Source-linked AI summary
Structure-based drug design with geometric deep learning
Clemens Isert, Kenneth Atz, Gisbert Schneider
TL;DR
Structure-based drug design needs reliable use of three-dimensional macromolecular information, and geometric deep learning offers symmetry-aware models for this setting. This review surveys representations and applications across property prediction, binding-site and pose prediction, and de novo design, while identifying generalization and experimental-validation challenges. The reviewed methods demonstrate broad potential, including generation of a novel MDM2-targeting ligand, but practical utility still requires stronger benchmarking and experimental testing.
Problem
Structure-based drug design requires methods that use detailed 3D macromolecular and ligand-interface information across prediction and design tasks.
Method
The review synthesizes symmetry-aware geometric deep-learning methods using 3D macromolecular representations for property prediction, binding-site and pose prediction, and de novo ligand design.
Results
The reviewed approaches span predictive and generative structure-based tasks, including a novel MDM2-targeting ligand DF-1 with Euclidean distance 0.48 to its closest ChEMBL molecule.
Takeaways & Limitations
Geometric deep learning provides a framework for structure-based drug discovery and design, but its real-world utility depends on meaningful benchmarking and experimental validation.
Takeaways & Limitations
Many models trained on PDBbind may memorize training data, show poor generalization, and perform similarly to ligand-only or protein-only descriptors.
Abstract
from arXiv · showhide
Structure-based drug design uses three-dimensional geometric information of macromolecules, such as proteins or nucleic acids, to identify suitable ligands. Geometric deep learning, an emerging concept of neural-network-based machine learning, has been applied to macromolecular structures. This review provides an overview of the recent applications of geometric deep learning in bioorganic and medicinal chemistry, highlighting its potential for structure-based drug discovery and design. Emphasis is placed on molecular property prediction, ligand binding site and pose prediction, and structure-based de novo molecular design. The current challenges and opportunities are highlighted, and a forecast of the future of geometric deep learning for drug discovery is presented.
1 Introduction
Structure-based drug design uses 3D macromolecular information, while geometric deep learning incorporates symmetry-aware neural architectures to support predictive and generative tasks. The review organizes methods around molecular representations, symmetry operations, and applications including property prediction, binding-site and pose prediction, and de novo design.
- Structure-based drug design leverages 3D structures of proteins and nucleic acids to rationalize ligand–macromolecule interactions.
- Geometric deep learning processes 3D molecular structures with neural architectures that incorporate symmetry information.
- The review covers molecular property prediction, binding-site and interface prediction, binding-pose generation, and structure-based de novo ligand design.
- Molecular representation: Macromolecular inputs are commonly represented as 3D grids, surfaces, or graphs, each with distinct geometric and neighborhood structures.
- Symmetry: Equivariance transforms outputs consistently with input transformations, whereas invariance preserves outputs under transformations such as rotation, translation, or reflection.
2 Molecular property prediction
Geometric deep-learning methods predict molecular and macromolecular properties by combining 3D grids, surfaces, graphs, coordinates, and ligand information. Architectures use geometric features such as distances, angles, edge types, and symmetry-aware message passing for affinity, function, activity, and structural predictions.
- Property-prediction methods estimate scalar quantities such as binding affinity, protein function, and docking-pose scores from macromolecular structures.
- Grid-based methods: 3D-grid methods use CNNs with voxelized protein–ligand features, while rotational limitations can be addressed through training-time 90° rotation augmentation.
- Multilevel representations: HoloProt combines sequence-, surface-, and structure-level representations with ligand graphs through multilevel message passing for affinity and protein-function prediction.
- Graph-based methods: Graph-based approaches encode distances, angles, edge types, or distance-based connectivity and often achieve translation- and rotation-invariant processing.
- Additional tasks: Geometric models also predict quantum-derived bond orders, protein force fields, protein structural deviation, and activity from 3D molecular information.
3 Binding site/interface prediction
Binding-site and interface prediction identifies regions of macromolecules that interact with drug-like ligands or other macromolecules. Geometric deep-learning methods address this task with grid, surface, and graph representations using pharmacophoric, geodesic, and relational features.
- Binding-site prediction targets regions that accommodate small drug-like ligands, while interface prediction targets interactions between macromolecules.
- Grid-based methods: DeepSite represents proteins as pharmacophore-annotated 3D voxel grids and predicts whether local subgrids lie near druggable binding sites.
- Grid-based methods: RNet extends grid-based binding-site prediction to RNA structures.
- Surface-based methods: MaSIF and dMaSIF use geodesic macromolecular surfaces for binding-site prediction and pocket-function classification.
- Graph-based methods: Graph-based methods use distances, angles, and relational features with spatial convolutions, graph transformers, or E(3)-invariant filters to predict interaction potentials and binding sites.
4 Binding pose generation/molecular docking
Binding-pose generation methods predict suitable conformations between ligands and macromolecular receptors or between two macromolecules. Equivariant and hybrid architectures estimate rigid-body transformations, ligand torsions, surface complementarity, and pose confidence.
- Docking methods generate binding conformations for small-molecule–receptor pairs or pairs of macromolecular structures.
- Equivariant docking: EquiDock uses SE(3)-equivariant message passing and optimal transport to predict rigid-body rotation and translation for blind protein–protein docking.
- Flexible ligand docking: EquiBind extends equivariant docking to flexible small-molecule ligands by modeling torsion-angle changes from randomly generated conformers.
- Graph-based and hybrid methods: Hybrid methods combine protein surfaces or meshes with molecular graphs to estimate shape complementarity or small-molecule binding poses.
- Graph-based and hybrid methods: The described approach uses residue-level and full-atomistic 3D graphs, with SE(3)-equivariant pose outputs and SE(3)-invariant torsion and confidence predictions.It substantially outperformed existing classical and deep-learning approaches on a common docking benchmark.
5 De novo design
Structure-based de novo design uses geometric deep learning to generate molecular structures directly from macromolecular binding sites. Recent approaches construct 3D molecular graphs or use equivariant diffusion, while their practical utility remains to be established.
- De novo design generates new molecular structures with desired biological and physical properties from scratch.
- Chemical language models commonly learn string-based molecular representations such as SMILES for de novo drug design.Ligand-based models have generated molecules with desired physicochemical and biological properties.
- Structure-based models generate potential ligands directly within macromolecular binding sites as 3D graphs.These models sample atoms sequentially from learned distributions and have been applied to multiple molecular properties.
- E(3)-equivariant diffusion models generate 3D molecular graphs by denoising normally distributed point sets, including within macromolecular binding sites.DiffSBDD and TargetDiff implement generation from scratch in binding sites, while DiffLinker connects fragments placed in the pocket.
- Although 3D graph-based methods can construct many novel molecules, their practical applications remain to be explored.
6 Outlook
The outlook emphasizes physics- and symmetry-informed models, stronger benchmarking for binding-affinity prediction, and experimental validation of generated structures. These developments are intended to improve assessment of generalization and real-world utility.
- Incorporating physics and symmetry into models tends to increase prediction accuracy, generalizability, and interpretability.
- Binding-affinity methods must address evidence that some PDBbind-trained architectures memorize training data and generalize poorly.
- Suitable benchmarking datasets and realistic train-test splits can assess binding-affinity generalization in lead-optimization scenarios.PDE10A inhibitor data with binding affinities and X-ray co-crystal structures is cited as one example.
- Experimental validation is paramount for comprehensively evaluating emerging generative models in real-world drug design.Collaboration with experimentalists is valuable when computational groups lack the expertise, equipment, or desire for synthesis and testing.
9 Structure-based de novo design of DF-1
DF-1 is presented as a ligand designed with a geometric deep-learning method for MDM2 and docked into the human MDM2 active binding site. Its highest-ranking docking pose is shown.
- DF-1 was designed using a geometric deep-learning method to target MDM2.
- DF-1 was docked into the active binding site of human MDM2 using GOLD software.The human MDM2 structure is identified by PDB-ID 4JRG.
- The highest-ranking docking pose of DF-1 is shown in Figure 1.