TxGemma من Google DeepMind: نماذج مفتوحة المصدر تحدث ثورة في تطوير الأدوية العلاجية

Google DeepMind has unveiled TxGemma, a suite of open-source large language models (LLMs) designed to accelerate and improve the efficiency of therapeutic drug development. Building upon the foundation of the Gemma family of models, TxGemma aims to tackle the high costs, lengthy timelines, and high failure rates associated with traditional drug discovery methods. By leveraging the power of AI to predict and understand the properties of therapeutic entities throughout the development process, TxGemma promises to shorten the time from lab to bedside and reduce development expenses.

From Tx-LLM to TxGemma: An Open Approach

Following the success and interest in their previous model, Tx-LLM, Google DeepMind has developed TxGemma as its open-source successor. TxGemma offers developers the ability to adapt the models to their own therapeutic data and tasks, fostering innovation and collaboration.

• Model Sizes: TxGemma is available in three sizes: 2B, 9B, and 27B parameters, catering to different computational resource requirements.
• “Predict” Versions: Each size includes a “predict” version specifically fine-tuned for narrow tasks sourced from the Therapeutic Data Commons, such as predicting molecule toxicity.
• Task Coverage: These tasks cover a broad range, including classification (e.g., predicting blood-brain barrier permeability), regression (e.g., predicting drug binding affinity), and generation (e.g., generating reactant sets from a given product).
• Performance: The largest TxGemma model (27B predict version) demonstrates strong performance, surpassing or matching the capabilities of Tx-LLM and even rivaling or outperforming specialized models designed for single tasks.

Conversational AI for Enhanced Insights

In addition to the “predict” versions, TxGemma includes 9B and 27B “chat” versions designed for conversational therapeutic data analysis.

• Instruction Tuning: These models incorporate general instruction tuning data, enabling them to explain their reasoning, answer complex questions, and engage in multi-turn discussions.
• Explainability: Researchers can probe TxGemma-Chat for explanations behind its predictions, gaining deeper insights into the underlying mechanisms. For example, asking why a particular molecule is predicted to be toxic and receiving an explanation based on its structure.
• Trade-off: This conversational capability comes with a slight performance trade-off on raw therapeutic tasks compared to the “predict” versions.

Fine-Tuning and Customization

Recognizing the importance of customization, Google DeepMind provides a fine-tuning example Colab notebook to help developers adapt TxGemma to their specific therapeutic data and tasks.

• TrialBench Dataset: The notebook uses the TrialBench dataset to demonstrate fine-tuning for predicting adverse events in clinical trials.
• Proprietary Data Integration: Fine-tuning allows researchers to leverage their proprietary data, creating models tailored to their unique research needs.
• Improved Accuracy: This customization can potentially lead to even more accurate predictions, aiding researchers in assessing the safety and effectiveness of potential new therapies.

Agentic-Tx: Orchestrating Workflows for Advanced Discovery

Going beyond single-step predictions, Google DeepMind is showcasing how TxGemma can be integrated into agentic systems like Agentic-Tx to tackle more complex research challenges.

• Gemini 2.0 Pro Powered: Agentic-Tx, powered by Gemini 2.0 Pro, is a therapeutics-focused agentic system equipped with 18 tools.
• Diverse Toolset: The toolset includes TxGemma for multi-step reasoning, general search tools (PubMed, Wikipedia, Web), and specific molecular, gene, and protein tools.
• State-of-the-Art Performance: Agentic-Tx achieves state-of-the-art results on reasoning-intensive chemistry and biology tasks, as demonstrated by benchmarks like Humanity’s Last Exam and ChemBench.
• Workflow Orchestration: A Colab notebook is provided to demonstrate how Agentic-Tx can orchestrate complex workflows and answer multi-step research questions.

Getting Started with TxGemma

TxGemma is readily accessible on both Vertex AI Model Garden and Hugging Face, allowing researchers and developers to easily explore and experiment with the models. Google encourages exploration, experimentation, and feedback to drive further improvements. As an open model, TxGemma is designed to be further improved through fine-tuning with specific therapeutic development use-cases. The community’s contribution is anticipated to significantly accelerate therapeutic discovery.

In conclusion, Google DeepMind’s TxGemma represents a significant step forward in leveraging AI for therapeutic drug development. By providing open-source, customizable models and agentic systems, Google is empowering researchers and developers to accelerate the discovery of new therapies and improve patient outcomes.

Source: Google