Leveraging binding-site structure for drug discovery with point-cloud methods

Vincent Mallet, Carlos Oliver, Nicolas Moitessier, Jérôme Waldispühl

Abstract

Computational drug discovery strategies can be broadly placed in two categories: ligand-based methods which identify novel molecules by similarity with known ligands, and structure-based methods which predict molecules with high-affinity to a given 3D structure (e.g. a protein). However, ligand-based methods do not leverage information about the binding site, and structure-based approaches rely on the knowledge of a finite set of ligands binding the target. In this work, we introduce TarLig, a novel approach that aims to bridge the gap between ligand and structure-based approaches. We use the 3D structure of the binding site as input to a model which predicts the ligand preferences of the binding site. The resulting predictions could then offer promising seeds and constraints in the chemical space search, based on the binding site structure. TarLig outperforms standard models by introducing a data-alignment and augmentation technique. The recent popularity of Volumetric 3DCNN pipelines in structural bioinformatics suggests that this extra step could help a wide range of methods to improve their results with minimal modifications.

Introduction

In in silico drug discovery, researchers typically choose between ligand-based methods, which prioritize chemical similarity to known actives, and structure-based methods, which utilize 3D protein shapes for docking. While ligand-based approaches are computationally efficient, they ignore the target protein's structural context. Conversely, structure-based methods are computationally intensive and often constrained by known chemical scaffolds. A significant challenge remains in leveraging 3D protein structures to guide the generation of novel chemical entities. This is complicated by the fact that 3D protein structures lack a natural orientation, making it difficult for standard neural networks to extract features efficiently.

Methods: The TarLig Pipeline

The authors propose TarLig, a deep learning framework that treats protein binding sites as 3D point clouds. The core innovation is the use of a "latent space bridge." Instead of predicting specific molecular structures directly, TarLig predicts a latent vector within a pre-trained Variational Autoencoder (VAE) space. This VAE maps millions of SMILES strings into a continuous, high-dimensional numerical representation.

Overview of the TarLig pipeline, showing the transition from protein structure to latent ligand space.

To address the rotation-invariance problem, TarLig employs Principal Component Analysis (PCA) to align every binding site to its own principal axes. Since PCA axes lack fixed directions, this results in eight possible orientations for any given pocket, which are used for data augmentation. The architecture utilizes a Volumetric 3D Convolutional Neural Network (3DCNN) and a Siamese setup to force the network to learn rotation-invariant representations.

Datasets and Evaluation

The model was trained on 30,964 protein structures and 3,362 unique ligands extracted from the Protein Data Bank (PDB). For benchmarking, the authors used the Directory of Useful Decoys, Enhanced (DUD-E), which contains 102 targets with associated active compounds and decoys.

Results

TarLig demonstrated significant improvements over baseline models. It achieved a Mean Square Error (MSE) of 0.085 on the test set, compared to a shuffled baseline of 0.210. Notably, the PCA-based alignment (USRPCAT) outperformed standard shape-matching tools (USRCAT) by approximately 9% in Enrichment Factor (EF) on the DUD-E benchmark.

Enrichment Factor (EF) performance across various actives cutoffs for the TarLig model.

The model also proved highly efficient, requiring only 1 to 6 hours of training on a single GPU, whereas steerable CNN alternatives required over 35 hours. Furthermore, the model performed well even for ligands with very few training examples, suggesting it successfully learned structural features rather than simple memorization.

Discussion and Future Work

While TarLig excels at identifying the general chemical neighborhood of active compounds, distinguishing between true actives and chemically similar decoys remains a challenge. Future work will focus on training embeddings specifically to separate compounds based on binding affinity. Additionally, integrating self-supervised learning on broader protein structural data and directly linking predicted latent vectors to generative models could further streamline the discovery of tailored drug candidates.

Summary updated

Abstract

Computational drug discovery strategies can be broadly placed in two categories: ligand-based methods which identify novel molecules by similarity with known ligands, and structure-based methods which predict molecules with high-affinity to a given 3D structure (e.g. a protein). However, ligand-based methods do not leverage information about the binding site, and structure-based approaches rely on the knowledge of a finite set of ligands binding the target. In this work, we introduce TarLig, a novel approach that aims to bridge the gap between ligand and structure-based approaches. We use the 3D structure of the binding site as input to a model which predicts the ligand preferences of the binding site. The resulting predictions could then offer promising seeds and constraints in the chemical space search, based on the binding site structure. TarLig outperforms standard models by introducing a data-alignment and augmentation technique. The recent popularity of Volumetric 3DCNN pipelines in structural bioinformatics suggests that this extra step could help a wide range of methods to improve their results with minimal modifications.
BibTeX
@misc{mallet2019leveraging,
  author = {Vincent Mallet and Carlos Oliver and Nicolas Moitessier and Jérôme Waldispühl},
  title = {{Leveraging binding-site structure for drug discovery with point-cloud methods}},
  year = {2019},
  eprint = {1905.12033},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.12033},
}
Figure