindle configuration. This blocks the progression of mitosis, prolongs the activation of mitosis, triggers apoptosis, or reverses the G0 phase of the cell cycle without cell division. Paclitaxel has been one of the most successful anti-cancer drugs in the last 30 years.[25] Ixabepilone is a novel cytotoxic compound derivate produced by myxomycetes.[26] Similar to paclitaxel, it also inhibits tubulin depolymerization and shows strong anti-tumor activity in P-glycoprotein-expressing multidrug-resistant tumors. These star drugs from natural products have inspired natural product chemists to continue their research for potential molecules with better activity and fewer side effects. Considering the significant advances of deep learning in drug discovery, it is particularly interesting to utilize deep learning to develop a new generation of tubulin inhibitors from the vast natural product library. In this study, we aim to adopt deep learning approaches to screen natural product libraries for potent β-microtubule inhibitors. The overall flowchart of this study is illustrated in Figure .

Figure 1

Flowchart of the AI-assisted discovery in natural products. A comprehensive hit dataset and a non-hit dataset were used to train a deep learning model. The trained model was used to screen a selected nature product dataset. The compounds were ranked by hit probability. The ranked compounds were further filtered to obtain the candidate dataset. Expert evaluated the candidates, conducted biotests, and identified three highly potent natural products of eleutherobin, bruceine D, and phorbol 12-myristate 13-acetate.

We assembled a hit dataset of 637 known β-tubulin inhibitors and a non-hit dataset of 2932 molecules, including tyrosine kinase inhibitors, small molecular immuno-oncology compounds, and angiogenesis-related compounds. The hit and non-hit datasets were used to train a directed message passing neural network (DMPNN).[27,28] An additional group of 4247 compounds was retrieved from public natural product libraries to form the natural product dataset. The trained DMPNN combined with various molecular fingerprints were adopted to virtual screen the natural product dataset. The performance of DMPNN, and three enhanced DMPNN were evaluated and additional machine learning algorithms were compared. All candidates screened by DMPNN were ranked by hit probability and further filtered by hit probability, Lipinski’s rule of drug-likeness, and Tanimoto similarity. Potential hits were manually checked, and three natural products were found as potent β-microtubule inhibitors. Among them, eleutherobin was identified as β-tubulin polymerization inhibitor reported in previous studies,[29,30] and for the first time, bruceine D (BD) and phorbol 12-myristate 13-acetate (PMA) were identified as active β-microtubule inhibitors by experimental validation. Current work highlights the significant potential of applying deep learning-based virtual screening approaches in drug discovery from natural products.

Results and Discussion

Identification of Eleutherobin, BD, and PMA as β-Microtubule Inhibitors

We first used the compounds of the hit dataset and non-hit dataset to train the adopted DMPNN model. The performance of DMPNN models on an independent testing set were summarized in Table and Figure . Additionally, we also systematically evaluated other machine learning algorithms (see Supporting Information, Tables S5 and S6 and Figures S1 and S2). The trained model was used to screen the natural product dataset for potential hits of β-microtubule inhibitors. The screened compounds were later filtered according to hit probability, Lipinski’s rule of drug-likeness, and molecular similarity. The qualified compounds were manually evaluated and surveyed in literatures. Those candidate compounds were further investigated for anti-tumor activity and β-tubulin polymerization inhibition activity. Finally, three natural products, including eleutherobin, BD, and PMA, were identified as β-microtubule inhibitors (Figure ).

Table 1

Performance of DMPNN Models (2D Normalized Features, Morgan Fingerprint with Bit Vector Features, Morgan Fingerprint Count Features, No Appended Features) in an Independent Testing Dataset

DMPNN models	AUC	ACC	precision	recall
2D normalized features	0.9962	0.9600	0.9832	0.9360
Morgan fingerprint with bit vector features	0.9867	0.9160	0.9906	0.8400
Morgan fingerprint count features	0.9859	0.9120	0.9813	0.8400
No appended features	0.9915	0.9280	0.9908	0.8640

Figure 2

ROC diagrams of four DMPNN models (2D normalized features, Morgan fingerprint with bit vector features, Morgan fingerprint count features, no appended features) in an independent testing dataset.

Figure 3

The structures of the three potential inhibitors are eleutherobin, bruceine D, and phorbol 12-myristate 13-acetate, which are screened by the DL model. Eleutherobin has been confirmed the capability of inducing β-tubulin polymerization, which is similar to paclitaxel.29,30 The successful identification of eleutherobin demonstrated the effectiveness of our deep learning approach in screening for β-tubulin polymerization inhibitors.

BD and PMA Inhibit Microtubule Assembly in MDA-MB-231 Cells

It has been shown that BD and PMA elicit anti-proliferative effects in multiple types of cancer.[31−34] Firstly, we found that BD and PMA exhibited potent anti-proliferative activity (IC50 values of 10.8 and 10.6 μM, respectively) in a dose-dependent manner in MDA-MB-231 cells by MTT assay (Figure a). Accordingly, 10 BD and 10 μM PMA were used in subsequent experiments. Next, we performed an immunofluorescence assay to confirm whether BD and PMA could inhibit microtubule assembly. We observed β-tubulin was abnormally accumulated, and the fluorescence intensity was significantly reduced in MDA-MB-231 cells after treatment with BD or PMA (Figure b,c). These results indicate that BD and PMA may be potential β-microtubule inhibitors.

Figure 4

Results of MTT assay and images of immunofluorescence confocal microscopy for bruceine D, and phorbol 12-myristate 13-acetate. (a) MTT assays performed to measure the anti-proliferative potency of bruceine D and phorbol 12-myristate 13-acetate against MDA-MB-231 cell. (b, c) Immunofluorescence confocal microscopy images of MDA-MB-231 cells treated with 10 μM bruceine D and 10 μM phorbol 12-myristate 13-acetate for 24 h, respectively. The nuclei and microtubules have been labeled with DAPI and β-tubulin, respectively. Representative images with quantification of β-tubulin intensity were shown. Scale bar, 20 μm. ***P < 0.001. Statistical significance was compared with respective control groups.

BD and PMA Induce Cell Cycle Arrest in MDA-MB-231 Cells

Previous studies have shown that BD and PMA are potent inducers of cell cycle arrest.[31,35] Moreover, microtubules are essential in the mitosis process, and microtubule inhibitors could disturb the progress of the cell cycle.[36] Therefore, we verified whether BD and PMA could similarly induce cell cycle arrest in MDA-MB-231 cells. Accordingly, we found that BD induced S phase arrest and PMA induced G0/G1 phase arrest (Figure a,b). Furthermore, as the key cell cycle regulators, the expression of CDK1, CDK2, and cyclin E was inhibited (Figure c,d), which further confirmed the above results. These results suggest that BD and PMA induce cell cycle arrest, although their respective periods of action are different.

Figure 5

Cell cycle analysis after bruceine D and phorbol 12-myristate 13-acetate treatment and Western blot analysis after bruceine D and phorbol 12-myristate 13-acetate treatment. (a, b) MDA-MB-231 cell treated with 10 μM bruceine D and 10 μM phorbol 12-myristate 13-acetate for 24 h, respectively. Cell cycle analysis was performed with propidium iodide. (c, d) Western blot analysis of CDK1, CDK2, and cyclin E in MDA-MB-231 cell treated with 10 μM bruceine D and 10 μM phorbol 12-myristate 13-acetate for 24 h, respectively. Relative CDK1, CDK2, and cyclin E expression levels were quantified by normalization to β-actin. **P < 0.01, ***P < 0.001. Statistical significance compared with respective control groups.

BD and PMA Induce Apoptosis in MDA-MB-231 Cells

To ascertain whether apoptosis was associated with the anti-tumor effects of BD and PMA, we evaluated their apoptotic ratio in MDA-MB-231 cells using annexin V/PI double staining. The results showed that a significant increase in early apoptotic cells in the presence of BD and PMA. Interestingly, PMA induced a higher rate of apoptosis compared with BD, indicating that PMA could elicit more obvious apoptosis (Figure a,b). Additionally, we detected the expression of apoptosis-related proteins such as Bax, Bcl-2, and caspase-3 in MDA-MB-231 cells, which suggested the activation of the apoptotic pathway (Figure c,d). Taken together, the above results demonstrate that BD and PMA elicit anti-proliferative effects via inducing apoptosis in MDA-MB-231 cells.

Figure 6

Apoptosis analysis of MDA-MB-231cells treated with bruceine D and phorbol 12-myristate 13-acetate and Western blot analysis. (a, b) MDA-MB-231 cells were treated with 10 μM bruceine D and 10 μM phorbol 12-myristate 13-acetate for 24 h, respectively. Apoptosis ratios were determined by flow cytometry analysis of annexin V/PI double staining. Representative images and quantification of apoptosis were shown. (c, d) Western blot analysis of caspase 3, cleaved caspase 3, Bax and Bcl-2 in MDA-MB-231 cell treated with 10 μM bruceine D and 10 μM phorbol 12-myristate 13-acetate for 24 h, respectively. Relative Bax and Bcl-2 expression levels were quantified by normalization to β-actin. **P < 0.01, ***P < 0.001. Statistical significance was compared with respective control groups.

Conclusions

In conclusion, we systematically developed a deep learning framework to screen natural products for potential β-microtubule inhibitors. In the obtained hits, eleutherobin was found in agreement with previous reports.[29,30] Another two compounds, BD and PMA were confirmed by experimental validation, demonstrating potential β-microtubule inhibition activity.

In addition, there is still room in the present work for further improvement with future efforts. Firstly, the datasets could be further expanded to include more molecules in the hit and non-hit datasets to train the deep learning model, as well as more candidates in the natural product dataset. Ideally, more molecules of diversified chemical structures would enable the model to further explore the chemical space, hence increasing the chances of discovering new valid compounds. With the availability of libraries of massive compounds and natural products, the pipeline proposed in this work can be easily applied to these libraries for better model training and screening. Secondly, the model can be pre-trained using other massive molecules before being trained by the specifically assembled hit and non-hit datasets. This approach could mitigate the cold start issues in deep learning. Thirdly, this work adopted the DMPNN model, which is not a generative model. It’s possible and easy to adopt generative models to replace the DMPNN model in our pipeline. Unlike the DMPNN, which is a non-generative model, generative models can be trained to learn the latent representation of molecules, which allows these models to screen molecules for desired properties. Moreover, using generative models could generate new molecular structures beyond the explored chemical space.

Overall, the results of this study demonstrate that our deep learning-based virtual screening pipeline could successfully identify three natural compounds as highly potent β-microtubule inhibitors. This work shows the encouraging potential of applying deep learning approaches in drug discovery, especially from abundant natural products.

Experimental Section

Hit Dataset Preparation

We systematically searched compounds databases for hits of β-microtubule inhibitors. In Selleck (a commercial database, selleckchem.com), we identified 22 microtubule-associated compounds. Furthermore, we surveyed published literature for FDA-approved drugs and compounds entering clinical trials and identified four compounds, including paclitaxel.[37−40] In addition to these 26 compounds entering clinical trials, we also considered compounds from pre-clinical trials which demonstrated as potential β-microtubule inhibitors. Specifically, in databases of ChEMBL, PDB, and ZINC15, we identified 611 active compounds with potential capability. In total, 637 compounds were included to assemble the hit dataset (see Supporting Information, Table S1).

Non-Hit Dataset Preparation

In order to establish the non-hit dataset, we searched ChEMBL, PDB, and Selleck databases for FDA-approved drugs and activate compounds that were not reported as active β-microtubule inhibitors. As a result, we collected 2932 compounds to assemble the non-hit compounds (see Supporting Information, Table S2).

Natural Products Dataset Preparation

We manually searched compounds of natural products from Selleck and Topscience databases. In total, 4247 compounds were selected, including sesquiterpene, diterpenoid, and those alkaloids which our group investigated in previous studies. Therefore, these compounds represent diverse chemical structures and bioactivities. The selected compounds formed the natural product dataset for later screening (see Supporting Information, Table S3).

Model Implementation

To utilize the information of molecular structures, a message-passing neural network (MPNN) framework was used in this study, namely, the directed message passing neural network (DMPNN)[27] was adopted. In DMPNN, properties of atoms and bonds are encoded as feature vectors with which multiple rounds of message passing operations are conducted over the molecular graph. In each round of message passing, the feature vectors were updated by aggregating messages from neighbors. After certain rounds of convolutional embeddings, the global descriptor in the form of a feature vector was obtained for the given molecule, with which molecular properties could be analyzed and predicted using conventional machine learning approaches. Thanks to the impactful success in search potent antibiotics, DMPNN has attracted significant attention with emerging applications of chemical property prediction, drug discovery, and structure characterization analysis.[41−43]

Extended on the learned descriptor for the input molecule by the DMPNN, we appended additional molecular fingerprints obtained using RDKit (http://rdkit.org/), namely the binary Morgan fingerprints, count-based Morgan fingerprints, and RDKit 2D normalized fingerprints to further enhance the information on the DMPNN descriptor. Using the basic DMPNN descriptor and the three enhanced fingerprint combinations, we classify the input molecules against their hit/non-hit labels. Finally, the hit probability indicating the likeliness of inhibiting β-tubulin for a given molecule was obtained for later filtering. For simplicity, the basic hyperparameters were adopted from the original implementation of DMPNN.[27,28]

Model Training and Screening

The hit and non-hit datasets were used to form the training dataset to train the DMPNN architecture. As described above, for each input molecule, using the basic DMPNN descriptor and the three enhanced descriptors, the hit probability and cross-entropy of binary classification were calculated against the ground truth. The weights in the DMPNN were updated using backpropagations. Once all molecules from the hit and non-hit datasets were input into the models, the DMPNN was trained and hence learned the capability of discriminating hits from non-hits. Following the training, we input all the molecules in the natural product dataset into the trained DMPNN, and the hit probabilities were obtained using the basic DMPNN descriptor and the three enhanced descriptors. We ranked all candidates and focused on the top molecules with a predicted hit probability larger than 0.8. This screening significantly narrowed the candidate dataset.

Molecule Similarity

In order to obtain diversified chemical structures in selected compounds, we removed molecules that were structurally similar to the molecules in the hit dataset. We calculated the Tanimoto similarity coefficients using RDKit. The coefficient of two given molecules is obtained by calculating the distance between the molecular fingerprints of the two molecules. Candidate compounds having a Tanimoto similarity coefficient larger than the cut-off of 0.4 to any of the molecules in the hit dataset were filtered.

Lipinski’s Rule of Drug-Likeness

The top molecules with optimal hit probability obtained in the initial screening using DMPNN were further analyzed using Lipinski’s rule of drug-likeness. In this study, we set filters as 250 ≤ molecular weight ≤ 500, logP ≤ 5, the amount of hydrogen bond donors ≤ 5, the amount of hydrogen bond acceptors ≤ 10. By applying the rules, we further focused on an even smaller group of candidates, allowing manual evaluations.

Cell Culture, Antibodies, and Reagents

Cells were purchased from American Type Culture Collection (ATCC, Manassas, VA, USA) and were cultured in DMEM with 10% fetal bovine serum and incubated with 5% CO2. Antibodies used in this study were as follows: caspase 3 (9662, CST), Bax (5023, CST), Bcl-2 (2870, CST), CDK1 (201008, Abcam), CDK2 (2546, CST), cyclin E (4129, CST), β-actin (3700, CST), β-tubulin (2128, CST). Compound BD and PMA were purchased from MedChemExpress.

Cell Viability Assay

Cell viability was measured by the MTT assay. MDA-MB-231 cell was dispensed in 96-well plates at a density of 7 × 103 cells/ml for 24 h. Then, cells were treated with different concentrations of compounds for 24 h.

Apoptosis and Cell Cycle Assays

For apoptosis assay, MDA-MB-231 cell was treated with 10 μM BD and 10 μM PMA for 24 h, respectively. Apoptosis ratios were determined by flow cytometry analysis of annexin V/PI double staining. For cell cycle detection, MDA-MB-231 cell was treated with 10 μM BD and 10 μM PMA for 24 h, respectively. Then, the cell cycle distribution was determined by flow cytometry analysis of PI staining.

Immunofluorescence Analysis

The MDA-MB-231 cell was incubated with β-tubulin (1200) in PBS containing 1% BSA incubated overnight at 4 °C, followed by the addition of fluorescent-labeled secondary antibodies (Alexa Fluor 488, ab150077) for 1 h at room temperature. Images were captured using a confocal laser canning microscopy (Zeiss).

Immunoblotting Analysis

Adherent and floating cells were collected and lysed by lysis buffer at 4 °C for 30 min. The protein content of the supernatant was quantified by Bio-Rad DC protein assay (Bio-Rad Laboratories, Hercules, CA, USA). Equal amounts of the total protein were separated by 15% SDS-PAGE and transferred to PVDF membranes, followed by primary antibodies and HRP-conjugated secondary antibodies. Quantification of immunoblot was performed by ImageJ 1.8.0.

Data Availability

Data are available within the article and supplementary files. All other data that support the findings of the study are available from the corresponding author upon reasonable request.

Code Availability

The source codes of DMPNN architecture are provided in the paper.[27] Python codes for the pipeline are available on GitHub (https://github.com/gracewang723/chemprop).