Google DeepMind has unveiled TxGemma, a powerful suite of open-source language models designed to support the discovery and development of new therapies. Available via Vertex AI Model Garden and Hugging Face, the models aim to boost research in molecular property prediction, drug candidate evaluation, and clinical trial outcome estimation.
Three Model Sizes with Advanced Capabilities
TxGemma, a successor to the October 2024 release Tx-LLM, is based on DeepMind’s Gemma family of models. It is trained on 7 million biomedical examples and offered in three sizes—2B, 9B, and 27B—each featuring both “predict” and “chat” versions. While the “predict” models are optimised for specific tasks such as determining molecular toxicity, the “chat” models support scientific reasoning and interpretation of results.
According to Shekoofeh Azizi, staff research scientist at DeepMind, “TxGemma is specifically trained to understand and predict the properties of therapeutic entities throughout the entire discovery process.” She added that it could help reduce drug development timelines and costs.
The flagship 27B predict model outperformed its predecessor and many task-specific models, scoring higher in 45 out of 66 benchmark tasks and matching or surpassing others in 50 of them.
Also read: CDSCO Tightens Rules for SEZ Drug Transfers
Support for Custom Research and Agentic Workflows
DeepMind has also provided resources for researchers to fine-tune TxGemma using proprietary data. For instance, a Colab notebook demonstrates how TxGemma can be adapted using the TrialBench dataset to predict adverse events in clinical trials.
In a further enhancement, DeepMind introduced Agentic-Tx, an orchestrated framework that integrates TxGemma with 18 specialised tools, including gene databases, search utilities, and molecular analysis modules. Powered by Gemini 2.0 Pro, this agentic system enables multi-step reasoning for complex therapeutic workflows.
Early testing on benchmarks such as ChemBench and Humanity’s Last Exam revealed that Agentic-Tx performs robustly in both chemistry and biomedical tasks.
“We’re excited to see how the community will use TxGemma to accelerate therapeutic discovery,” Azizi said, inviting researchers to contribute by refining the models and sharing feedback.
