Literature · google
TxGemma 9B Predict
A 9B-parameter language model fine-tuned on Therapeutics Data Commons for therapeutic property prediction.
What TxGemma 9B Predict is
TxGemma 9B Predict is a decoder-only language model developed by Google, tailored for therapeutic property prediction and drug discovery tasks. Built by fine-tuning the 9-billion parameter Gemma 2 architecture, the model is trained to interpret and evaluate diverse biochemical entities, including small molecules, proteins, nucleic acids, cell lines, and diseases.
Unlike conversational variants designed for dialogue, this model is optimised specifically for raw predictive accuracy. It operates using structured, task-specific prompt formats to assess biochemical properties, drug interactions, and development milestones across the preclinical and translational pipeline.
How it works
TxGemma 9B Predict is based on the Gemma 2 decoder-only transformer architecture and was trained using JAX. It underwent instruction fine-tuning on diverse biomedical datasets curated from the Therapeutics Data Commons (TDC), spanning 66 therapeutic tasks and over 15 million biomedical data points.
The model consumes text-based inputs that combine natural language task instructions and context with domain-specific encodings, such as simplified molecular-input line-entry system (SMILES) strings, amino acid sequences, or nucleotide sequences. It produces text-based outputs specifying the predicted class or property value.
Key facts
- Base architecture
- Gemma 2 (decoder-only Transformer)
- Parameter count
- 9 billion
- Training framework
- JAX
- Fine-tuning data
- Therapeutics Data Commons (TDC)
- Input format
- Text (natural language, SMILES, amino acid and nucleotide sequences)
- Output format
- Text
What you can use it for
- Molecular Property Prediction
- Predicting key pharmacological attributes from chemical SMILES strings, such as blood-brain barrier permeability.
- Drug-Target Interaction Assessment
- Evaluating potential interactions between candidate therapeutic compounds and specific protein or biological targets.
- Target Identification
- Screening and prioritising biological targets across diverse disease contexts and cell line profiles.
- Clinical Trial Outcome Forecasting
- Predicting clinical trial approval and progression probabilities using structured therapeutic task prompts.
- Fine-Tuning Foundation
- Serving as an initialised checkpoint for downstream parameter adaptation on private or specialised therapeutic datasets.
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
- Surpasses or matches best-in-class performance on 50 out of 66 therapeutic benchmark tasks in the Therapeutics Data Commons.
- Outperforms specialist single-task models on 26 distinct therapeutic tasks.
- Improves predictive accuracy across 45 out of 66 benchmark tasks compared to its predecessor, Tx-LLM.
- Processes multiple biological data modalities in raw text, including SMILES, nucleotide sequences, and protein sequences.
- Maintains higher raw prediction performance than conversational variants by focusing on structured prompts.
Watch out for
- Requires structured prompt templates following Therapeutics Data Commons conventions to achieve optimal predictive performance.
- Lacks the interactive multi-turn dialogue and rationale-explaining features found in conversational variants.
- Trained exclusively on public benchmarks, necessitating independent downstream validation on representative operational datasets.
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 Predict 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 TxGemma 9B Predict differ from TxGemma-Chat?
TxGemma 9B Predict is specialised for predictive accuracy using narrow, template-formatted prompts, whereas conversational variants trade some raw predictive performance to support multi-turn dialogue and reasoning explanations.
What types of biochemical inputs can the model process?
The model accepts text prompts containing natural language alongside chemical SMILES strings, amino acid sequences, and nucleotide sequences.
How was the model evaluated?
Performance was validated across 66 therapeutic tasks from the Therapeutics Data Commons using benchmark-standard splits, including random, scaffold, cold-start, combination, and temporal data partitions.
How to cite it
Wang, Eric, Samuel Schmidgall, Paul F. Jaeger, Fan Zhang, Rory Pilgrim, Yossi Matias, Joelle Barral, David Fleet, and Shekoofeh Azizi. TxGemma: Efficient and Agentic LLMs for Therapeutics. arXiv:2504.06196 (2025).
Written from TxGemma 9B Predict'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.