Re-ranking binding mode predictions using graph neural networks with protein-ligand contact map

Re-ranking binding mode predictions using graph neural networks with protein-ligand contact map

Glauco Endrigo1 , Eric Allison Philot2 and Ana Lígia Scott2

1Department of Mathematics, Computing, and Cognition (CMCC, Federal University of ABC, São Paulo, SP (email:[email protected])

Docking programs are able to generate ligand conformations similar to crystallographically determined protein/ligand complex structures for at least one of the targets (Guedes, et. al, 2018). However, scoring functions are less successful at distinguishing the crystallographic conformation from the set of docked poses. In that sense, a significant limitation in molecular docking arises from the lack of confidence in scoring functions' ability to provide accurate binding energies (Ramírez and Caballero, 2018). In this study, we propose employing a Graph-Convolutional Neural Network (GCNN) to learn ligand-protein contact information and re-rank docking poses from conventional methods (Morrone et al., 2020). For each sample in the PDBbind-core dataset (Liu et al., 2015), docking calculations are performed to generate possible binding poses of the protein-ligand complex using the AutoDock Vina software. A pose is labeled as the positive class if the root-mean-square deviation (RMSD) concerning its experimentally verified binding structure is less than 2 Å or labeled as a negative sample if the RMSD is greater than 4 Å. Samples with an RMSD between 2 and 4 Å are omitted (Morrone et al., 2020). After obtaining the poses, the representation of protein-ligand complexes is done in such a way that protein atoms whose minimum distance to ligand atoms is greater than 8 Å are excluded. The ligands, initially represented by smiles are transformed into graphs using the RDKit cheminformatics library. The amino acid sequence representing the protein is transformed into a graph using Pcons4 (Michel, et al., 2018).

This work is supported by CAPES

References

Guedes, I.A., Pereira, F.S.S. and Dardenne, L.E. (2018) ‘Empirical Scoring Functions for Structure-Based Virtual Screening: Applications, Critical Aspects, and Challenges’, Frontiers in Pharmacology, 9. Available at: https://www.frontiersin.org/articles/10.3389/fphar.2018.01089
Michel, M., Menéndez Hurtado, D. and Elofsson, A. (2018) ‘PconsC4: fast, accurate and hassle-free contact predictions’, Bioinformatics, 35(15), pp. 2677–2679. Available at: https://doi.org/10.1093/bioinformatics/bty1036.
Morrone, J.A. et al. (2020) ‘Combining Docking Pose Rank and Structure with Deep Learning Improves Protein-Ligand Binding Mode Prediction over a Baseline Docking Approach’, Journal of Chemical Information and Modeling, 60(9), pp. 4170–4179. Available at: https://doi.org/10.1021/acs.jcim.9b00927.
Ramírez, D. and Caballero, J. (2018) ‘Is It Reliable to Take the Molecular Docking Top Scoring Position as the Best Solution without Considering Available Structural Data?’, Molecules : A Journal of Synthetic Chemistry and Natural Product Chemistry, 23(5), p. 1038. Available at: https://doi.org/10.3390/molecules23051038.