Literature · google
TxGemma 9B Chat
A conversational 9B-parameter language model for drug discovery and therapeutic property prediction with reasoning.
What TxGemma 9B Chat is
TxGemma 9B Chat is a conversational open language model designed for drug discovery and therapeutic development, developed by Google. Built on the Gemma 2 architecture, it is fine-tuned to process and comprehend information spanning diverse therapeutic entities, including small molecules, proteins, nucleic acids, diseases, and cell lines.
The model addresses the need for adaptable tools in therapeutic development that can perform multi-turn interactions. Unlike standard predictive variants that require strict single-task prompts, this chat edition supports natural dialogue, allowing researchers to request therapeutic property predictions and ask the model to explain the structural or biological rationale behind its answers.
Trained on instructional datasets curated from the Therapeutics Data Commons alongside general conversational data, TxGemma 9B Chat functions both as an interactive drug discovery assistant and as a pre-trained foundation for downstream fine-tuning on specialized scientific datasets.
How it works
TxGemma 9B Chat is a decoder-only transformer based on the Gemma 2 9B architecture. Training was carried out using JAX, enabling hardware acceleration across TPUs. The model undergoes instruction fine-tuning combining domain-specific therapeutic data from the Therapeutics Data Commons (TDC) with general instruction-tuning corpora to maintain conversational abilities.
The model takes text inputs formatted with natural language instructions, context, task questions, and biological or chemical identifiers such as SMILES strings, amino acid sequences, and nucleotide sequences. It produces textual completions or conversational turns demarcated by dedicated chat role tokens.
Key facts
- Architecture
- Decoder-only Transformer (based on Gemma 2)
- Parameter count
- 9 billion (9B)
- Training framework
- JAX
- Base model
- Gemma 2 9B
- Training data
- Therapeutics Data Commons (commercially licensed subset) and general instruction-tuning data
- Input modalities
- Text (including SMILES, amino acid sequences, nucleotide sequences)
- Output modality
- Text
- Model version
- 1.0.0
What you can use it for
- Molecular Property Prediction and Explanation
- Predict chemical and pharmacological attributes, such as blood-brain barrier permeability from SMILES representations, and query the model across multiple chat turns to explain its reasoning based on molecular structure.
- Target Identification and Interaction
- Evaluate relationships between potential drug candidates and biological targets, including drug-target interaction assessment using sequence and chemical text data.
- Clinical Trial Outcome Prediction
- Predict drug development endpoints, such as clinical trial approval prospects, within structured therapeutic discovery pipelines.
- Agentic Drug Discovery Workflows
- Serve as a specialized tool within agentic multi-model discovery systems to answer queries and perform step-by-step therapeutic evaluations.
Run it here
We list this model but cannot run it yet: excluded (gated): HAI-DEF terms, as above..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Capable of multi-turn natural language dialogue and explaining the reasoning underlying its therapeutic predictions.
- Exhibits strong data efficiency, outperforming or matching best-in-class results on 50 of 66 Therapeutics Data Commons benchmarks across the model family.
- Accommodates diverse biological and chemical representations, including small molecule SMILES, amino acid sequences, and nucleotide sequences.
- Can serve directly as a pre-trained base model for further instruction fine-tuning on private or institutional datasets.
Watch out for
- Conversational flexibility comes at the expense of raw predictive performance compared to non-chat TxGemma prediction variants.
- Trained exclusively on public datasets with commercial licenses from the Therapeutics Data Commons, omitting non-commercially licensed datasets included in the reference research paper.
- Requires downstream task-specific validation to ensure predictions are reliable and representative of real-world use contexts.
Inputs and outputs
You provide
Prompt
The model continues from here.
You get back
The model's completion.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use TxGemma 9B Chat commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
How does the Chat variant differ from the standard TxGemma 9B model?
The Chat variant is trained on both therapeutic data and general instruction-tuning datasets, enabling multi-turn conversation and natural-language explanations of predictions. This increased conversational flexibility results in a slight reduction in raw predictive performance compared to the dedicated prediction model.
What types of biochemical inputs can the model accept?
The model accepts text containing small molecule SMILES strings, amino acid sequences, nucleotide sequences, and natural language descriptions of diseases or cell lines.
What template format should be used for conversational prompts?
TxGemma Chat models rely on a specific chat template structured with role tokens ('user' and 'model'), which can be automatically applied using the tokenizer's built-in chat template utility.
Which benchmark tasks was TxGemma validated against?
TxGemma models were validated against a suite of 66 therapeutic development tasks from the Therapeutics Data Commons, evaluated across standard data split regimes including random, scaffold, cold-start, combination, and temporal splits.
How to cite it
Wang, Eric and Schmidgall, Samuel and Jaeger, Paul F. and Zhang, Fan and Pilgrim, Rory and Matias, Yossi and Barral, Joelle and Fleet, David and Azizi, Shekoofeh. TxGemma: Efficient and Agentic LLMs for Therapeutics. arXiv:2504.06196, 2025.
Written from TxGemma 9B Chat's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.