Google drops TxGemma to help AI design better drugs

TxGemma is Google's new collection of open language models built specifically for drug discovery and therapeutics development. It is designed to help researchers predict molecular properties, screen for toxicity, and evaluate clinical trial risks throughout the research pipeline.

The collection is fine-tuned from Google DeepMind's Gemma 2 using 7 million domain-specific training examples. It serves as the open successor to Tx-LLM, a generalist therapeutics model Google unveiled last October, scaling those capabilities down into models developers can actually download, run, and fine-tune on their own data.

What it does

TxGemma comes in three parameter sizes: 2B, 9B, and 27B, matching different compute budgets from local setups to GPU clusters.

Every size includes a specialized predict version tailored for specific tasks drawn from the Therapeutic Data Commons, such as forecasting whether a candidate molecule carries toxic properties.

For researchers who need more than a raw score, the 9B and 27B sizes also come in chat versions. Trained with general instruction-tuning data, these models can engage in multi-turn dialogue and explain their reasoning, such as breaking down the chemical substructures behind a toxicity call. Google notes that this conversational ability comes with a small performance trade-off compared to the pure predict models.

Performance and model variants

TxGemma puts up serious numbers against both generalists and single-task specialists. The 27B predict model matches or beats Google's previous Tx-LLM on 64 out of 66 benchmark tasks, directly outperforming it on 45.

When pitted against models built strictly for a single specific task, the 27B model held its own or won on 50 tasks, beating the specialists outright on 26 of them.

To handle multi-step workflows, Google also showed off Agentic-Tx, a framework powered by Gemini 2.0 Pro and equipped with 18 specialized tools. This setup achieved state-of-the-art marks on heavy chemical and biological reasoning benchmarks, including ChemBench and Humanity's Last Exam.

How to try it

TxGemma weights are available right now on Hugging Face and Google Cloud Vertex AI Model Garden.

Google has also published interactive Colab notebooks covering basic inference, agent workflows, and a fine-tuning guide that uses the TrialBench dataset to predict adverse events in clinical trials.

Why it matters

Developing a new therapeutic is brutally expensive, and roughly 90% of candidates fail after entering clinical trials. Giving labs open weights that understand chemistry and clinical data at this level is a massive win for biopharma startups and academic labs that cannot afford massive proprietary discovery platforms. Teams sitting on proprietary assay data can now fine-tune a 9B or 27B model locally without sending sensitive molecules to a closed third-party API.

That said, pure language models are still an assistive screening tool rather than a replacement for physical wet labs or rigorous molecular dynamics simulations. But as an open foundation to accelerate target identification and toxic hazard screening, TxGemma gives the community a remarkably capable starting point.