Source-linked AI summary
Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets
Xingang Peng, Shitong Luo, Jiaqi Guan, Qi Xie, Jian Peng, Jianzhu Ma
TL;DR
Pocket-conditioned molecular generation must satisfy geometric constraints while preserving detailed chemical structure, which earlier graph- or coordinate-focused methods incompletely address. Pocket2Mol combines an E(3)-equivariant pocket model with autoregressive prediction and tractable conditional coordinate sampling without MCMC. Experiments report better binding affinities, drug-related properties, and structural realism than prior methods.
Problem
Pocket-based molecular generation must satisfy multiple pocket-imposed geometric constraints while retaining chemical details such as bond types and functional groups.
Method
Pocket2Mol uses an E(3)-equivariant generative network that models pocket geometry and bonding, jointly predicts molecular components, and samples coordinates from a tractable conditional distribution without MCMC.
Results
Pocket2Mol-generated molecules show higher binding affinity and drug-related properties, with more realistic substructures than state-of-the-art models.
Takeaways & Limitations
Pocket2Mol provides an efficient pocket-conditioned approach for generating 3D drug candidates with improved reported affinities, chemical properties, and structural realism.
Takeaways & Limitations
The substructure analysis assumes that sampled molecules should preserve distributions of basic functional groups and substructures from the test set.
Abstract
from arXiv · showhide
Deep generative models have achieved tremendous success in designing novel drug molecules in recent years. A new thread of works have shown the great potential in advancing the specificity and success rate of in silico drug design by considering the structure of protein pockets. This setting posts fundamental computational challenges in sampling new chemical compounds that could satisfy multiple geometrical constraints imposed by pockets. Previous sampling algorithms either sample in the graph space or only consider the 3D coordinates of atoms while ignoring other detailed chemical structures such as bond types and functional groups. To address the challenge, we develop Pocket2Mol, an E(3)-equivariant generative network composed of two modules: 1) a new graph neural network capturing both spatial and bonding relationships between atoms of the binding pockets and 2) a new efficient algorithm which samples new drug candidates conditioned on the pocket representations from a tractable distribution without relying on MCMC. Experimental results demonstrate that molecules sampled from Pocket2Mol achieve significantly better binding affinity and other drug properties such as druglikeness and synthetic accessibility.
1. Introduction
Pocket-based drug design addresses the need to generate molecules that satisfy protein-pocket geometry, beyond earlier representations and sampling strategies. Pocket2Mol combines an E(3)-equivariant model with efficient conditional sampling and chemical-bond generation.
- Prior approaches: Earlier generative models represented molecules as 1D descriptors, 2D graphs, or 3D structures and sampled candidates by perturbing compact latent representations.
- Motivation: Protein-pocket geometry is important because molecules act by binding specific pockets, whereas pocket-agnostic design may reduce downstream cell-experiment success.
- Pocket2Mol: Pocket2Mol develops an E(3)-equivariant generative model with a geometric pocket network, efficient conditional coordinate sampling, and chemical-bond sampling.
- Pocket2Mol: The model learns chemical and geometric pocket constraints, jointly predicts frontier atoms, positions, atom types, and bonds, and avoids conventional MCMC through tractable relative-coordinate distributions.
- Results: Pocket2Mol is reported to generate candidates with higher binding affinity, drug-likeness, and more realistic substructures than state-of-the-art models.
2. Related Work
Prior pocket-based molecular-generation methods range from SMILES and graph generation to 3D approaches using voxelized densities or search-based optimization. Related work also highlights challenges in rotational equivariance and efficient atom-position prediction.
- Pocket-conditioned generation: Some pocket-conditioned methods generate only SMILES strings or 2D molecular graphs, limiting their direct treatment of three-dimensional molecular structure.
- Pocket-conditioned generation: Li et al. use 3D pockets to evaluate candidates and guide Monte Carlo Tree Search rather than incorporating pocket structures directly into training.
- 3D generation: Masuda et al. generate voxelized atomic densities with a 3D CNN that is not rotation-equivariant and is difficult to scale to large protein pockets.
- Equivariant networks: Vector neurons and geometric vector perceptrons support equivariant processing of scalar and vector features, addressing limitations of architectures built for equivariant layer features.
- Atom-position generation: Autoregressive atom-position methods either predict long-range distances, which introduce errors, or use local spherical coordinates, whose Euclidean transformation is inefficient and nonstraightforward.
3. Method
Pocket2Mol models protein pockets and molecular fragments with an E(3)-equivariant network, then generates molecules autoregressively by predicting frontier atoms, positions, elements, and bonds. Training masks molecular atoms and teaches the model to recover their locations and chemical structure from pocket and fragment context.
- Generation Procedure: Pocket2Mol learns context-specific atom and bond distributions autoregressively from existing molecular fragments.The model predicts the randomly masked part of a training molecule from the remaining structure.
- Generation Procedure: Generation selects frontier atoms, samples a focal atom, predicts a new atom’s relative position, then samples its element and bonds.Frontiers are atoms that can covalently connect to new atoms; generation stops when no frontier remains.
- E(3)-Equivariant Neural Network: The E(3)-equivariant architecture uses scalar and vector features with GVP and vector-neuron blocks to capture pocket geometry and predict atom positions.Vector features enable direct geometric position prediction while preserving equivariance.
- Encoder: Pocket and fragment atoms are represented in a KNN graph with chemical, residue, bond, and geometric features propagated through message-passing and update modules.Protein features include element, amino-acid, and backbone or side-chain identity; molecular features include element, valence, and bond counts.
- Position Predictor: Position prediction represents Gaussian-component means, covariance vectors, and mixture probabilities using separate GVL blocks.Because their vector outputs are equivariant, GVL blocks directly represent the mean and covariance vectors.
- Element-and-Bond Predictor: Bond prediction uses separate scalar and vector attention pathways, with vector attention based on Frobenius inner products and geometry-aware edge biases.The bias incorporates the relationship between neighboring edges to capture chemical and geometric constraints such as valence compatibility.
- Training: Training randomly masks molecule atoms and recovers their positions, element types, and bonds, using protein atoms as frontiers when all molecular atoms are masked.Negative ambient-space positions are assigned the additional Nothing element label during element prediction.
4. Results
Pocket2Mol outperforms CVAE and AR on general molecular properties and produces structures whose chemical and geometric patterns more closely match the data. Its generated molecules also show realistic ring distributions and low RMSD, although diversity is lower than another method.
- General properties: Pocket2Mol outperforms CVAE and AR on binding affinity, drug potential, and similarity to the training set.The reported metrics include Vina Score, High Affinity, QED, SA, LogP, Lipinski, and training-set similarity.
- General properties: Pocket2Mol has lower diversity than another method, while diversity does not account for constraints introduced by the protein pocket.The paper distinguishes this metric from pocket-conditioned structural quality.
- Sub-structure analysis: Pocket2Mol produces ring-size ratios more similar to the test set than CVAE and AR, including the common five- and six-membered rings.The authors attribute the improvement to jointly learning atom and bond distributions and predicting bonds during generation.
- Sub-structure analysis: Pocket2Mol-generated molecules have lower KL divergence for bond-angle and dihedral-angle distributions than the other methods.The result indicates closer agreement with the geometric attributes of the data.
- Sub-structure analysis: Pocket2Mol generally produces the lowest RMSD between generated structures and RDKit-predicted conformers.For each molecule, 20 RDKit conformers were sampled and the minimum RMSD was selected.
5. Conclusion
Pocket2Mol generates 3D drug-like molecules conditioned on protein pockets using an E(3)-equivariant network and an efficient sampling algorithm. Generated molecules show better affinities, chemical properties, and structural realism.
- Pocket2Mol combines a graph neural network for 3D pocket features with an efficient pocket-conditioned algorithm for sampling 3D drug candidates.The model captures chemical and geometrical features of protein pockets.
- Generated molecules show better affinities and chemical properties than the evaluated alternatives.
- Generated molecules also contain more realistic and accurate structures.
Software
The implementation uses Ubuntu Linux with V100 GPUs and is written mainly in Python 3.8 with PyTorch 1.9.0. The code is provided as supplementary material.
- Experiments run on Ubuntu Linux using V100 GPUs.
- The code is implemented mainly with Python 3.8 and PyTorch 1.9.0.
- The authors uploaded the code as supplementary material.
A. The formula of the original geometric vector perceptron
The original geometric vector perceptron maps scalar-vector pairs to scalar-vector pairs using learnable transformations and nonlinearities. The formulation uses row-wise norms, Hadamard products, sigmoid activation, and Leaky ReLU in this model.
- The GVP takes a scalar-vector pair as input and outputs another scalar-vector pair.It is denoted as (x′, vector x′) = GVP(x, vector x).
- The formulation uses learnable parameters, row-wise norms, and Hadamard products.
- The model uses sigmoid and Leaky ReLU activations in the GVP formulation.Leaky ReLU is the chosen arbitrary activation function.
B. The hyper-parameters of Pocket2Mol and training
Pocket2Mol uses specified hidden and edge-feature dimensions across its encoder and predictors, with six message-passing layers. Training uses batch size 8, learning-rate decay on validation stagnation, and 475,000 total iterations.
- The encoder uses 6 message-passing layers with 256 scalar and 64 vector hidden dimensions.Edge embeddings use 64 scalar and 64 vector dimensions.
- The frontier predictor uses 128 scalar and 32 vector hidden dimensions.
- The position predictors use 128 scalar and vector hidden dimensions.
- Training uses batch size 8 and an initial learning rate of 2 × 10^-4.
- The learning rate is multiplied by 0.6 when validation loss fails to decrease for 8 validation iterations, with validation every 5000 iterations.
- Training runs for 475,000 total iterations.
C. The Proof of the equivariant of the vector-attention module
The vector-attention module generalizes the vector inner product to a Frobenius inner product while preserving E(3)-equivariance. Under rotations or reflections, attention scores remain invariant and the module outputs remain equivariant.
- The module generalizes the vector inner product of queries and keys to a Frobenius inner product.
- Queries, keys, and values are vector features processed by the vector-attention module A.
- Under any rotation or reflection matrix R ∈ R3×3, the attention scores calculated by Eq. 9 are invariant.
- Because the attention scores are invariant, the attention-module output vectors remain equivariant under the transformation.