Source-linked AI summary
Generating 3D Molecules Conditional on Receptor Binding Sites with Deep Generative Models
Matthew Ragoza, Tomohide Masuda, David Ryan Koes
TL;DR
The paper addresses the unresolved problem of generating 3D molecules directly from protein binding pockets. It uses a receptor-conditioned variational autoencoder over atomic density grids and reports mutation-sensitive outputs, substantial validity and stability, and favorable binding-related properties, while noting representation and energetic limitations.
Problem
Generating 3D molecular structures directly from protein binding pockets remains an unsolved challenge despite prior receptor-conditioned SMILES generation and affinity-guided 3D sampling.
Method
A conditional variational autoencoder learns to sample ligand density distributions conditioned on receptor binding-site density, with generated densities converted into molecules through atom fitting and bond inference.
Results
All multi-residue mutations and some important single-residue mutations significantly changed generated-molecule properties, with Arg136 affecting all assessed relevant properties for posterior and prior molecules.
Takeaways & Limitations
The results demonstrate receptor-conditioned 3D molecular generation and support using mutated-receptor responses to assess whether the model conditions its outputs on receptor structure.
Takeaways & Limitations
Generated molecules can have unstable internal geometries because the model lacks explicit bond information and relies on local density maxima for atom locations.
Abstract
from arXiv · showhide
The goal of structure-based drug discovery is to find small molecules that bind to a given target protein. Deep learning has been used to generate drug-like molecules with certain cheminformatic properties, but has not yet been applied to generating 3D molecules predicted to bind to proteins by sampling the conditional distribution of protein-ligand binding interactions. In this work, we describe for the first time a deep learning system for generating 3D molecular structures conditioned on a receptor binding site. We approach the problem using a conditional variational autoencoder trained on an atomic density grid representation of cross-docked protein-ligand structures. We apply atom fitting and bond inference procedures to construct valid molecular conformations from generated atomic densities. We evaluate the properties of the generated molecules and demonstrate that they change significantly when conditioned on mutated receptors. We also explore the latent space learned by our generative model using sampling and interpolation techniques. This work opens the door for end-to-end prediction of stable bioactive molecules from protein structures with deep learning.
1 Introduction
Structure-based drug discovery seeks molecules that bind specific protein targets, but directly generating 3D molecules from binding pockets remains unresolved. This work introduces receptor-conditional deep generative modeling and examines mutation-conditioned outputs and latent-space structure.
- Drug discovery searches chemical space for compounds that bind specific biological targets, motivating computational methods that sample and screen candidates in silico.
- Deep learning has generated molecules with drug-like properties and 3D conformations, but prior approaches generally used 2D representations or generated conformers without directly conditioning on protein binding pockets.
- Distance-matrix approaches require triangle-inequality enforcement and cubic-time Hungarian matching, whereas density grids are permutation invariant and avoid expensive matching algorithms.
- Atomic density grids preserve holistic molecular shape information, but converting generated densities into unambiguous discrete atoms and bonds remains a central challenge.
- The paper demonstrates receptor-conditional 3D molecular generation, evaluates conditioning on mutated receptors, and explores latent-space sampling and interpolation.
2 Methods
The method represents molecular atoms through property-based type vectors and processes docked complexes with a conditional variational autoencoder pipeline. The pipeline encodes complex and receptor densities, samples a latent representation, and decodes a ligand density into a 3D structure.
- Atoms are typed using element, aromaticity, hydrogen-bond donor and acceptor status, and formal charge properties.
- The pipeline converts docked protein-ligand complexes into atomic density grids before conditional variational autoencoding.
- The CVAE encodes full complex density through its input branch and receptor density through its conditional branch.
- A sampled latent vector is combined with the receptor-conditioned vector, and the decoder converts them into ligand density for atom fitting and bond addition.
2.2 Atomic density grids
Atomic density grids represent molecules as property-specific spatial fields generated by summing truncated Gaussian atom kernels. The implementation uses finite cubic grids with fixed resolution and randomized spatial transformations.
- Atoms are represented as continuous densities with a truncated Gaussian kernel evaluated from atom-to-grid-point distance and atomic radius.
- Grid values sum each atom’s kernel contribution, weighted by the atom’s type-vector value in the corresponding channel.
- Molecules are represented by atom-type and coordinate matrices, which the gridding function maps to multi-channel 3D density grids.
- The grids use 23.5 Å cubic side lengths at 0.5 Å resolution, producing spatial dimensions NX = NY = NZ = 48.
- Training and evaluation apply random translations and rotations after centering grids on the input molecule.
2.3 Atom fitting algorithm
The model generates ligand density grids with a CVAE, then converts those densities into discrete 3D molecules through atom fitting and bond inference. Its training objective combines reconstruction, prior-matching, and steric-clash terms.
- Atom fitting algorithm: The inverse conversion from a reference density grid to a discrete 3D structure is formulated as an optimization problem.The objective minimizes the squared difference between the reference grid and the grid generated from fitted atoms and coordinates.
- Atom fitting algorithm: Initial atom locations are detected at grid points with the largest density values, then refined by gradient descent against the density representation.The procedure uses libmolgrid to backpropagate grid-value gradients to atomic coordinates.
- Atom fitting algorithm: Bond inference adds bonds and hydrogens to fitted atoms while applying atom-type constraints to construct valid molecules.The procedure uses customized OpenBabel bond-perception routines.
- Conditional variational autoencoder: The CVAE learns p(lig|rec) by sampling a latent interaction variable and decoding it with a receptor encoding, while an input encoder approximates the latent posterior.The decoder generates ligand density grids conditioned on receptor information.
- Conditional variational autoencoder: Training combines reconstruction, KL-divergence, and steric-clash losses to produce realistic ligand densities with reduced receptor overlap.Steric clash is measured from overlap between generated ligand and receptor densities; real complexes were empirically checked to lack such overlap.
2.6 Training data set
Training data came from CrossDocked2020, which expands structural examples through cognate and non-cognate docking while retaining pocket-similarity splits and quality filters.
- Training data set: CrossDocked2020 contains 18,450 bound protein-ligand crystal structures that were clustered by pocket similarity and combinatorially docked.Each ligand was re-docked to its known receptor and cross-docked to receptors with similar pockets.
- Training data set: The first pocket-similarity split was used to construct training and test sets, excluding poses above 2 Å RMSD from the cognate crystal pose.Molecules that could not be sanitized with RDKit were also omitted.
2.7 Test target selection
The test set comprised ten targets selected from distinct pocket clusters, each with at least five unique ligands and evaluated using top-ranked docked poses.
- Test target selection: Ten CrossDocked2020 test targets were randomly selected, with each target drawn from a different pocket cluster.Only targets containing at least five unique ligands were considered.
- Test target selection: The evaluation used the top-ranked docked pose for each ligand associated with the selected targets.The selected proteins and associated ligands are listed in Table 2.
2.8 Sampling methods
Sampling combines latent variables with receptor encodings and supports posterior, prior, and interpolated bias-controlled modes. Each evaluated method generated 100 samples per test complex.
- Sampling methods: Posterior sampling encodes a real protein-ligand complex before drawing latent samples, whereas prior sampling draws from a standard normal distribution without reference-ligand bias.Both modes combine the latent vector with the conditional receptor encoding before decoding a ligand density.
- Sampling methods: The variability factor λvar scales the standard deviations used to sample the latent space.This setting controls the level of sampling variance.
- Sampling methods: The bias factor λbias controls reference-molecule influence by linearly interpolating between prior and posterior distribution parameters.This provides an intermediate sampling mode rather than selecting only prior or posterior sampling.
- Sampling methods: 100 samples were generated for each protein-ligand complex for every evaluated sampling method.
2.9 Evaluation metrics
The evaluation defines validity, novelty, and uniqueness using connectedness, RDKit sanitization, training-set membership, and duplicate-generation criteria.
- Validity requires a single connected fragment that RDKit can sanitize while checking valency and attempting aromatic-bond Kekulization.
- Novelty requires that a molecule’s canonical SMILES string was absent from the training set.
- Uniqueness requires that a molecule’s canonical SMILES string was not generated previously during test evaluations.
3 Results
Generated molecules were generally novel and diverse, with posterior samples more valid and reference-like than prior samples. Sampling controls altered diversity and similarity, while receptor mutations changed generated structures and predicted interaction properties.
- Properties of generated molecules: 98.5% of posterior molecules and 90.9% of prior molecules were valid, while 100% of all generated molecules were novel.
- Properties of generated molecules: Posterior molecules averaged 0.33 Tanimoto similarity to the reference, whereas prior molecules were highly dissimilar, with 25% exceeding 0.15 similarity.
- Properties of generated molecules: 30.8% of posterior and 17.3% of prior molecules had lower minimized Vina energy than the reference; 15.4% and 15.9%, respectively, had higher predicted affinity.
- Properties of generated molecules: Generated molecules showed similar median bond angles but greater angle variance and prevalent small strained angles, especially for carbon-oxygen-carbon bonds.
- Controlling sampling variability and bias: Increasing posterior variability increased diversity, reduced reference similarity, and tended to produce less energetically stable molecules.
- Controlling sampling variability and bias: For prior samples, increasing variability increased molecular weight, complexity, and energy without a relationship to reference similarity.
- Controlling sampling variability and bias: Increasing the bias factor shifted generated molecules toward the posterior, increasing reference similarity and moving molecular-weight, RMSD, Vina-energy, and affinity distributions toward posterior values.
- Conditioning on mutated receptors: Mutating receptor residues changed generated molecules and their predicted properties, with Arg136 affecting similarity, predicted affinity, and Vina-energy change, while Arg58-to-alanine showed no significant differences.
4 Discussion
The model generates valid, novel, and unique 3D molecules conditioned on receptor binding pockets, and its outputs respond to receptor mutations. Generated molecules show energetic and grid-representation limitations that motivate future improvements.
- Discussion: Over 90% of generated molecules were valid, novel, and unique, while over 80% moved less than 2 Å RMSD after UFF minimization.The authors note that validity, novelty, and uniqueness do not fully evaluate 3D conformational quality; the RMSD result indicates energetic stability within the binding site.
- Discussion: A significant number of generated molecules had lower Vina energy and higher predicted binding affinity than the reference molecule.
- Discussion: Generated molecules tended to have more strain than real molecules, so their internal bond lengths and angles were relaxed with UFF.The authors attribute this probably to the model’s lack of explicit bond information and propose adding an energetic term to training.
- Discussion: Conditioning on shikimate kinase mutants produced chemically relevant changes, with all multi-residue mutations and some single-residue mutations significantly altering generated-molecule properties.Arg136 mutations significantly changed all relevant properties, while effects varied for Asp34 and Arg58 substitutions and charge inversions.
- Discussion: Generation quality was sensitive to density-grid hyperparameters, including grid resolution, atomic radius, and random-rotation augmentation.Overlapping density peaks can make nearby atoms difficult to resolve, while smaller kernels can improve atom distinction and chemical realism.
- Discussion: Future work proposes training setups that emphasize the conditional receptor, including independent rotations and receptor variation within binding-pocket clusters.The authors also propose using higher-RMSD ligand poses with lower-RMSD labels to perform instantaneous minimization for docking.
- Discussion: The project provides code to support reproduction and extension of the approach.
Conflicts of interest
The authors report no conflicts of interest.
- Conflicts of interest: The authors declare that there are no conflicts of interest.
Supplementary Materials: Generating 3D Molecules Conditional on Receptor Binding Sites with Deep Generative Models
The supplementary materials document atom typing, atom fitting, bond inference, and evaluations of generated-molecule properties. They also examine diversity, shape similarity, drug-likeness, geometry, variability, prior–posterior bias, and conditioning on mutant receptors.
- Supplementary Materials: The supplementary materials identify the authors and their University of Pittsburgh affiliations.
- Atom typing scheme: The atom typing scheme represents element, formal charge, aromaticity, hydrogen-bond acceptor and donor status, and bonded-hydrogen count, using placeholders for out-of-range values.
- Bond inference algorithm: Bond inference builds a connected molecule by adding nearby bonds and hydrogens, removing invalid valences and strained geometry, and assigning hybridization, aromaticity, and bond orders.
- Atom fitting algorithm: Atom fitting iteratively adds candidate atoms at high-density grid points and optimizes their coordinates by gradient descent against a reference density.
- Supplementary evaluations: The supplementary figures compare generated and real molecules across atomic-property distributions, shape similarity, drug-likeness, energy changes, bond lengths, bond angles, and torsion angles.
- Supplementary evaluations: Figure S3 measures per-target diversity as the inverse of expected Tanimoto fingerprint similarity, while Figures S7–S9 compare minimized geometric distributions.
- Latent-space controls: Figures S10–S12 vary posterior or prior variability factors and interpolate between prior and posterior distributions to examine controllability and reference-molecule bias.
- Mutant-receptor conditioning: Figures S13 and S14 evaluate posterior and prior molecules conditioned on mutant receptors using Vina energy and CNN affinity relative to the conditional receptor.