Introducing TxGemma: Open models to improve therapeutics development- Google Developers Blog

What this post added

Introduced TxGemma, a collection of open models specifically trained for therapeutic development, built upon Google DeepMind's Gemma. TxGemma models are fine-tuned from Gemma 2 and are available in 2B, 9B, and 27B sizes, with 'predict' and 'chat' versions. The 'predict' versions are tailored for tasks like molecule toxicity prediction, blood-brain barrier crossing classification, and binding affinity regression. The 'chat' versions offer conversational capabilities for explaining predictions and engaging in multi-turn discussions. A fine-tuning example Colab notebook is provided for adapting TxGemma to custom therapeutic data. Additionally, demonstrated Agentic-Tx, a therapeutics-focused agentic system powered by Gemini 2.0 Pro, which integrates TxGemma and other tools for complex research problems, achieving state-of-the-art results on reasoning-intensive chemistry and biology tasks.