Generative modeling for molecular design and discovery

Addressing generalization limits in machine learning models for chemical reactions, this work introduces a pretraining scheme that enables accurate predictions for unseen chemistry and expands computational discovery.

FLOWR combines continuous and categorical flow matching with equivariant optimal transport and a separate protein pocket encoder to improve validity, pose accuracy and interaction recovery at substantially faster inference speeds.

A general-purpose large language model can be trained to create a specialized model, SmileyLlama, that can generate drug-like molecules with user-specified properties that are relevant to medicinal chemistry through a process that might be extended to other scientific domains.

The authors propose the TANGO reward function, which enables the generation of property-optimized small molecules with predicted synthesis routes, incorporating a small set of shared precursors.

This study presents a dynamic routing-guided framework to model and interpret salt–solvent chemistry, which effectively handles long-tailed data and captures the full spectrum of formulations, shaping the conductivity atlas for non-aqueous electrolytes.

PropMolFlow is a flow-matching method for property-guided molecule generation that matches diffusion model performance while generating stable, valid structures more quickly and enabling the discovery of molecules with under-represented property values.

A persistent gap from theoretical molecules to experimentally viable compounds has hindered the practical adoption of generative algorithms. This study proposes SynGFN as a bridge linking molecular design and synthesis, accelerating exploration and producing diverse, synthesizable, high-performance molecules.

This study presents ECloudGen, which uses latent diffusion to generate electron clouds from protein pockets and decodes them into molecules. The adopted two-stage training expands the chemical space accessible to generative drug design.

A pharmacophore-oriented 3D molecular generation method, which efficiently generates diverse, drug-like molecules is proposed. It successfully identifies new inhibitors in real-world drug discovery.

Combining conformal prediction machine learning with molecular docking, a method to efficiently screen multi-billion-scale libraries is developed, enabling the discovery of a dual-target ligand modulating the A_2A adenosine and D₂ dopamine receptors.

A physics-based training pipeline is developed to help tackle the challenges of data scarcity. The framework aligns large language models to a physically consistent initial state that is fine-tuned for learning polymer properties.

This work applies diffusion models to conditional molecule generation and shows how they can be used to tackle various structure-based drug design problems

DeepBlock is a deep learning framework for ligand generation, inspired by the DNA-encoded compound library technique, that enhances ligand design with building blocks and a rule-based reconstruction algorithm, achieving better drug properties.

The downselection of compounds for synthesis is a key challenge in molecular design cycles that typically relies on expert chemist intuition. Fromer and Coley propose a cost-aware method to automatically select compounds and synthetic routes.

A method is developed for the directional optimization of multiple properties without prior knowledge on their nature. Using a large ligand dataset, diverse metal complexes are found along the Pareto front of vast chemical spaces.

An optimization algorithm is used to discover guest molecules based on knowing only the structure of the host. The molecules are represented as 3D volumes, optimized to improve host–guest interaction and converted into SMILES using a transformer model.

A diffusion model that generates chemical reactions in 3D with all desired symmetries preserved is established and shown to reduce transition state search from days to seconds and complement intuition-based reaction exploration with generative AI.

SurfGen is a structure-based drug design approach that delves into topological and geometric deep learning techniques for interaction learning, echoing the classical lock-and-key model.

GaUDI is a guided diffusion method for the design of molecular structures that features a flexible and scalable target function and that achieves high validity of generated molecules.

A generative deep learning model of molecular structure is combined with supervised deep learning models of molecular properties to achieve high-throughput (multi-)property-driven design of organic molecules.

The Absolut! framework can generate synthetic three-dimensional antibody–antigen structures to assist machine learning and dataset construction for antibody design. Most importantly, the relative machine learning performance learnt on Absolut! datasets is shown to transfer to experimental datasets.

The application of machine learning techniques to small-molecule drug discovery has not yet yielded a true leap forward in the field. This Perspective discusses how a renewed focus on data and validation could help unlock machine learning’s potential.

Autoencoders are versatile tools for molecular informatics with the opportunity for advancing molecule and drug design. In this Review, the authors highlight the active areas of development in the field and explore the challenges that need to be addressed moving forward.

Nature Computational Science is calling all researchers who develop and use generative models for molecular design and discovery to publish with us.

As artificial intelligence (AI) proliferates, synthetic chemistry stands to benefit from its progress. Despite hidden variables and ‘unknown unknowns’ in datasets that may impede the realization of a digital twin for the laboratory flask, there are many opportunities to leverage AI and large datasets to advance synthesis science.