Source-linked AI summary
Generating 3D Molecules for Target Protein Binding
Meng Liu, Youzhi Luo, Kanji Uchino, Koji Maruhashi, Shuiwang Ji
TL;DR
Designing molecules that bind specific proteins is a fundamental drug-discovery challenge with complex 3D and chemical constraints. GraphBP sequentially places atom types and continuous positions using geometry-aware context representations, local coordinates, and flow modeling; experiments show effective generation of molecules with target-site binding ability.
Problem
Designing molecules that bind specific proteins is a fundamental and challenging drug-discovery problem involving complex binding-site context and enormous chemical and continuous 3D spaces.
Method
GraphBP sequentially places atoms at target binding sites, using a 3D graph neural network, a local spherical coordinate system, and a flow model for atom types and relative positions.
Results
27.0% of molecules generated by GraphBP had higher predicted binding affinity than reference molecules, an absolute improvement of 11.1% over LiGAN.
Takeaways & Limitations
GraphBP captures 3D geometry and chemical interactions, places atoms without discretizing 3D space, preserves equivariance, and generates molecules that bind strongly to target proteins.
Abstract
from arXiv · showhide
A fundamental problem in drug discovery is to design molecules that bind to specific proteins. To tackle this problem using machine learning methods, here we propose a novel and effective framework, known as GraphBP, to generate 3D molecules that bind to given proteins by placing atoms of specific types and locations to the given binding site one by one. In particular, at each step, we first employ a 3D graph neural network to obtain geometry-aware and chemically informative representations from the intermediate contextual information. Such context includes the given binding site and atoms placed in the previous steps. Second, to preserve the desirable equivariance property, we select a local reference atom according to the designed auxiliary classifiers and then construct a local spherical coordinate system. Finally, to place a new atom, we generate its atom type and relative location w.r.t. the constructed local coordinate system via a flow model. We also consider generating the variables of interest sequentially to capture the underlying dependencies among them. Experiments demonstrate that our GraphBP is effective to generate 3D molecules with binding ability to target protein binding sites. Our implementation is available at https://github.com/divelab/GraphBP.
1. Introduction
GraphBP addresses the challenging, underexplored task of generating molecules that bind specific protein sites by modeling their 3D and chemical context. It sequentially places atoms with continuous coordinates while preserving equivariance and modeling dependencies among generated variables.
- Structure-based drug design seeks molecules that bind specific target proteins and is a fundamental, challenging drug-discovery problem.
- Machine-learning methods have rarely been explored for generating molecules that bind specific protein binding sites.
- The task requires capturing both the binding site's 3D geometric structure and chemical features across enormous chemical and continuous 3D spaces.
- GraphBP generates 3D molecules by sequentially placing atom types and positions at a target binding site using a 3D graph neural network, local coordinates, and a flow model.
- GraphBP generates variables sequentially to capture dependencies and, according to the authors, outperforms baselines in generating molecules with binding affinity to target sites.
2. Preliminaries and Related Work
Prior work spans 1D, 2D, and 3D molecular generation, while structure-based generation for protein binding sites remains challenging and under-explored. Flow models provide invertible transformations, and autoregressive flows model each variable conditioned on preceding dimensions.
- 1D/2D molecule generation: Molecules are generated as 1D SMILES strings or 2D molecular graphs using sequence methods and deep generative models.
- 3D molecule generation: 3D molecular generation methods construct geometries from 2D graphs, bags of atoms, or no initial molecular structure.
- Structure-based drug design: Generating 3D molecules that bind specific protein binding sites remains challenging and under-explored.
- Structure-based drug design: LiGAN represents protein–ligand complexes as 3D atomic density grids, generates density grids with GANs or VAEs, and fits atoms to obtain molecular geometries.
- Structure-based drug design: LiGAN does not preserve equivariance and discretizes continuous 3D space into grids.
- Autoregressive flow models: Autoregressive flow models use invertible transformations in which each data dimension is conditioned on previous dimensions.
3. Method
GraphBP generates ligand atoms sequentially within a protein binding site, encoding geometric and chemical context, selecting local reference atoms, and placing atoms in an equivariant local coordinate system.
- Problem formulation: The problem is represented as learning the conditional distribution p(M|P), where molecules and binding sites contain atom types and 3D Cartesian coordinates.
- Sequential generation: GraphBP autoregressively places one atom at each step using the binding site and previously generated ligand atoms as context.At the first step, the context is only the binding site.
- Encoding the context: A context graph and 3D graph neural network encode geometric structure and chemical interactions from the current binding-site and ligand context.Atoms are connected using a cutoff distance, and distance embeddings support rotation- and translation-invariant representations.
- Selecting a local reference atom: Auxiliary contact and focal atom classifiers select local reference atoms from the binding site initially and from previously generated ligand atoms thereafter.The selected atom defines the local region used for generating the next ligand atom.
- Placing a new atom: Three context atoms define a local spherical coordinate system in which GraphBP generates invariant distance, angle, and torsion variables for the new atom.The reference atom is paired with its closest and second-closest context atoms to construct the coordinate system.
- Placing a new atom: Because the local coordinate system transforms with the context and the generated geometric variables are continuous, atom coordinates remain equivariant while placement remains flexible.The flow model generates atom type and relative placement variables sequentially, incorporating dependencies among them.
4. Experiments
Experiments evaluate GraphBP on structure-based drug design and random molecular geometry generation. GraphBP generates valid, higher-affinity, diverse molecules and its sequential variable generation improves geometry modeling.
- Experimental Setup: GraphBP is evaluated on CrossDocked2020 using the same train/test split and 10 target proteins as LiGAN.The test set contains 90 protein-ligand reference pairs, with 100 generated molecules per reference binding site.
- Structure-Based Drug Design: 27.0% of GraphBP-generated molecules have higher predicted binding affinity than reference molecules, an absolute 11.1% improvement over LiGAN.GraphBP also generates more valid molecules than the considered LiGAN baselines.
- Structure-Based Drug Design: GraphBP generates more diverse molecules than LiGAN-posterior, with 27.0% versus 15.4% exceeding reference binding affinity.LiGAN-posterior has higher average affinity improvement but lower variance, consistent with slight modifications of reference molecules.
- Structure-Based Drug Design: Examples show generated 3D molecules predicted to bind more strongly than their corresponding reference molecules.The examples are associated with labeled protein PDB IDs and reference ligand IDs.
- Ablation Studies: The ablation comparison is reported in Table 2, while training-loss comparisons between GraphBP and ablation models are shown in Figure 4.The geometry evaluation uses molecular validity and MMD distances between generated and training bond-length distributions.
- Ablation Studies: Sequentially generating atom type, distance, angle, and torsion improves generation performance and fits training data better than independent generation.The sequential method performs best on random molecular geometry generation by modeling dependencies among these variables.
5. Conclusions
GraphBP generates 3D molecules for target protein binding while modeling geometric and chemical context, continuous atom placement, and equivariant generation. It outperforms recent baselines in generating molecules predicted to bind strongly to target proteins.
- Conclusions: GraphBP generates 3D molecules for target protein binding while capturing geometric structures and chemical interactions of protein-ligand complexes.The framework places atoms without discretizing 3D space and preserves equivariance during generation.
- Conclusions: GraphBP outperforms recent baselines significantly in generating 3D molecules that bind strongly to target proteins.This conclusion summarizes the reported effectiveness of the proposed framework.
A. The Detailed Derivation of Lap
The appendix derives the atom-placement loss for GraphBP's flow-based generation. It explains latent-variable mappings, Gaussian priors, and dequantization for discrete atom types.
- Loss Derivation: The detailed derivation expands the atom-placement loss introduced in Eq. (18).The derivation uses the product of vector elements across dimensions and autoregressive flow properties.
- Flow Variables: Latent variables are obtained by inverting the mappings for atom type, distance, angle, and torsion, with Gaussian priors for each latent variable.The priors are denoted pZa, pZd, pZθ, and pZϕ.
- Dequantization: Dequantization converts discrete atom types into continuous values during training so the flow model can optimize their likelihood.An argmax operation converts the dequantized values back to discrete atom types.
B. Dataset Details
The dataset represents ligand and binding-site chemistry using distinct sets of atom types. Ligands use 27 possible atom types, while binding sites use 19.
- Atom Types: Ligands contain 27 possible atom types, including common elements such as B, C, N, O, F, P, S, Cl, Br, I, and Au.The full ligand atom-type inventory is provided in the dataset details.
- Atom Types: Binding sites contain 19 possible atom types, including C, N, O, Na, Mg, P, S, Cl, K, Ca, Zn, I, Cs, and Hg.The binding-site inventory differs from the ligand atom-type inventory.
C. More Experimental Results
Figure 5 compares bond-length distributions from generated molecules with those from training molecules, while the sequential strategy consistently outperforms ablation variants.
- Sequential generation consistently outperforms ablation variants in modeling bond-length distributions.The comparison evaluates whether modeling dependencies among variables improves generated molecular geometries.
- Figure 5 visualizes bond-length distributions for generated molecules and training molecules.