README.md

---
license: other
license_name: health-ai-developer-foundations
license_link: https://developers.google.com/health-ai-developer-foundations/terms
language:
  - en
pipeline_tag: text-generation
library_name: transformers
tags:
  - therapeutics
  - drug-development
extra_gated_heading: Access TxGemma on Hugging Face
extra_gated_prompt: >-
  To access TxGemma on Hugging Face, you're required to review and
  agree to [Health AI Developer Foundation's terms of use](https://developers.google.com/health-ai-developer-foundations/terms).
  To do this, please ensure you're logged in to Hugging Face and click below.
  Requests are processed immediately.
extra_gated_button_content: Acknowledge license
---

# TxGemma model card

**Model documentation:** [TxGemma](https://developers.google.com/health-ai-developer-foundations/txgemma)

**Resources:**

*   Model on Google Cloud Model Garden: [TxGemma](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/txgemma)
*   Model on Hugging Face: [TxGemma](https://huggingface.co/collections/google/txgemma-release-67dd92e931c857d15e4d1e87)
*   GitHub repository (supporting code, Colab notebooks, discussions, and
    issues):
    [TxGemma](https://github.com/google-gemini/gemma-cookbook/tree/main/TxGemma)
*   Quick start notebook: [notebooks/quick\_start](https://github.com/google-gemini/gemma-cookbook/blob/main/TxGemma/%5BTxGemma%5DQuickstart_with_Hugging_Face.ipynb)
*   Support: See
    [Contact](https://developers.google.com/health-ai-developer-foundations/txgemma/get-started.md#contact).

Terms of use: [Health AI Developer Foundations terms of
use](https://developers.google.com/health-ai-developer-foundations/terms)

**Author:** Google

## Model information

This section describes the TxGemma model and how to use it.

### Description

TxGemma is a collection of lightweight, state-of-the-art, open language models
built upon Gemma 2, fine-tuned for therapeutic development. It comes in 3 sizes,
2B, 9B, and 27B.

TxGemma models are designed to process and understand information related to
various therapeutic modalities and targets, including small molecules, proteins,
nucleic acids, diseases, and cell lines. TxGemma excels at tasks such as
property prediction, and can serve as a foundation for further fine-tuning or as
an interactive, conversational agent for drug discovery. The model is fine-tuned
from Gemma 2 using a diverse set of instruction-tuning datasets, curated from
the [Therapeutics Data Commons (TDC)](https://tdcommons.ai/).

TxGemma is offered as both a prediction model that expects a narrow form of
prompting and for the 9B and 27B version, conversational models that are more
flexible and can be used in multi-turn interactions, including to explain its
rationale behind a prediction. This conversational model comes at the expense of
some raw prediction performance. See our [manuscript](https://arxiv.org/abs/2504.06196)
for more information.

**Key Features**

*   Versatility: Exhibits strong performance across a wide range of therapeutic
    tasks, outperforming or matching best-in-class performance on a significant
    number of benchmarks.
*   Data Efficiency: Shows competitive performance even with limited data
    compared to larger models, offering improvements over its predecessors.
*   Conversational Capability (TxGemma-Chat): Includes conversational variants
    that can engage in natural language dialogue and explain the reasoning
    behind their predictions.
*   Foundation for Fine-tuning: Can be used as a pre-trained foundation for
    specialized use cases.

**Potential Applications**

TxGemma can be a valuable tool for researchers in the
following areas:

*   Accelerated Drug Discovery: Streamline the therapeutic development process
    by predicting properties of therapeutics and targets for a wide variety of
    tasks including target identification, drug-target interaction prediction,
    and clinical trial approval prediction.

### How to use

Below are some example code snippets to help you quickly get started running the
model locally on GPU. If you want to use the model to run inference on a large
number of inputs, we recommend that you create a production version using [Model
Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/txgemma).

#### Formatting prompts for therapeutic tasks

```python
import json
from huggingface_hub import hf_hub_download

# Load prompt template for tasks from TDC
tdc_prompts_filepath = hf_hub_download(
    repo_id="google/txgemma-27b-predict",
    filename="tdc_prompts.json",
)
with open(tdc_prompts_filepath, "r") as f:
    tdc_prompts_json = json.load(f)

# Set example TDC task and input
task_name = "BBB_Martins"
input_type = "{Drug SMILES}"
drug_smiles = "CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21"

# Construct prompt using template and input drug SMILES string
TDC_PROMPT = tdc_prompts_json[task_name].replace(input_type, drug_smiles)
print(TDC_PROMPT)
```

The resulting prompt is in the format expected by the model:

```none
Instructions: Answer the following question about drug properties.
Context: As a membrane separating circulating blood and brain extracellular fluid, the blood-brain barrier (BBB) is the protection layer that blocks most foreign drugs. Thus the ability of a drug to penetrate the barrier to deliver to the site of action forms a crucial challenge in development of drugs for central nervous system.
Question: Given a drug SMILES string, predict whether it
(A) does not cross the BBB (B) crosses the BBB
Drug SMILES: CN1C(=O)CN=C(C2=CCCCC2)c2cc(Cl)ccc21
Answer:
```

#### Running the model on predictive tasks

```python
# pip install accelerate transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model directly from Hugging Face Hub
tokenizer = AutoTokenizer.from_pretrained("google/txgemma-27b-predict")
model = AutoModelForCausalLM.from_pretrained(
    "google/txgemma-27b-predict",
    device_map="auto",
)

# Formatted TDC prompt (see "Formatting prompts for therapeutic tasks" section above)
prompt = TDC_PROMPT

# Prepare tokenized inputs
input_ids = tokenizer(prompt, return_tensors="pt").to("cuda")

# Generate response
outputs = model.generate(**input_ids, max_new_tokens=8)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

Alternatively, you can use the `pipeline` API, which provides a simple way to
run inference while abstracting away complex details of loading and using the
model and tokenizer:

```python
# pip install transformers
from transformers import pipeline

# Instantiate a text generation pipeline using the model
pipe = pipeline(
    "text-generation",
    model="google/txgemma-27b-predict",
    device="cuda",
)

# Formatted TDC prompt (see "Formatting prompts for therapeutic tasks" section above)
prompt = TDC_PROMPT

# Generate response
outputs = pipe(prompt, max_new_tokens=8)
response = outputs[0]["generated_text"]
print(response)
```

### Examples

See the following Colab notebooks for examples of how to use TxGemma:

*   To give the model a quick try, running it locally with weights from Hugging
    Face, see [Quick start notebook in
    Colab](https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/TxGemma/%5BTxGemma%5DQuickstart_with_Hugging_Face.ipynb),
    which includes some example eval tasks from TDC.
*   For a demo of how to fine-tune TxGemma in Hugging Face, see our [Fine-tuning notebook in Colab](https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/TxGemma/%5BTxGemma%5DFinetune_with_Hugging_Face.ipynb).
*   For a demo of how TxGemma can be used as a tool as part of a larger agentic
    workflow powered by Gemini 2 see the [Agentic workflow notebook in Colab](https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/TxGemma/%5BTxGemma%5DAgentic_Demo_with_Hugging_Face.ipynb).

### Model architecture overview

*   TxGemma is based on the Gemma 2 family of lightweight, state-of-the-art open
    LLMs. It utilizes a decoder-only transformer architecture.
*   Base Model: Gemma 2 (2B, 9B, and 27B parameter versions).
*   Fine-tuning Data: Therapeutics Data Commons, a collection of
    instruction-tuning datasets covering diverse therapeutic modalities and
    targets.
*   Training Approach: Instruction fine-tuning using a mixture of therapeutic
    data (TxT) and, for conversational variants, general instruction-tuning
    data.
*   Conversational Variants: TxGemma-Chat models (9B and 27B) are trained with a
    mixture of therapeutic and general instruction-tuning data to maintain
    conversational abilities.

### Technical Specifications

*   Model type: Decoder-only Transformer (based on Gemma 2)
*   Key publication: [TxGemma: Efficient and Agentic LLMs for Therapeutics](https://arxiv.org/abs/2504.06196)
*   Model created: 2025-03-18 (From the TxGemma Variant Proposal)
*   Model Version: 1.0.0

### Performance & Validation

TxGemma's performance has been validated on a comprehensive
[benchmark](https://tdcommons.ai/) of 66 therapeutic tasks derived from TDC.

### Key performance metrics

*   Aggregated Improvement: Improves over the original [Tx-LLM
    paper](https://arxiv.org/abs/2406.06316) on 45 out of 66 therapeutic tasks.

*   Best-in-Class Performance: Surpasses or matches best-in-class performance on
    50 out of 66 tasks, exceeding specialist models on 26 tasks. See [Table A.11](https://arxiv.org/abs/2504.06196)
    of the TxGemma paper for the full breakdown.

### Inputs and outputs

*   **Input:** Text. For best performance, text prompts should be formatted
    according to the TDC structure, including
    instructions, context, question, and, optionally, few-shot examples. Inputs
    can include SMILES strings, amino acid sequences, nucleotide sequences, and
    natural language text.

*   **Output:** Text.

## Dataset details

### Training dataset

**Therapeutics Data Commons**: A curated collection of instruction-tuning
datasets covering 66 tasks spanning the discovery and development of safe and
effective medicine. This includes over 15 million data points across different
biomedical entities. Released TxGemma models are only trained on datasets with
commercial licenses, whereas models in our publication are also trained on
datasets with non-commercial licenses.

**General Instruction-Tuning Data**: Used for TxGemma-Chat in combination with
TDC.

### Evaluation dataset

**Therapeutics Data Commons**: The same 66 tasks used for training are used
for evaluation, following TDC's recommended methodologies for data splits
(random, scaffold, cold-start, combination, and temporal).

## License

The use of TxGemma is governed by the [Health AI Developer Foundations terms of
use](https://developers.google.com/health-ai-developer-foundations/terms).

## Implementation information

This section contains details about the model internals.

### Software

Training was done using [JAX](https://github.com/jax-ml/jax).

JAX allows researchers to take advantage of the latest generation of hardware,
including TPUs, for faster and more efficient training of large models.

## Use and limitations

### Intended use

*   Research and development of therapeutics.

### Benefits

TxGemma provides a versatile and powerful tool for accelerating therapeutic
development. It offers:

*   Strong performance across a wide range of tasks.
*   Data efficiency compared to larger models.
*   A foundation for further fine-tuning from private data.
*   Integration into agentic workflows.

### Limitations

*   Trained on public data from TDC.
*   Task-specific validation remains an important aspect of downstream model
    development by the end user.
*   As with any research, developers should ensure that any downstream
    application is validated to understand performance using data that is
    appropriately representative of the intended use setting for the specific
    application (e.g., age, sex, gender, condition, scanner, etc.).

## Citation

```bibtex
@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},
}
```

Find the paper [here](https://arxiv.org/abs/2504.06196).