PaperlayerView exact revision on arXiv

Source-linked AI summary

An Equivariant Generative Framework for Molecular Graph-Structure Co-Design

Zaixi Zhang, Qi Liu, Chee-Kong Lee, Chang-Yu Hsieh, Enhong Chen

arXiv:2304.12436v1q-bio.BMcs.LG

TL;DR

Molecule design spans an enormous chemical space, while existing approaches often do not jointly model 2D topology and 3D geometry. MolCode generates molecular graphs and structures together, and experiments report effectiveness across molecule design tasks. The paper concludes that topology and geometry provide complementary information for molecular generation.

  • Problem

    The chemical space is too large for exhaustive searches, while many molecule-generation methods do not unify 2D topology and 3D geometry or explicitly model bonds.

    Source: Introduction · p. 1

  • Method

    MolCode uses an autoregressive flow to generate atom types, bonds, and coordinates sequentially, with equivariant graph encoding and bond-aware attention.

    Source: Introduction · p. 2, Sequential Generation with Flow Models · p. 3

  • Results

    Experiments demonstrate MolCode's effectiveness across de novo molecule design, targeted molecule discovery, and structure-based drug design, with reported advantages over existing generative models for from-scratch generation.

    Source: Introduction · p. 2, Conclusion · p. 6

  • Takeaways & Limitations

    The paper concludes that 2D topology and 3D geometry contain intrinsically complementary information for molecular representation and generation.

    Source: Conclusion · p. 6

    Applying MolCode to significantly larger systems with more diverse atom types will require modifications for scalability and robustness; wet-lab validation remains future work.

    Source: Conclusion · p. 6

Abstract

from arXiv · show

Designing molecules with desirable physiochemical properties and functionalities is a long-standing challenge in chemistry, material science, and drug discovery. Recently, machine learning-based generative models have emerged as promising approaches for \emph{de novo} molecule design. However, further refinement of methodology is highly desired as most existing methods lack unified modeling of 2D topology and 3D geometry information and fail to effectively learn the structure-property relationship for molecule design. Here we present MolCode, a roto-translation equivariant generative framework for \underline{Mol}ecular graph-structure \underline{Co-de}sign. In MolCode, 3D geometric information empowers the molecular 2D graph generation, which in turn helps guide the prediction of molecular 3D structure. Extensive experimental results show that MolCode outperforms previous methods on a series of challenging tasks including \emph{de novo} molecule design, targeted molecule discovery, and structure-based drug design. Particularly, MolCode not only consistently generates valid (99.95$\%$ Validity) and diverse (98.75$\%$ Uniqueness) molecular graphs/structures with desirable properties, but also generate drug-like molecules with high affinity to target proteins (61.8$\%$ high-affinity ratio), which demonstrates MolCode's potential applications in material design and drug discovery. Our extensive investigation reveals that the 2D topology and 3D geometry contain intrinsically complementary information in molecule design, and provide new insights into machine learning-based molecule representation and generation.

Introduction

Molecule generation must address an enormous chemical space and the complementary information in molecular topology and geometry. MolCode jointly generates 2D graphs and 3D structures, aiming to capture their relationship for molecule design.

Introduction

  • The chemical space is estimated at 10^607, precluding exhaustive computational or experimental searches of possible molecular candidates.Machine-learning methods have accelerated exploration of chemical compound space, and deep generative models have been used to generate molecules and optimize properties.Source: Introduction · p. 1
  • 2D graph-only methods leave molecular 3D structures unknown, although the same graph can have conformations with different quantum properties.Intramolecular interactions and rotations of structural motifs contribute to this conformational variation.Source: Introduction · p. 1
  • Many 3D point-cloud methods do not explicitly model chemical bonds and valency, which may yield invalid molecules and inhibit realistic substructures.The passage gives benzene rings as an example of realistic substructures.Source: Introduction · p. 1
  • MolCode jointly generates molecular 2D graphs and 3D structures, using each representation to inform generation of the other.Its autoregressive flow generates atom types, bonds, and coordinates sequentially; the 3D geometry informs bond generation, while bonds constrain coordinate prediction.Source: Introduction · p. 2, Figure 1 · p. 2

Results

MolCode sequentially generates molecular graphs and 3D structures, with reported results across random generation, property targeting, and pocket-conditioned drug design. The findings include high validity and uniqueness, targeted property outcomes, and ligand affinity gains.

Sequential Generation with Flow Models

  • MolCode sequentially predicts atom types, bond types, and 3D coordinates, using equivariant graph features, bond-encoding attention, and local spherical relative coordinates.The local coordinates are used to ensure coordinate equivariance and likelihood invariance.Source: Sequential Generation with Flow Models · p. 3

De novo Molecule Design

  • 99.95% Validity and 98.75% Uniqueness; without bond information, these fell to 92.12% Validity and 94.32% Uniqueness, respectively.Even without a validity check, MolCode achieved 94.60% Validity.Source: De novo Molecule Design · p. 4
  • Generated bond-length and bond-angle distributions align well with QM9, and MolCode obtains lower KL divergence than other methods and its variant without bond information.The reported examples include carbon-carbon and carbon-oxygen bond lengths, and carbon-carbon-carbon and carbon-carbon-oxygen angles.Source: De novo Molecule Design · p. 4, Figure 2 · p. 4

Targeted Molecule Discovery

  • For targeted HOMO-LUMO gap and isotropic polarizability discovery, fine-tuned MolCode achieved Good Percentages of 87.76% and 38.40%, respectively, although the biased datasets were only 3.20% and 2.04% of QM9, respectively.Source: Targeted Molecule Discovery · p. 5
  • MolCode captures QM9 atom, bond, and ring counts; lower-gap bias favors more nitrogen/oxygen atoms, double bonds, and six-membered rings, while polarizability bias favors larger molecules.The paper associates the lower-gap motifs with aromatic rings and conjugated systems, and larger atom, bond, and ring counts with prerequisites for high polarizability.Source: Targeted Molecule Discovery · p. 5

Structure-based Drug Design

  • 61.8% of protein pockets had generated molecules with higher affinity than their reference molecules on average; generated molecules also showed higher QED and SA.The paper reports lower Vina scores than baseline methods and gives QED and SA as additional ligand-quality measures.Source: Structure-based Drug Design · pp. 5–6
  • Examples of higher-affinity generated ligands are structurally diverse and differ substantially from their corresponding test-set reference molecules.The paper presents these examples as evidence of generating diverse, novel molecules rather than merely reproducing known molecules.Source: Structure-based Drug Design · p. 6

Conclusion

MolCode concurrently designs molecular graphs and 3D structures, with experiments demonstrating effectiveness across three molecule-design tasks. The authors identify complementary information in 2D topology and 3D geometry, and outline extensions to larger systems and experimental validation.

Conclusion

  • MolCode concurrently designs 2D molecular graphs and 3D structures, with experiments demonstrating effectiveness in de novo design, targeted discovery, and structure-based drug design.Source: Conclusion · p. 6
  • Unified modeling of 2D topology and 3D geometry improves molecular generation performance, and the authors find these representations intrinsically complementary.Source: Conclusion · p. 6
  • Scaling to larger systems with diverse atom types requires modifications for scalability and robustness; proposed extensions also include chemical priors and wet-lab validation.The paper proposes fragment-by-fragment generation as a way to incorporate chemical priors and potentially speed generation.Source: Conclusion · p. 6

Methods

MolCode generates atom types, bonds, and 3D coordinates sequentially in a unified molecular graph, conditioning each step on the intermediate structure. Its methods combine equivariant geometric encoding, bond-aware attention, validity checks, and task-specific training and sampling procedures.

Methods

  • The CrossDocked dataset is filtered to about 160k structures, clustered at 30% sequence identity, and used to train on 100,000 protein–ligand pairs and test on 100 proteins.For QM9 targeted discovery, molecules with HOMO-LUMO gaps below 4.5 eV or isotropic polarizabilities above 91 Bohr3 form the fine-tuning sets.Source: Methods · pp. 6–7
  • MolCode uses an autoregressive flow to generate each new atom’s type, bond types, and 3D coordinates; dequantization converts discrete atom and bond types into continuous values for the flow.The molecule is represented as a 3D graph containing atom types, adjacency information including a no-edge type, and Cartesian coordinates.Source: Methods · p. 7
  • New-atom coordinates are generated as distance, angle, and torsion relative to a local spherical coordinate system, yielding roto-translation equivariance.An equivariant encoder conditions generation on the intermediate graph; SphereNet is used for QM9 and EGNN for CrossDocked2020.Source: Methods · pp. 7–8
  • Bond-encoded multi-head self-attention combines atom embeddings with covalent-bond information to capture bonding and global molecular context.The encoder node embeddings come from the intermediate 3D graph, while bond embeddings enter the attention mechanism.Source: Methods · p. 8
  • Focal atoms guide where new atoms are placed; sampling stops when no atom exceeds the 0.5 classification threshold or no new bond is formed, and bonds violating valency are rejected.Prim’s algorithm sets generation orders using interatomic distances, while the first atom is randomly sampled during training.Source: Methods · pp. 8–9
  • Sampling temperatures are selected by grid search on random-generation validity and uniqueness, with lower variance for bond lengths and angles than for flexible torsions.MolCode is trained with Adam; the described schedules use 100 epochs for random generation and 40 or 80 epochs for the two targeted properties.Source: Methods · p. 9

1 Supplementary Information

The supplementary information describes MolCode’s training and generation procedures and presents a substructure analysis comparing generated molecules with the training set.

1 Supplementary Information

  • Table S1 reports KL divergences for bond lengths and bond angles between generated molecules and the training set.The caption identifies bond lengths in the upper part and bond angles in the lower part; it provides no cell values.Source: Table S1 · p. 14
  • Training uses batches of molecules, orders each molecule’s atoms, and adds a binary cross-entropy loss for focal-atom selection.The algorithm loops over molecules and atoms while obtaining atom variables and reference atoms.Source: §1 · p. 16
  • Generation initializes a one-carbon molecule at the origin and iteratively generates atoms, bond types, and 3D coordinates subject to valency checks.The algorithm samples bond types and rejects candidates that violate the stated valency condition.Source: §1 · p. 16
Loading 2304.12436v1…