TxGemma

TxGemma is a family of open therapeutics foundation models from Google, built on the Gemma-2 architecture and released in March 2025. It targets the drug development pipeline, where predicting the properties of a candidate therapeutic — its efficacy, toxicity, absorption, and interactions — remains slow, costly, and heavily fragmented across single-task specialist models. TxGemma reframes these tasks as text-to-text problems, letting one generalist model reason over small molecules, proteins, nucleic acids, diseases, and cell lines expressed as strings such as SMILES and amino acid sequences.

The models come in three sizes (2B, 9B, and 27B parameters) and two flavors. Prediction variants are available at all three sizes and expect a narrow, structured prompt for a specific therapeutic task. Conversational chat variants, released at 9B and 27B, retain that predictive skill while also supporting multi-turn dialogue — answering open questions, explaining the rationale behind a prediction, and following instructions expressed in natural language.

TxGemma is the successor to Google's Tx-LLM (arXiv:2406.06316) and was developed by researchers at Google DeepMind and Google Research. It is distributed as part of the Health AI Developer Foundations program, with open weights on Hugging Face and Vertex AI Model Garden under Google's custom terms of use.

#Key Features

• Generalist therapeutic prediction: A single text-to-text model handles dozens of property-prediction tasks spanning small molecules, proteins, nucleic acids, diseases, and cell lines, replacing a patchwork of specialist models.
• Predict and chat variants: Prediction models (2B/9B/27B) serve high-throughput narrow prompting, while conversational models (9B/27B) add flexible multi-turn reasoning and can explain their predictions.
• Efficient scaling: The 27B prediction model matches or exceeds much larger general-purpose systems on therapeutic benchmarks while remaining compact enough to fine-tune and self-host.
• Fine-tuning foundation: Released as open weights, TxGemma is designed to be adapted with proprietary data for specific drug-discovery programs.
• Agentic-Tx system: An accompanying agentic workflow couples TxGemma with a general-purpose Gemini reasoning model and 18 tools — including PubMed search and molecular analysis utilities — for multi-step research tasks.

#Technical Details

TxGemma fine-tunes the decoder-only Gemma-2 transformer on TxT, a curated instruction-tuning mixture of roughly 7 million training examples drawn from the Therapeutics Data Commons (TDC), spanning 66 therapeutic tasks; chat variants additionally include general instruction-tuning data to preserve conversational ability. Across the 66 tasks, the 27B prediction model is superior or comparable to a state-of-the-art generalist on 64 tasks (better on 45) and competitive with best-in-class specialist models on 50 tasks (better on 26). Within Agentic-Tx, the system delivers a 52.3% relative improvement over o3-mini (high) on the Chemistry and Biology subset of Humanity's Last Exam and 26.7% on GPQA Chemistry. Inputs and outputs are text, with molecular entities encoded as SMILES, sequences, or natural language.

#Applications

TxGemma serves computational chemists, pharmacologists, and drug-discovery teams who need fast property estimates for candidate therapeutics — toxicity, binding, ADME, and clinical-trial-related endpoints — early in a program. The prediction variants suit high-throughput virtual screening and triage, while the chat variants support interactive hypothesis exploration and explanation. Because the weights are open, groups can fine-tune the models on internal assay data to specialize them for particular targets or chemistries.

#Impact

TxGemma packages a broad slice of the Therapeutics Data Commons benchmark suite into openly released, fine-tunable models, lowering the barrier to applying large language models across the therapeutic pipeline. Its efficiency-oriented design shows that mid-sized, task-specialized LLMs can rival far larger systems on domain benchmarks. TxGemma is released under Google's Health AI Developer Foundations terms of use, a custom license permitting research and commercial development with health-specific restrictions rather than an OSI-approved open-source license, and its benchmark results are in-silico evaluations that require task-specific validation before real-world use.