Language Models for Molecule Generation
We build AI-enhanced algorithms for molecular design with the goal of reaching superhuman performance. Our aim is to create systems that generate drug candidate molecules with properties specified by medicinal chemists, either in a single step or through iterative optimization. We have developed language models that understand 2D molecular graphs and basic molecular properties, and we combined them with evolutionary algorithms to generate property-conditioned molecules beyond existing databases. We are currently extending these models with 3D molecular understanding and developing more realistic, practically relevant benchmarks for molecular design.
PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization
We introduce PMO-Dock, a molecular optimization benchmark with 25 tasks testing docking, target specificity, and generalization. Accepted at NeurIPS 2026, Evaluations and Datasets track; an earlier version appeared at ICLR 2026 GEM.
CoordToken: Accurate Tokenization of 3D Small Organic Molecules
We introduce CoordToken, a tokenizer for 3D small organic molecules using finite scalar quantization, improving reconstruction accuracy for conformer generation.
Downstream Transfer of Scaling in Molecular Language Models
We study how molecular language model scale transfers to property prediction, property-conditioned generation, and docking-based molecular optimization. Gains depend on the downstream task.
Towards Scaling Laws for Language Model Powered Evolutionary Algorithms: Case Study on Molecular Optimization
We show that evolutionary algorithms enhanced with LLM-based molecule generation can scale for simple molecular optimization tasks in both directions: LLM size and number of optimization steps
Towards Molecular Conformer Generation with Language Models
We trained language models that directly generate the 3D structures of drug-like molecules. We also show improvements with the scale of the models.
Small Molecule Optimization with Large Language Models
Mol-E combines molecular language models (ChemLactica and Chemma) with evolutionary search for molecular optimization. The September 2026 revision evaluates task-agnostic and task-informed optimization, alongside multi-property docking objectives.
Exploring the Recall of Language Models: Case Study on Molecules
We present a benchmark for evaluating language model recall using organic molecules, probing when models reproduce memorized molecular structures.
BARTSmiles: large-scale generative masked language models for molecular representations
A BART-like encoder-decoder model trained on 1.7 billion SMILES. Demonstrated competitive performance after fine-tuning on property prediction, chemical reaction prediction and retrosynthesis tasks.
Datasets Construction and Development of QSAR Models for Predicting Micronucleus In Vitro and In Vivo Assay Outcomes
We construct toxicology datasets and develop QSAR models for predicting micronucleus assay outcomes, in collaboration with molecular biology and toxicology researchers.
Improved molecular representations for property prediction based on VAEs
We show that property predictors operating on the latent space of VAEs can improve the downstream performance on related property prediction tasks.