Streamlining Generative Models for Structure-Based Drug Design

Rafal Karczewski, Markus Heinonen, Vikas Garg

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: drug design, binding, docking, graph neural networks, generalization bounds

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Generative models for structure-based drug design (SBDD) aim to generate novel 3D molecules for specified protein targets $\textit{in silico}$. The prevailing paradigm focuses on model expressivity - typically with powerful Graph Neural Network (GNN) models - but is agnostic to binding affinity during training, potentially overlooking better molecules. We address this issue with a two-pronged approach: learn an economical surrogate for affinity to infer an unlabeled molecular graph, and optimize for labels conditioned on this graph and desired molecular properties (e.g., QED, SA). The resulting model FastSBDD achieves state-of-the-art results as well as streamlined computation and model size (up to 1000x faster and with 100x fewer trainable parameters compared to existing methods), paving way for improved docking software. We also establish rigorous theoretical results to expose the representation limits of GNNs in SBDD contexts and the generalizability of our affinity scoring model, advocating more emphasis on generalization going forward.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7461