Overview
Chemistry is the science of transformation: breaking and forming bonds to turn one substance into another. Materials science extends this logic to the solid state, asking which arrangements of atoms yield the properties we need. Both fields share a fundamental computational challenge. The chemical space of drug-like molecules alone contains an estimated $10^{60}$ members, and the space of possible crystal structures is similarly vast. No brute-force search can traverse these spaces. Discovery AI offers a different approach: learn the structure of chemical space from data, then generate, evaluate, and optimize candidates in a closed loop.
This chapter covers four interconnected capabilities. First, we build molecular generators that propose novel compounds with desired properties, using techniques from de novo design through equivariant diffusion models. Second, we fold proteins together with candidate ligands using Boltz-1 and Chai-1, predicting binding poses without expensive docking simulations. Third, we design crystalline materials and predict their properties using universal force fields (MACE-MP-0, CHGNet) that generalize across the periodic table. Fourth, we assemble these components into a closed-loop drug discovery campaign that integrates structure-based generation, cofolding, Bayesian optimization, and provenance tracking.
The generative model foundations from Chapter 34 provide the diffusion and flow-matching theory underlying molecular generators. The optimization machinery from Chapter 45 drives our acquisition loops. The provenance system from Chapter 47 ensures every generated molecule, predicted structure, and optimization decision is fully traceable. And the biology chapter (Chapter 48) provides the protein structure prediction context that ligand cofolding builds upon.
Prerequisites
This chapter assumes familiarity with molecular representations (SMILES strings, molecular graphs), basic organic chemistry concepts (functional groups, rings, stereochemistry), and the generative modeling techniques from Chapter 34 (diffusion models, score matching). For materials sections, basic knowledge of crystal lattices and unit cells is helpful. All code uses RDKit for molecular manipulation and PyTorch for model inference. The Bayesian optimization sections build on Chapter 45.
Learning Outcomes
- Generate novel molecules using SMILES-based, graph-based, and 3D equivariant diffusion approaches, and assess their synthesizability.
- Predict protein-ligand complex structures using Boltz-1 and Chai-1 cofolding models, and extract binding affinity estimates.
- Apply universal machine-learned force fields (MACE-MP-0, CHGNet) to screen crystal structures and predict material properties.
- Plan retrosynthetic routes with AiZynthFinder and score candidates for practical synthesizability.
- Build a closed-loop drug discovery pipeline that integrates structure-based generation, cofolding, multi-objective Bayesian optimization, and full provenance tracking.
Sections
49.1 Molecular Generation and Drug Design
De novo molecular generation from SMILES autoencoders through equivariant diffusion models. Scaffold hopping with REINVENT. Structure-based generation with DiffSBDD. Synthesizability scoring and retrosynthesis with AiZynthFinder. ChemCrow for LLM-augmented chemistry.
49.2 Protein-Ligand Cofolding
From rigid docking to learned cofolding. DiffDock-L for diffusion-based pose prediction. Boltz-1 and Chai-1 for joint protein-ligand structure prediction. Binding affinity estimation. Practical cofolding pipelines with confidence filtering.
49.3 Materials Design and Universal Force Fields
Crystal structure generation with MatterGen. Universal machine-learned interatomic potentials: MACE-MP-0 and CHGNet. Property prediction across the periodic table. Autonomous chemistry laboratories and the self-driving lab paradigm.
49.4 Building a Drug Discovery Campaign
Recipe: closed-loop drug discovery combining DiffSBDD for structure-based generation, Boltz-1 for cofolding validation, multi-objective Bayesian optimization for lead optimization, and the Discovery Workbench provenance system for full auditability.
Bibliography
Molecular Generation
Introduces EDM, applying diffusion models to 3D molecular generation with E(3) equivariance, enabling direct generation of atomic coordinates and types.
DiffSBDD: generates 3D molecules conditioned on protein binding pockets using equivariant diffusion, producing chemically valid ligands with favorable predicted binding.
REINVENT uses reinforcement learning to fine-tune a SMILES generator toward molecules satisfying multi-parameter objectives, a production-grade de novo design tool.
Protein-Ligand Structure Prediction
An open-source model for predicting biomolecular complex structures including protein-ligand, protein-protein, and protein-nucleic acid interactions.
A multi-modal foundation model for molecular structure prediction that handles proteins, small molecules, DNA, RNA, and their complexes in a unified framework.
Frames molecular docking as a generative problem over the product space of translations, rotations, and torsion angles, solved by a diffusion process.
Materials Science and Force Fields
MACE-MP-0: a universal machine-learned interatomic potential trained on the Materials Project, providing DFT-quality predictions across the periodic table at a fraction of the cost.
CHGNet incorporates charge information into a graph neural network potential, improving accuracy for materials with varying oxidation states.
A diffusion model for generating novel stable crystal structures conditioned on desired properties like symmetry, chemistry, and target properties.
Tools and Synthesis Planning
A template-based retrosynthesis tool using Monte Carlo tree search to find synthetic routes from commercially available building blocks.
An LLM agent equipped with chemistry-specific tools (RDKit, web searches, reaction predictors) that can plan and execute multi-step chemical reasoning tasks.
The foundational open-source toolkit for cheminformatics, providing SMILES parsing, molecular descriptors, fingerprints, substructure search, and 2D/3D coordinate generation.
A deep learning library for chemistry and materials science, providing featurizers, datasets, and model architectures for molecular property prediction and generation.