A systematic study of data transfer techniques (joint training, self-training, pre-training plus fine-tuning) applied to Transformer-based retrosynthesis. Pre-training on USPTO-Full followed by fine-tuning on USPTO-50K achieves the best results, improving top-1 accuracy from 35.3% to 57.4%.
DrugAssist fine-tunes Llama2-7B-Chat on over one million molecule pairs for interactive, dialogue-based molecule optimization across six molecular properties.
DrugChat is a prototype system that bridges molecular graph neural networks with large language models for interactive, multi-turn question answering about drug compounds. It trains only a lightweight linear adaptor between a frozen GNN encoder and Vicuna-13B using 143K curated QA pairs from ChEMBL and PubChem.
DrugEx v2 introduces Pareto-based multi-objective optimization and evolutionary exploration strategies into an RNN reinforcement learning framework for de novo drug design toward multiple protein targets.
Jablonka et al. show that fine-tuning GPT-3 on natural language chemistry questions achieves competitive or superior performance to dedicated ML models across 15 benchmarks, with particular strength in low-data settings and inverse molecular design.
Galactica trains a decoder-only Transformer on a curated 106B-token scientific corpus spanning papers, proteins, and molecules, achieving strong results on scientific QA, mathematical reasoning, and citation prediction.
The Grammar VAE replaces character-level decoding with context-free grammar production rules, using a stack-based masking mechanism to guarantee that all generated SMILES strings are syntactically valid. Applied to molecular optimization and symbolic regression, it learns smoother latent spaces and finds better molecules than character-level baselines.
LatentGAN decouples molecular generation from SMILES syntax by training a Wasserstein GAN on latent vectors from a pretrained heteroencoder, enabling de novo design of drug-like and target-biased compounds.
LlaSMol fine-tunes Mistral, Llama 2, and other open-source LLMs on SMolInstruct, a 3.3M-sample instruction tuning dataset covering 14 chemistry tasks. The Mistral-based model outperforms GPT-4 and Claude 3 Opus across all tasks.
Ertl et al. train a character-level LSTM on 509K bioactive ChEMBL SMILES and generate one million novel, diverse molecules whose physicochemical properties, substructure features, and predicted bioactivity closely match the training distribution.
Introduces a memory unit that modifies the RL reward function to penalize previously explored chemical scaffolds, substantially increasing the diversity of generated molecules while maintaining relevance to known active ligands.
Proposes MolecularRNN, a graph recurrent model that generates molecular graphs atom-by-atom with 100% validity via valency-based rejection sampling, then shifts property distributions using policy gradient reinforcement learning.