Geometric Deep Learning for Structure-Based Ligand Design

Alexander S. Powers, Helen H. Yu, Patricia Suriana, Rohan V. Koodli, Tianyu Lu, Joseph M. Paggi, Ron O. DrorACS 2023 · DOI

TL;DR

This paper introduces FRAME, a novel deep learning framework that uses 3D protein-ligand structures to intelligently expand small molecules by adding chemical fragments. FRAME outperforms traditional docking methods in generating drug-like molecules with improved predicted affinity and selectivity, offering a powerful new tool for accelerating drug discovery.

What Problem Does This Paper Address?

Designing new drugs is a long and expensive process, with a major bottleneck being the identification of promising candidate molecules. A key strategy in drug discovery is to modify existing molecules (ligands) to improve their properties, such as how strongly they bind to a target protein (affinity) or how specifically they bind to the intended target versus others (selectivity). This modification often involves adding chemical groups, called fragments, to the ligand. For lead optimization, one or two fragments are added iteratively. In fragment-based drug discovery, multiple fragments are added to a small starting molecule. The challenge lies in predicting where to add these fragments and which fragments to choose to achieve the desired improvements, as the space of possible chemical modifications is vast and predicting outcomes is difficult. Existing computational methods, like docking, can be slow and may have inherent biases.

Methods

The core of FRAME is a comprehensive framework that leverages deep learning and 3D structural information of protein-ligand complexes. The process is framed as a sequence of steps:

  1. Attachment Point Selection: An SE(3)-equivariant neural network (the "Attachment Location Model") predicts where on the current ligand a new fragment should be attached. It does this by scoring potential attachment points (hydrogen atoms) on the ligand based on the 3D context of the protein pocket.
  2. Fragment Selection and Geometry Prediction: Another SE(3)-equivariant neural network (the "Fragment Scoring Model") evaluates candidate fragments from a user-defined library. It scores not only the fragment itself but also its predicted 3D geometry when attached to the ligand.

These models are trained on existing high-quality protein-ligand structures, learning to recognize patterns associated with favorable interactions and synthetic feasibility without explicit rules. The expansion process can be iterative, continuing until a user-defined goal (e.g., molecular weight) is met or the model predicts no further viable attachment points.

Datasets

The study utilized curated datasets of high-resolution 3D protein-ligand structures. An initial set was obtained from the PDBBind database and filtered to include only drug-like, high-affinity ligands, excluding biomolecules like lipids and carbohydrates. This resulted in a dataset of 4200 ligands. For training the models, "reference trajectories" were created by sequentially removing fragments from these ligands. Two specific datasets were derived from these trajectories: one for evaluating attachment point prediction and another for scoring candidate fragments. The data was split into training (70%), validation (15%), and testing (15%) sets, with careful consideration to ensure no protein in one set had more than 30% sequence identity with proteins in other sets to assess generalizability. A custom fragment library was also created, comprising 900 unique fragments.

Key Insights and Ideas

  • Learning from Structure, Not Rules: FRAME's key innovation is learning to predict favorable fragment additions directly from 3D protein-ligand structures, bypassing the need for hard-coded chemical rules or explicit knowledge of interaction types (like hydrogen bonds or pi-pi stacking).
  • SE(3)-Equivariant Networks for Geometry: The use of SE(3)-equivariant neural networks is crucial. These networks are inherently invariant to rotation and translation of the input structures, allowing them to accurately capture the precise 3D geometry of the ligand within the protein pocket, which is fundamental for predicting interactions.
  • Iterative Expansion as a Sequence: Framing ligand expansion as a sequence of discrete steps (attachment, fragment selection, geometry) makes the problem tractable for machine learning and allows for flexible application to various drug design tasks.
  • Orthogonal Hypothesis Generation: FRAME provides a complementary approach to traditional docking methods. By learning from successful ligand designs, it can generate hypotheses that might be missed by physics-based scoring functions.

Key Results

  • Accurate Attachment Point Prediction: The attachment location model achieved high precision (95%) and recall (92%) in identifying actual attachment points used in reference ligands, significantly outperforming a random baseline.
  • Effective Fragment Ranking: The fragment scoring model successfully ranked fragments that form key interactions, with the fine-tuned version identifying reference fragments as the top choice 45% of the time and within the top 10 fragments 65% of the time, outperforming random selection and docking scores.
  • Improved Affinity and Selectivity: FRAME-expanded ligands showed improved predicted binding affinity (median docking score of -7.54 kcal/mol vs. -4.36 kcal/mol for starting molecules) and better selectivity compared to starting molecules and iterative docking methods.
  • Drug-Like Properties: FRAME generated molecules with more drug-like chemical properties (e.g., synthetic accessibility, number of rings, rotatable bonds) compared to iterative docking and random expansion methods.
  • Robustness: The framework demonstrated robustness to perturbations in the starting molecule's position and orientation.

Limitations and Future Work

  • Not a Replacement for Chemists: The authors emphasize that FRAME is a tool for hypothesis generation and not a substitute for medicinal chemists.
  • Occasional Failures: FRAME can still produce suboptimal ligands in some cases, leading to worse docking scores or increased synthetic complexity. This is attributed to the autoregressive nature of the approach and potential for occasional mistakes.
  • Computational Estimates: The evaluation relies on computational estimates (docking scores, predicted logP, etc.), and experimental validation is ultimately necessary.
  • Future Directions: The authors suggest exploring more sophisticated search strategies beyond greedy selection, combining FRAME with global structure generation methods (like diffusion models), and applying the framework to specific, real-world drug design projects. They also note that the fragment library can be easily adapted for different applications.

Summary updated

Abstract

A pervasive challenge in drug design is determining how to expand a ligand-a small molecule that binds to a target biomolecule-in order to improve various properties of the ligand. Adding single chemical groups, known as fragments, is important for lead optimization tasks, and adding multiple fragments is critical for fragment-based drug design. We have developed a comprehensive framework that uses machine learning and three-dimensional protein-ligand structures to address this challenge. Our method, FRAME, iteratively determines where on a ligand to add fragments, selects fragments to add, and predicts the geometry of the added fragments. On a comprehensive benchmark, FRAME consistently improves predicted affinity and selectivity relative to the initial ligand, while generating molecules with more drug-like chemical properties than docking-based methods currently in widespread use. FRAME learns to accurately describe molecular interactions despite being given no prior information on such interactions. The resulting framework for quality molecular hypothesis generation can be easily incorporated into the workflows of medicinal chemists for diverse tasks, including lead optimization, fragment-based drug discovery, and <i>de novo</i> drug design.
BibTeX
@inproceedings{powers2023geometric,
  author = {Alexander S. Powers and Helen H. Yu and Patricia Suriana and Rohan V. Koodli and Tianyu Lu and Joseph M. Paggi and Ron O. Dror},
  title = {{Geometric Deep Learning for Structure-Based Ligand Design}},
  booktitle = {ACS},
  year = {2023},
  doi = {10.1021/acscentsci.3c00572},
}
Figure