Google Creates Tx-LLM for Drug Discovery and Therapeutic Development

Google Research and Google DeepMind have introduced Tx-LLM, a new large language model (LLM) for drug discovery and therapeutic development, built from PaLM-2. Tx-LLM leverages Google's generative AI technology to address medical questions.

Trained on 709 datasets for 66 tasks across drug discovery stages, Tx-LLM evaluates efficacy, safety, targets, and manufacturing ease. It uses the Therapeutics Instruction Tuning (TxT) collection, combining free-text instructions with small molecule representations, like SMILES strings, to refine its performance.

SMILES (Simplified Molecular Input Line Entry System) represents molecules using printable characters. TxT helps Tx-LLM tackle classification, regression, and generation tasks in drug development. For drug synergy prediction, prompts include instructions, context, and questions.

Tx-LLM outperformed or matched state-of-the-art (SOTA) models in 43 of 66 tasks and exceeded them in 22 tasks. The researchers noted positive transfer between diverse datasets, enhancing performance across different drug types. They highlighted Tx-LLM's potential as an end-to-end tool for therapeutic development.

Med-PaLM 2, released last year, improved upon its predecessor by generating more comprehensive medical answers. AI is increasingly crucial in drug discovery. In December, Absci partnered with AstraZeneca in a $247 million deal to develop AI-based cancer treatments. Similarly, IBM and Boehringer Ingelheim are collaborating on biologic drug discovery using AI. Other notable companies include Genesis, Daewoong Pharmaceutical, and AION Labs.