PaperlayerView exact revision on arXiv

Source-linked AI summary

A 3D Generative Model for Structure-Based Drug Design

Shitong Luo, Jiaqi Guan, Jianzhu Ma, Jian Peng

arXiv:2203.10446v2q-bio.BMcs.LG

TL;DR

Structure-based drug design requires generating molecules that fit specific protein binding sites, but string- and graph-based methods lack the spatial information needed for protein–molecule interactions in 3D. The paper introduces a 3D generative model that estimates atom-occurrence densities and samples atoms autoregressively, producing diverse drug-like molecules with high target-specific binding affinity and supporting linker design.

  • Problem

    Generating molecules that bind specific protein binding sites is challenging because feasible chemical and conformational spaces are enormous, while common string- and graph-based models lack 3D spatial information.

  • Method

    A 3D generative model estimates the probability density of atom types at 3D coordinates conditioned on a binding site and uses autoregressive sampling to build molecules sequentially.

  • Results

    The model generates diverse drug-like molecules with high binding affinity, fairly high QED and SA scores, and applicability to linker design; in examples, 8% and 18% of molecules exceed reference-ligand affinity.

  • Takeaways & Limitations

    The approach supports structure-based molecule generation and linker design directly from 3D protein binding sites without explicit optimization for drug-likeness or synthesizability.

  • Takeaways & Limitations

    The model does not guarantee successful generation of valid molecules in every case.

Abstract

from arXiv · show

We study a fundamental problem in structure-based drug design -- generating molecules that bind to specific protein binding sites. While we have witnessed the great success of deep generative models in drug design, the existing methods are mostly string-based or graph-based. They are limited by the lack of spatial information and thus unable to be applied to structure-based design tasks. Particularly, such models have no or little knowledge of how molecules interact with their target proteins exactly in 3D space. In this paper, we propose a 3D generative model that generates molecules given a designated 3D protein binding site. Specifically, given a binding site as the 3D context, our model estimates the probability density of atom's occurrences in 3D space -- positions that are more likely to have atoms will be assigned higher probability. To generate 3D molecules, we propose an auto-regressive sampling scheme -- atoms are sampled sequentially from the learned distribution until there is no room for new atoms. Combined with this sampling scheme, our model can generate valid and diverse molecules, which could be applicable to various structure-based molecular design tasks such as molecule sampling and linker design. Experimental results demonstrate that molecules sampled from our model exhibit high binding affinity to specific targets and good drug properties such as drug-likeness even if the model is not explicitly optimized for them.

1 Introduction

Structure-based drug design requires generating molecules for specific protein binding sites, but existing string- and graph-based methods lack the spatial information needed for 3D protein interactions. The paper introduces a 3D atom-occurrence model with autoregressive sampling and reports diverse, drug-like molecules with high target affinity.

  • Motivation: Structure-based drug design searches an enormous chemical and conformational space to identify molecules binding specific protein sites.The computational burden arises from both the space of synthetically feasible chemicals and the conformational freedom of compounds and proteins.
  • Motivation: String- and graph-based generative methods lack spatial information about molecule–protein interactions and therefore cannot target specific protein structures.These approaches learn desirable molecular distributions but do not represent how molecules fit their targets in 3D space.
  • Motivation: Existing 3D methods either handle only small organic molecules or rely on voxelization, leaving high-quality drug-scale generation for specific binding sites challenging.Voxelization-based approaches can produce drug-scale molecules, but their sampling quality is limited by the voxel representation.
  • Method: The proposed model estimates p(e, r|C), the probability density of atom types at 3D coordinates conditioned on a protein binding site.A rotationally invariant graph neural network processes the binding-site context so the distribution remains equivariant to rotations and translations.
  • Method: Autoregressive sampling adds atoms sequentially to address dependencies between atoms and capture multiple feasible molecular structures.Independent sampling would not generally yield valid molecules or adequately represent the multimodal chemical space.
  • Results: Experiments report diverse drug-like molecules with high target affinity, fairly high QED and SA scores, and applicability to linker design.The reported drug-likeness and synthesizability emerge even though the model is not specifically optimized for those properties.

2 Related Work

Prior 3D molecular-generation methods model molecules directly in space, but their ability to produce drug-scale structures is limited. Voxel-based approaches extend to drug-scale molecules by converting molecular generation into image generation, while introducing voxelization-related quality limitations.

  • 3D Molecular Generation: Autoregressive and reinforcement-learning methods generate realistic 3D molecules but can handle only small organic molecules.These methods sequentially place atoms using distance constraints or rewards derived from atomic potential energy.
  • 3D Molecular Generation: Voxel-based non-autoregressive methods generate drug-scale molecules by representing molecular structures as 3D images on voxelized meshgrids.This formulation enables the use of convolutional VAE and GAN image-generation techniques.
  • 3D Molecular Generation: Voxelization-based generation can support specific protein pockets, but voxelization heavily limits the quality of molecular sampling.The representation transforms molecular generation into image generation while constraining spatial precision.

3 Method

The method models atom occurrence in a protein binding site's 3D space and samples molecules autoregressively, conditioning each new atom on the protein and previously generated atoms. Cloze-filling training learns atom positions, chemical elements, and when generation should stop.

  • 3.1 3D Generative Model Design: The model represents a binding site as protein atoms with attributes and 3D coordinates, then predicts atom-type occurrence probabilities at query coordinates.The binding site is C = {(a_i, r_i)}, and the model estimates p(e|r, C).
  • 3.1 3D Generative Model Design: A context encoder builds structure-aware atom representations with a k-nearest-neighbor graph, while a spatial classifier aggregates nearby context to predict p(e|r, C).The encoder is designed for context awareness and rotational and translational invariance.
  • 3.2 Sampling: The joint distribution p(e, r|C) combines atom type and coordinate, enabling sampling methods such as MCMC or discretization in the three-dimensional coordinate space.The distribution uses an unknown normalizing constant, while the coordinate dimension remains only three.
  • 3.2 Sampling: Atoms are generated one at a time from p(e, r|C_t), where each context includes the protein atoms and previously sampled atoms, thereby modeling inter-atom dependencies.An auxiliary frontier network stops sampling when all existing atoms are classified as non-frontier; OpenBabel then obtains generated bonds.
  • 3.2 Sampling: The stochastic autoregressive path naturally diverges, producing diverse samples while conditioning new placements on previously sampled atoms.This design addresses both dependency between atoms and multimodality in feasible chemical space.
  • 3.3 Training: Cloze-filling training masks a random molecule portion and uses three losses to learn occupancy, chemical-element prediction, and frontier classification.Positive coordinates come from masked atoms, negative coordinates from ambient space, and the full loss is L = L_BCE + L_CAT + L_F.

4 Experiments

The experiments evaluate molecule generation and linker prediction using structural, chemical, synthetic-accessibility, similarity, recovery, and binding-affinity measures. The model generates diverse, drug-like molecules for binding sites and outperforms or matches graph- and voxel-based baselines on key task metrics.

  • Experimental setup: The evaluation covers molecule design for binding sites and linker prediction between two disconnected fragments.The model is tested on both generating complete molecules and constructing substructures that connect fragments in a binding-site context.
  • Experimental setup: The molecule-design evaluation measures binding affinity, drug-likeness, synthesizability, high-affinity sample percentage, and molecular diversity.Vina measures binding affinity, QED measures drug-likeness, normalized SA measures synthesizability, and diversity averages pairwise Tanimoto similarities over Morgan fingerprints.
  • 4.1 Molecule Design: The model generates diverse molecules with higher binding affinity and desirable drug-likeness and synthesizability, outperforming liGAN on all reported metrics.The paper highlights especially better QED performance than the CNN-based voxel-generation baseline, interpreting this as more realistic drug-like molecules.
  • 4.1 Molecule Design: In the 2hcj example, 8% of generated molecules exceed the reference molecule’s binding affinity, while average QED and SA are 0.483 and 0.663.The reported 8% is below the median 18.5% across binding sites.
  • 4.1 Molecule Design: In the 4rlu example, average QED and SA are 0.728 and 0.785, and 18% of sampled molecules exceed the reference molecule’s binding affinity.The visualized high-affinity molecules generally resemble the reference structure and share important substructures.
  • 4.2 Linker Prediction: For linker prediction, the method matches DeLinker on Vina score but clearly outperforms it on Similarity and Percentage of Recovery.The authors interpret these results as evidence that the method links fragments more realistically, while examples further demonstrate suitable linker generation.

5 Conclusions and Discussions

The paper concludes that its 3D generative model and autoregressive sampler can design drug-like molecules for specific binding sites, while acknowledging that valid generation is not guaranteed.

  • 5 Conclusions and Discussions: The model estimates atom-occurrence densities in 3D space and uses autoregressive sampling to generate drug-like molecules for specific binding sites.Its effectiveness is demonstrated through experiments on structure-based molecule design.
  • 5 Conclusions and Discussions: The method does not guarantee successful generation of valid molecules in every case.The authors suggest incorporating graph representations and techniques such as valency checking and property optimization as future work.

A.1 Molecule Design

The model generates 3D molecules that are generally more realistic and diverse than liGAN samples, while also producing novel molecules with higher binding affinity than the reference molecule.

  • Molecule Design: Our generated molecules are generally more realistic than liGAN outputs, which often contain bonds that are too short and angles that are too sharp.The comparison uses six highest-binding-affinity molecules from each method for each binding site.
  • Molecule Design: Our generated molecules are more diverse, whereas liGAN’s 3D atom configurations are often similar.
  • Molecule Design: The model generates novel molecules that are obviously different from the reference molecule and achieve higher binding affinity.

A.2 Linker Prediction

The method also supports linker prediction, where examples indicate that its predicted linkers are generally more likely to recover or resemble the original structures than the baseline comparison.

  • Linker Prediction: The method is generally more likely than DeLinker to produce linkers that recover or resemble the original structure.DeLinker is graph-based and does not generate 3D linker structures, so the comparison displays 2D molecular graphs.
  • Linker Prediction: The examples use five randomly selected test cases and five representative predicted molecules per case, spanning best, worst, and intermediate similarity.

B Additional Model Details

The model samples atoms sequentially in 3D space, using Metropolis-Hastings for the first atom and meshgrid-based probability evaluation after atoms have been placed.

  • Additional Model Details: Metropolis-Hastings samples the first atom from the marginal distribution p(r|C), with coordinate-element pairs selected by highest joint probability.
  • Additional Model Details: The initial sampling draws 1,000 samples from a protein-atom Gaussian mixture model with 1Å standard deviation, using a 0.1Å proposal and 500 steps.
  • Additional Model Details: After atoms are placed, the method evaluates non-normalized joint probabilities on 0.1Å meshgrids restricted to regions 1.0Å–2.0Å from predicted frontier atoms.The restriction saves memory, and frontier atoms are predicted by the frontier network.
  • Additional Model Details: Beam search generates 100 different molecules for each binding site with beam width 300.

B.2 Hyperparameters

The model uses shared hyperparameters across molecule design and linker prediction, with specified graph, classifier, optimizer, batch, and training-duration settings.

  • Hyperparameters: The context encoder uses k-NN neighborhood size 48, six message-passing layers, and hidden-feature dimension 256.
  • Hyperparameters: The spatial classifier uses feature dimension 128 and aggregates 32 nodes.
  • Hyperparameters: Training uses Adam with learning rate 0.0001, batch size 4, and 1.5 million iterations requiring about 2 days on GPU.
Loading 2203.10446v2…