Infinite Physical Monkey: Do Deep Learning Methods Really Perform Better in Conformation Generation?

Abstract

Conformation Generation is a fundamental problem in drug discovery and cheminformatics. And organic molecule conformation generation, particularly in vacuum and protein pocket environments, is most relevant to drug design. Recently, with the development of geometric neural networks, the data-driven schemes have been successfully applied in this field, both for molecular conformation generation (in vacuum) and binding pose generation (in protein pocket). The former beats the traditional ETKDG method, while the latter achieves similar accuracy compared with the widely used molecular docking software. Although these methods have shown promising results, some researchers have recently questioned whether deep learning (DL) methods perform better in molecular conformation generation via a parameter-free method. To our surprise, what they have designed is some kind analogous to the famous infinite monkey theorem, the monkeys that are even equipped with physics education. To discuss the feasibility of their proving, we constructed a real infinite stochastic monkey for molecular conformation generation, showing that even with a more stochastic sampler for geometry generation, the coverage of the benchmark QM-computed conformations are higher than those of most DL-based methods. By extending their physical monkey algorithm for binding pose prediction, we also discover that the successful docking rate also achieves near-best performance among existing DL-based docking models. Thus, though their conclusions are right, their proof process needs more concern.

Conformer Generation for Structure-Based Drug Design: How Many and How Good?

“…This speculation is supported by reports that resampling RDKit ensembles using clustering can substantially improve its coverage metric on GEOM-QM9 and GEOM-Drugs. 34 Energy Minimization Is a Valuable Postprocessing Step. Energy minimizing-generated conformers generally improve their ability to recapitulate bioactive conformers (Figures 3 and 4), and selecting the lowest-energy conformer generally performs better than a random conformer (Figures 5, S6, 7, and 11).…”

“…33 Deep generative models significantly outperform RDKit at this particular task, but an extended sampling and clustering approach using RDKit achieves a highly competitive performance. 34 It is not clear that it is a fair comparison to compare methods that utilize different amounts of sampling, 35 so here we evaluate RDKit and a deep generative model using identical sampling and ensemble formation criteria. As the direct molecular conformation generation (DMCG) 14 was found to perform best at the task of reconstituting the ensembles of the GEOM-Drugs subset of GEOM, we evaluate it here at the task of bioactive conformation recovery.…”

“…DMCG is an end-to-end generative model with a variational encoder/ decoder architecture that learns all network parameters from the training data distribution, GEOM-Drugs. 31 However, our main goal is not to extend previous evaluations 26,27,34 but to explore the impact of various choices made in the conformer generation process, such as the size of the ensemble, the criteria for including conformers in the ensemble, and the use (or not) of energy minimization, on the ultimate end point of the common structure-based tasks of pharmacophore search and molecular docking.…”

Uni-Mol: A Universal 3D Molecular Representation Learning Framework

“…We leave the details of molecular conformation generation to Appendix C.4, as paper [57] pointed out that the current benchmark for molecular conformation generation could be wrong.…”

Integrating Expert Knowledge with Deep Learning Improves QSAR Models for CADD Modeling

“…[5,12,13] only compares their method with other GNN methods), or the task is not on QSAR modeling (such as quantum mechanics dataset QM9 used in [9]). Recently, some authors expressed doubts about deep learning performance over traditional methods in molecular tasks [14,15]. Ultimately, it remains a mystery whether GNNs are consistently better than methods that rely on traditional descriptor in CADD [16,17].…”