Google DeepMind has launched TxGemma, a suite of open-source AI models specifically designed to enhance therapeutic development and drug discovery.

This new series is a successor to Tx-LLM, which was introduced in October 2024, and is trained on 7 million examples, available in three sizes: 2B, 9B, and 27B.

Among these models, the 27B predict model has shown superior performance, excelling in 45 out of 66 benchmark tasks and matching or surpassing task-specific models in 50 tasks.

TxGemma aims to improve drug candidate assessment, molecule property prediction, and clinical trial outcome estimation by leveraging large language model capabilities applied to biomedical data.

The models are accessible through Vertex AI Model Garden and Hugging Face, allowing researchers to experiment, fine-tune with proprietary data, and share their findings.

The release includes tools for fine-tuning, illustrated by a Colab notebook that demonstrates how to adapt TxGemma for adverse event prediction in clinical trials.

TxGemma's chat-based versions, equipped with instruction tuning, can assist researchers in interpreting predictions and explaining toxicity assessments based on molecular structures.

Each model version includes specialized 'predict' models for tasks like assessing molecular toxicity and 'chat' models for conversational analysis.

DeepMind also introduced Agentic-Tx, an orchestrated system powered by Gemini 2.0 Pro, which integrates TxGemma with 18 tools for multi-step reasoning tasks in biology and chemistry.

Agentic-Tx has demonstrated strong performance on challenging benchmarks such as ChemBench and Humanity’s Last Exam, showcasing its ability to manage complex therapeutic workflows.

DeepMind expresses excitement about the potential community applications of TxGemma to accelerate therapeutic discovery.

The TxGemma models are built on existing technologies Gemma and Gemini, enhancing their capabilities for health-related applications.