Literature · google
TxGemma 27B Chat
A 27-billion-parameter conversational language model for therapeutic discovery and drug property prediction.
What TxGemma 27B Chat is
TxGemma 27B Chat is a conversational language model designed for drug discovery and development, developed by Google. Built upon the Gemma 2 architecture, it is tailored to process and reason across a variety of therapeutic targets and modalities, including small molecules, proteins, nucleic acids, cell lines, and diseases.
Unlike standard prediction-only checkpoints that require rigid prompting structures, the conversational variant can engage in multi-turn natural language dialogue. This enables researchers to query drug-related properties and receive step-by-step rationales explaining the structural and biochemical mechanisms behind the predictions.
The model addresses key bottlenecks in therapeutic research, functioning as both an interactive assistant and a base foundation for downstream fine-tuning across predictive biomedical tasks.
How it works
TxGemma 27B Chat is built on a decoder-only transformer architecture derived from Gemma 2. The model was trained in JAX using instruction fine-tuning on a mixture of therapeutic datasets (TxT) alongside general instruction-tuning data to preserve conversational fluency.
The therapeutic training data originates from the Therapeutics Data Commons (TDC), comprising tasks spanning over 15 million data points across various biomedical entities. Released variants are restricted to training datasets that carry commercial licenses. It accepts inputs formatted with context, instructions, and questions containing representations such as SMILES strings, amino acid sequences, and nucleotide sequences.
Key facts
- Architecture
- Decoder-only Transformer (based on Gemma 2)
- Parameter count
- 27B
- Base model
- Gemma 2
- Input modalities
- Text (natural language, SMILES strings, amino acid sequences, nucleotide sequences)
- Output modality
- Text
- Training software
- JAX
- Release version
- 1.0.0
What you can use it for
- Molecular Property Prediction
- Predicting chemical and biological characteristics of drug candidates, such as blood-brain barrier penetration from molecular SMILES strings.
- Interactive Scientific Reasoning
- Conducting multi-turn discussions where the model explains the structural rationale behind its biochemical predictions.
- Target and Drug Interaction Analysis
- Assessing relationships and binding predictions between small molecules, protein targets, and nucleic acids.
- Clinical Trial Approval Forecasting
- Assisting therapeutic evaluation workflows by predicting downstream clinical trial approval outcomes.
Run it here
We list this model but cannot run it yet: excluded (gated): HAI-DEF + Gemma terms: hosting is 'Distribution' and requires flowing restrictions down to users..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Maintains conversational capability allowing multi-turn dialogues and rationalisation of predictions.
- Trained across more than 15 million biomedical data points covering small molecules, proteins, and nucleic acids.
- Matches or exceeds best-in-class performance on 50 of 66 therapeutic benchmark tasks from Therapeutics Data Commons.
- Serves as a pre-trained base for integration into larger agentic workflows or further fine-tuning on proprietary data.
Watch out for
- Conversational versatility comes at the expense of raw predictive performance compared to non-chat TxGemma variants.
- Fine-tuning is restricted to publicly available TDC datasets that have commercial licenses.
- Downstream applications require external validation to ensure appropriateness for target operational distributions.
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 27B 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.
What input formats does TxGemma 27B Chat accept?
It accepts standard text inputs which can include natural language, chemical SMILES strings, amino acid sequences, and nucleotide sequences.
How does the Chat variant differ from standard TxGemma models?
The Chat variant is trained on both therapeutic and general instruction datasets, allowing it to converse in multi-turn dialogues and explain reasoning, though with slightly reduced raw predictive accuracy compared to dedicated prediction models.
What benchmark was used to evaluate this model?
Performance was evaluated using 66 therapeutic tasks curated from the Therapeutics Data Commons across varied data split strategies, including scaffold, cold-start, temporal, combination, and random splits.
How to cite it
@article{wang2025txgemma, title={TxGemma: Efficient and Agentic LLMs for Therapeutics}, author={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}, year={2025}}
Written from TxGemma 27B 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.