Google DeepMind releases TxGemma: Open AI Models for Drug Discovery

In a move that could accelerate and reshape therapeutic development, Google DeepMind has launched TxGemma, a suite of open AI models designed to make drug discovery faster, smarter, and more cost-efficient. The announcement was made by Shekoofeh Azizi, Staff Research Scientist, on March 25, 2025.

What is TxGemma?

TxGemma is a collection of lightweight, high-performing open models that are fine-tuned from DeepMind’s Gemma 2 architecture. These models are built specifically to assist in therapeutic development, from identifying promising drug targets to predicting outcomes of clinical trials.

With over 7 million therapeutic examples used in training, TxGemma models excel in prediction and conversational analysis, potentially helping researchers cut down the years and billions traditionally spent in drug development.

The need for efficient drug development is greater than ever. Nearly 90% of drug candidates fail beyond Phase 1 trials, costing companies time, resources, and human capital. TxGemma offers a powerful solution by using large language models (LLMs) to better understand molecular properties, toxicity, blood-brain barrier crossing, binding affinities, and more.

In early benchmarks, the TxGemma 27B Predict model either outperformed or matched Google’s previous state-of-the-art model, Tx-LLM, across 64 of 66 tasks. It even held its ground against specialized single-task models, outperforming them on 26 out of 66 tasks.

Who is it for and how can it be used?

The TxGemma release includes three model sizes: 2B, 9B, and 27B, each available in two versions:

• Predict Version: Tailored for specific scientific tasks like classification (e.g., toxicity), regression (e.g., binding affinity), and generation (e.g., molecular synthesis).
• Chat Version: Optimized for conversational reasoning, allowing researchers to interact with the model, for example, asking why a molecule was flagged as toxic.

To support real-world application, Google has released Colab notebooks for developers to fine-tune TxGemma on their own data, using datasets like TrialBench. This enables more accurate, proprietary predictions tailored to specific drug development pipelines.

What’s new with Agentic-Tx?

TxGemma is also part of Agentic-Tx, an advanced agentic AI system built on Gemini 2.0 Pro. With access to 18 specialized tools, Agentic-Tx can handle multi-step reasoning using data from PubMed, Wikipedia, and various gene/protein databases.

Agentic-Tx has achieved state-of-the-art performance on benchmarks like ChemBench and Humanity’s Last Exam, showcasing its ability to answer complex biology and chemistry questions.

Where to access it?

TxGemma is now available on:

•    Vertex AI Model Garden
•    Hugging Face

Researchers and developers are encouraged to try out the inference models, fine-tuning tutorials, and agent system demos. The goal is to foster an open, collaborative ecosystem for improving therapeutic AI research.