Literature · google
TxGemma 27B Predict
A 27B language model predicting drug, target, and molecular properties across diverse therapeutic modalities.
What TxGemma 27B Predict is
TxGemma 27B Predict is a 27-billion-parameter language model developed by Google for therapeutic research and drug development. Built upon Gemma 2, it is designed to understand and process biological and chemical entities, including small molecules, proteins, nucleic acids, cell lines, and diseases.
The model addresses a broad spectrum of predictive challenges across the drug discovery pipeline, including target identification, drug-target interaction screening, and clinical trial approval prediction. In benchmark evaluations covering 66 therapeutic tasks, it matched or surpassed best-in-class results on 50 tasks and outperformed dedicated specialist models on 26.
Unlike conversational variants within the TxGemma family, the predict version is specialised for maximum raw predictive performance. It relies on structured prompt templates to generate discrete therapeutic predictions and can also serve as a pre-trained base for task-specific fine-tuning.
How it works
TxGemma 27B Predict uses a decoder-only transformer architecture based on Gemma 2. It was trained using JAX on TPU hardware via instruction fine-tuning on therapeutic datasets curated from the Therapeutics Data Commons (TDC). The released model weights are trained on TDC datasets that permit commercial use.
The model processes text inputs structured into instructions, background biological context, and questions. Inputs can combine natural language with molecular representations, such as SMILES strings, amino acid sequences, and nucleotide sequences, generating text-based classification or regression predictions.
Key facts
- Architecture
- Decoder-only Transformer (Gemma 2)
- Parameter count
- 27 billion
- Base model
- Gemma 2 27B
- Model version
- 1.0.0
- Creation date
- 18 March 2025
- Training software
- JAX
- Primary training data
- Therapeutics Data Commons (TDC)
- Input formats
- Text (SMILES, amino acid sequences, nucleotide sequences, natural language)
- Output format
- Text
What you can use it for
- Blood-brain barrier penetration prediction
- Forecast whether a small molecule crosses the blood-brain barrier given its SMILES representation.
- Drug-target interaction screening
- Evaluate binding or functional interactions between candidate small molecules and target proteins.
- Target identification and disease profiling
- Examine relationships between molecular targets, biological disease mechanisms, and cell lines.
- Clinical trial outcome forecasting
- Predict the likelihood of clinical trial approval to help prioritise therapeutic candidates.
- Foundation for downstream fine-tuning
- Provide a pre-trained starting checkpoint for custom fine-tuning on proprietary laboratory assay 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
- Matches or exceeds best-in-class performance across 50 of 66 therapeutic benchmark tasks.
- Outperforms dedicated specialist models on 26 separate benchmark evaluations.
- Handles multiple biological representations within a single text-based framework, including SMILES and biopolymer sequences.
- Offers competitive data efficiency compared with larger general-purpose language models.
Watch out for
- Requires structured prompt templates rather than open natural language dialogue to achieve optimal predictive accuracy.
- Does not generate natural-language explanations or rationales for its predictions, unlike chat-tuned variants.
- Trained on public benchmark datasets, necessitating independent task-specific validation on intended 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 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 27B Predict differ from TxGemma-Chat models?
TxGemma 27B Predict is optimised for raw prediction accuracy using narrow prompt templates. TxGemma-Chat incorporates general instruction-tuning data to support conversational dialogue and explain prediction rationales, which comes at the expense of some predictive performance.
What input structure is required for inference?
Prompts should follow the Therapeutics Data Commons template structure, providing task instructions, scientific context, a specific question, and molecular identifiers such as SMILES strings or sequence data.
How was the model evaluated?
It was evaluated across 66 therapeutic tasks from the Therapeutics Data Commons using standard benchmark split strategies, including random, scaffold, cold-start, combination, and temporal splits.
Can the model be integrated into automated or agentic workflows?
Yes, it can run as an individual inference step using standard Hugging Face pipelines or be invoked as a tool within larger agentic workflows.
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, 2025, arXiv:2504.06196.
Written from TxGemma 27B 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.