Google DeepMind has launched TxGemma, a collection of open AI models aimed at improving drug discovery and clinical trial predictions. This initiative is designed to address the slow and costly nature of drug development, where 90% of drug candidates fail to advance beyond phase 1 trials. TxGemma utilizes large language models (LLMs) to enhance predictions across the research pipeline, from identifying drug targets to assessing clinical outcomes.
The model family includes three sizes: 2B, 9B, and 27B parameters, with specialized Predict versions for tasks like classifying molecule properties and estimating drug binding affinity. In benchmark tests, the 27B Predict model excelled, outperforming or matching specialized models in 64 out of 66 tasks. TxGemma also features an interactive component, TxGemma-Chat, which allows researchers to ask questions and receive detailed explanations about predictions.
To cater to specific research needs, a fine-tuning example is provided for customization. Additionally, Google DeepMind introduces Agentic-Tx, which integrates TxGemma into multi-step research workflows, enhancing research capabilities. TxGemma is now accessible on Vertex AI Model Garden and Hugging Face, enabling researchers to experiment and provide feedback.
