PaperlayerView exact revision on arXiv

Source-linked AI summary

GraphVF: Controllable Protein-Specific 3D Molecule Generation with Variational Flow

Fang Sun, Zhihao Zhan, Hongyu Guo, Ming Zhang, Jian Tang

arXiv:2304.12825v1q-bio.BMcs.AIcs.LG

TL;DR

Existing 3D molecular generation methods do not take into account useful 2D pharmacophoric patterns, which are critical for binding. GraphVF uses a flow-based generative model integrating 3D geometric and 2D topological constraints, encoding 2D molecular topology patterns in the prior distribution. GraphVF is reported to achieve high binding affinity, outperform state-of-the-art methods by a large margin, and produce more realistic substructure distributions.

  • Problem

    Existing 3D molecular generation methods do not take into account useful 2D pharmacophoric patterns, which are critical for binding.

    Source: §1 · p. 1

  • Method

    GraphVF uses a flow-based generative model integrating 3D geometric and 2D topological constraints, encoding 2D molecular topology patterns in the prior distribution.

    Source: §1 · p. 2

  • Results

    GraphVF is reported to achieve high binding affinity, outperform state-of-the-art methods by a large margin, and produce more realistic substructure distributions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    GraphVF generates protein-binding 3D molecules with desired ligand substructures and biochemical properties.

    Source: §6 · p. 9

Abstract

from arXiv · show

Designing molecules that bind to specific target proteins is a fundamental task in drug discovery. Recent models leverage geometric constraints to generate ligand molecules that bind cohesively with specific protein pockets. However, these models cannot effectively generate 3D molecules with 2D skeletal curtailments and property constraints, which are pivotal to drug potency and development. To tackle this challenge, we propose GraphVF, a variational flow-based framework that combines 2D topology and 3D geometry, for controllable generation of binding 3D molecules. Empirically, our method achieves state-of-the-art binding affinity and realistic sub-structural layouts for protein-specific generation. In particular, GraphVF represents the first controllable geometry-aware, protein-specific molecule generation method, which can generate binding 3D molecules with tailored sub-structures and physio-chemical properties. Our code is available at https://github.com/Franco-Solis/GraphVF-code.

1 INTRODUCTION

GraphVF addresses two gaps in protein-pocket 3D generation: preserving 2D pharmacophoric patterns and controlling chemical substructures or properties. It combines 2D topology and pocket geometry through a variational flow, and reports improved binding and structural realism.

1 INTRODUCTION

  • Existing 3D generation methods do not account for 2D pharmacophoric patterns, which can be important to ligand activity and binding affinity.The introduction illustrates this with DMT’s additional methyl groups and their attractive charge interaction with Asp-231.Source: §1 · p. 1
  • Pocket-specific 3D controllability over properties such as solubility, polarizability, and toxicity remains unexplored in prior work, despite their relevance to drug quality.The passage connects these properties to in-vivo exposure, including ADME, and efficacy in clinical trials.Source: §1 · p. 2
  • GraphVF encodes 2D topology patterns in a prior and uses a flow model conditioned on 3D pocket geometry to generate ligands with specified structures and properties.The authors describe this as the first method for generating 3D molecules with specified chemical substructures or physicochemical properties.Source: §1 · p. 2
  • GraphVF reports high binding affinity, a large-margin advantage over state-of-the-art methods, and more realistic substructure distributions.The paper also presents its VAE-style interface as a way to impose customized substructure and physicochemical constraints.Source: §1 · p. 2

2 RELATED WORK

Prior molecule-generation work spans 2D graphs, unbounded 3D molecules, and protein-specific 3D ligands. The surveyed approaches differ in how they represent molecules and control generation, with protein-specific geometric methods lacking explicit topology perception and targeted control.

2 RELATED WORK

  • As geometric deep learning and probabilistic generative models developed in recent years, de novo molecule generation evolved drastically; the paper identifies three major tasks: 2D, 3D unbounded, and 3D protein-specific generation.Source: §2 · p. 2

3D Molecule Generation for Target Protein Binding

  • Figure 1 depicts GraphVF sampling ligand atoms and bonds autoregressively through a pocket-geometry-parameterized flow, using a prior encoded from 2D topology.During controllable generation the reference ligand is absent; the latent encoding is formed by averaging molecules with specified qualities or selected using Bayesian optimization.Source: Figure 1 · p. 2

2D molecule generation

  • 2D generation methods create valid molecular graphs, using approaches such as autoregressive atom-and-bond generation, scaffold-guided generation, or property optimization.VAE-based methods can support zero-shot optimization when an optimal latent embedding is acquired without retraining.Source: §2 · p. 2

3D Non-Protein Specific Molecule Generation

  • Unbounded 3D generation targets valid molecules with reasonable conformations; surveyed methods include equivariant representations, hierarchical structure generation, and iterative denoising.The passage notes that EDM can be configured for property control through retraining, while RetMol is retrieval-based and needs no task-specific fine-tuning.Source: §2 · pp. 2–3

3D Molecule Generation for Target Protein Binding

  • Protein-specific 3D methods generate ligands from pocket geometry, but the surveyed approaches cannot explicitly perceive ligand pharmacophoric topology or control specific substructures and physicochemical properties.Examples include Pocket2Mol’s spatial autoregression, GraphBP’s normalizing flow, and DiffBP’s non-autoregressive diffusion.Source: §2 · p. 3
  • The roadmap compares three molecule-generation tasks on the horizontal axis and qualitative overall controllability on the vertical axis.Its dashed baseline represents task-inherent controllability; the distance above it represents control over customized properties.Source: §2 · p. 3

3 PRELIMINARIES

The preliminaries specify the protein–ligand graph-generation task, describe geometry encoding with bond-aware message passing, and introduce flow-based generation from a prior. The prior can also encode 2D structural information as a latent distribution.

3.1 Problem Setup

  • Training uses binding-pose protein–ligand pairs, while generation receives a protein target and aims to produce a tightly binding 3D ligand.Proteins and ligands are graphs with atom-type and position node features and bond-type edge features; the ligand’s 3D geometry and 2D topology are denoted R3D and R2D.Source: §3.1 · p. 3

3.2 Geometry Graph Encoding

  • EchNet adapts SchNet’s distance-based message passing by incorporating bond types, addressing the stated inadequacy of distance alone for modeling covalent bonds.Its message-passing equation concatenates a bond-type embedding with distance encoding; the passage gives characteristic bond lengths as motivation.Source: §3.2 · p. 3

3.3 Generative Flow Model

  • A flow maps a simple prior to a complex data distribution through invertible transformations, and autoregressive computation supports tractable Jacobian determinants.For GraphVF, a junction-tree encoder turns canonical 2D substructures into a latent encoding used in the prior; samples are transformed into 3D binding-molecule graphs.Source: §3.3 · pp. 3–4, Figure 3 · p. 4

4 THE PROPOSED METHOD

GraphVF encodes 2D ligand topology in a variational prior and uses a flow model to generate protein-binding 3D ligands. Its prior can be steered to control generated substructures and physicochemical properties.

4.1 2D Topology Prior Encoding

  • GraphVF extracts canonical scaffold fragments into a junction tree, pools them into a root embedding, and maps that embedding to a Gaussian topology prior.Source: §4.1.1 · p. 4

4.1.2 Junction Tree Message Passing.

  • Tree message passing aggregates child-to-parent messages bottom-up through a GRU, yielding the junction-tree root representation.Source: §4.1.2 · pp. 4–5

4.2.1 Autoregressive Generation via Flow.

  • Ligands are generated autoregressively at the binding site: GraphVF builds a radius graph, selects a focal atom, and uses spherical coordinates to preserve equivariance.Source: §4.2.1 · p. 5

4.2.2 Parameterization of the Flow Transformation.

  • At each step, GraphVF samples atom type, bonds, and position from a variational prior and transforms these variables into ligand features through sequential flow parameterizations.Generation continues until the focal classifier finds no eligible atom.Source: §4.2.1 · p. 5, §4.2.2 · p. 6

4.2.3 The Objective Function.

  • Training combines the flow negative log-likelihood with a KL regularizer aligning the topology prior to a standard Gaussian, alongside focal-classifier supervision.Focal labels come from a ring-first expert trajectory.Source: §4.2.3 · p. 6

4.3 Controllable generation

  • The variational prior enables zero-shot control of selected generated-molecule properties without retraining the model.Source: §4.3 · p. 6
  • For qualitative attributes such as a substructure, mean aggregation of structural encodings from molecules with that attribute supplies a controllable prior.Source: §4.3 · p. 6
  • For numerical attributes such as free energy or enthalpy, a sparse Gaussian process models latent encodings against property values for Bayesian optimization.Source: §4.3 · pp. 6–7

5 EXPERIMENTS

GraphVF is evaluated for protein-pocket generation and controllable generation of sub-structures and molecular properties. Experiments compare molecule quality and geometric distributions with prior methods, then test targeted control.

5 EXPERIMENTS

  • Experiments use the CrossDocked dataset of 22.5 million protein-ligand pairs and compare GraphVF with GraphBP and Pocket2Mol using 100 generated molecules per pocket.The evaluation includes binding affinity, drug-related properties, novelty, diversity, and generation time.Source: §5.1 · p. 7

5.1 3D Molecular Generation Conditioned on Protein Pocket

  • GraphVF outperforms GraphBP and Pocket2Mol on HA, Lipinski, LogP, Novelty, and Diversity, and generates molecules two orders of magnitude faster than Pocket2Mol.It also shows reasonably good QED, LogP, and SA without explicit guidance from the variational prior.Source: §5.1 · p. 7, Table 1 · p. 8
  • Masking the 2D encoder and using a standard Gaussian prior drops HA from 31.1% to 26.3%, while the full model balances its 2D and 3D data sources.The ablation tests whether the KL term collapses the variational distribution toward a standard Gaussian.Source: §5.1 · p. 7

5.2 Sub-structure Analysis

  • GraphVF best matches the test-set dihedral-angle distributions, is comparable to GraphBP and Pocket2Mol on bond lengths and angles, and produces the closest CC-bond and CCCC-dihedral distributions.The ablation also reports that the 2D encoder helps capture sub-structure geometry.Source: §5.2 · p. 7, Table 2 · p. 8, Figure 6 · p. 9

5.3 Controllable Generation for Specified Chemical Sub-structures

  • Priors aggregated from training ligands containing specified motifs make GraphVF more likely to generate those sub-structures; its standard-prior rates are closest to test-set rates among the compared models.The motif cases include alkenyl, imine, five-membered sulfur rings, and six-membered oxygen rings.Source: §5.3 · pp. 7–8, Table 3 · p. 9

5.4 Controllable Molecular Generation for Specified Molecular Properties

  • Bayesian optimization selects latent encodings for five predicted QM properties, and GraphVF achieves higher energy than all baselines in the property-generation experiments.Property predictions use DimeNet, while a sparse Gaussian process fits the property-encoding relationship.Source: §5.4 · p. 9, Table 4 · p. 9

6 CONCLUDING REMARKS

GraphVF combines 2D structure semantics with 3D pocket geometry for controllable binding-molecule generation. The paper reports superior pocket-based generation performance and generation of molecules with desired ligand sub-structures and biochemical properties.

6 CONCLUDING REMARKS

  • GraphVF integrates 2D structure semantics and 3D pocket geometry, with reported superior performance to state-of-the-art pocket-based 3D molecule-generation strategies.The framework is also reported to generate binding molecules with desired ligand sub-structures and biochemical properties.Source: §6 · p. 9
  • The authors conclude that domain constraints can improve molecule-design quality and support finer-grained control over biochemical properties.They describe this as a potential for generating binding ligands using sophisticated domain knowledge.Source: §6 · p. 9

Appendices

The appendix describes parsing ligand graphs into canonical sub-structures and junction trees, then encoding whole molecules and their atom, bond, and geometric features in a latent Gaussian representation.

A.1 Parsing and pooling

  • The parsing procedure identifies rings, non-ring covalent atom pairs, and pivot atoms, yielding 427 canonical sub-structures from 100,000 CrossDocked reference ligands.Each sub-structure becomes a junction-tree node, with edges between intersecting sub-structures.Source: §A.1 · p. 11

A.2 Latent encoding

  • The molecule’s structural encoding is represented by two 34-dimensional vectors parameterizing a diagonal Gaussian, whose channels encode atom type, bond type, and three continuous geometric variables.The atom-type, bond-type, and (d, θ, ϕ) values are sampled from the first 27, next 4, and final 3 channels, respectively.Source: §A.2 · p. 11
  • The appendix reports that end-to-end training of the encoding parameters is preferable because a pretrained GRU led to serious degradation in generated-molecule quality.Source: §A.2 · p. 11

B IMPLEMENTATION DETAILS

Implementation details specify the model and training configuration, generation constraints, analysis discretization, and property-optimization procedure.

B IMPLEMENTATION DETAILS

  • GraphVF uses six Echnet layers, 20 tree-GRU layers, and six variational-flow layers, trained for 40 epochs with batch size 4.Training uses Adam with learning rate 1e-4 and weight decay 1e-6.Source: B IMPLEMENTATION DETAILS · p. 11
  • Generation samples 100 molecules per test pocket, resamples molecules below 15 atoms, truncates those above 50 atoms, and applies validity constraints on bonds, elements, and geometry.The constraints include allowed elements C, N, O, P, S, and Cl, and a maximum generated-bond length of 10 Å.Source: B IMPLEMENTATION DETAILS · p. 11
  • Sub-structure distributions are discretized into distance, angle, and dihedral bins, while motif-prior sampling drops training molecules with more than 16 atoms.Property optimization uses Bayesian optimization on one tenth of the training data; DimeNet predictions are regularized using the 3σ principle.Source: B IMPLEMENTATION DETAILS · p. 11

C ALGORITHMS FOR TRAINING AND GENERATION

The training and generation algorithms use receptor–ligand data and a junction-tree encoder to parameterize a latent prior, then generate ligand atoms and bonds autoregressively.

C ALGORITHMS FOR TRAINING AND GENERATION

  • The algorithm descriptions identify separate training and generation procedures for GraphVF.Their pseudocode is provided as Algorithms 1 and 2.Source: C ALGORITHMS FOR TRAINING AND GENERATION · pp. 12–13
  • The training algorithm samples receptor–ligand pairs, encodes the ligand with a junction-tree encoder, and uses its output to parameterize a Gaussian latent prior.The algorithm also dequantizes bond types and includes focal-classification and KL loss terms.Source: C ALGORITHMS FOR TRAINING AND GENERATION · p. 12
  • Generation takes receptors and optional property cues as input, initializes an empty ligand, and samples a focal atom at each step until none is eligible.The algorithm then appends the generated atom and its associated bonds to the ligand.Source: C ALGORITHMS FOR TRAINING AND GENERATION · p. 13

D CHEMICAL STRUCTURE ILLUSTRATION

The illustrations and procedures cover DMT binding, focal-atom selection, and conversion between spherical and Cartesian coordinates for generated atoms.

D CHEMICAL STRUCTURE ILLUSTRATION

  • Figure 7 compares serotonin and DMT structures and shows DMT bound to 5-HT2A, highlighting the NHMe2–Asp-231 interaction.Source: Figure 7 · p. 13

E.1 Focal atom selection

  • During generation, atoms whose focal scores exceed 0.5 are eligible, one is sampled randomly, and generation ends when none qualify.At the first step, focal scores are evaluated among receptor atoms; later steps evaluate known ligand atoms.Source: §E.1 · p. 14
  • Training uses teacher forcing: the nearest receptor atom is focal first, then the preceding atom in the expert trajectory, with labels scored by mean-reduced binary cross-entropy.Source: §E.1 · p. 14

E.2 Construction of the spherical coordinate system (SCS)

  • The spherical coordinate system expresses an atom’s position relative to focal atom f and its two nearest neighbors c and e, and can be converted back to Cartesian coordinates.The conversion is SE(3)-equivariant.Source: §E.2 · p. 14

E.3 Equivariance of the generation process

GraphVF’s generation probability is SE(3)-equivariant: rotating or translating the known structure correspondingly transforms the generated position distribution.

E.3 Equivariance of the generation process

  • At each generation step, the generation probability is equivariant to orthogonal transformations and translations of the known structure.Source: §E.3 · p. 14
  • The proof composes an SE(3)-invariant flow over spherical coordinates with their equivariant Cartesian reconstruction, yielding equivariant generated coordinates.The flow is parameterized by known atom types, bond types, and coordinates; both transformed expressions share the same latent distribution.Source: §E.3 · pp. 14–15

F RING-FIRST TRAVERSAL ALGORITHM

The ring-first traversal algorithm organizes molecular graph labels so atoms in the same ring or functional group remain consecutive, unlike the illustrated DFS labeling.

F RING-FIRST TRAVERSAL ALGORITHM

  • Ring-first labeling preserves local chemical semantics by keeping atoms from the same ring or functional group consecutive, whereas DFS can mix them.The figure illustrates the distinction between the desirable ring-first labeling and DFS labeling.Source: F RING-FIRST TRAVERSAL ALGORITHM · p. 15, Figure 8 · p. 15
  • The proposed minimal-ring Floyd algorithm repeatedly contracts minimal rings, applies DFS to the resulting tree, then expands rings to restore the original graph with consecutive labels.Source: F RING-FIRST TRAVERSAL ALGORITHM · p. 15
  • The traversal has O(n^5) time complexity, and Numba just-in-time compilation accelerates it by about 50×.The authors report improving training with ring-first traversal without losing efficiency.Source: F RING-FIRST TRAVERSAL ALGORITHM · p. 15
Loading 2304.12825v1…