TLDR: GALAX is a novel AI framework that combines Large Language Models (LLMs) with Graph Neural Networks (GNNs) using reinforcement learning, guided by a Graph Process Reward Model (GPRM). It generates explainable, disease-relevant subgraphs from multi-omic data and biological knowledge, enabling more accurate and interpretable identification of therapeutic targets in precision medicine without requiring explicit intermediate reasoning annotations. A new benchmark, Target-QA, was also introduced to facilitate its evaluation.
In the evolving landscape of precision medicine, understanding complex biological interactions is crucial for developing effective treatments. Traditional methods often fall short, either by overlooking the intricate network of molecular features, lacking quantitative reasoning, or failing to fully leverage the rich information embedded in biological data and scientific literature. This challenge is particularly evident in identifying disease-critical pathways and therapeutic targets, especially in complex diseases like cancer.
A new framework called GALAX (Graph-Augmented Language Model with Explainability) has been introduced to bridge these gaps. Developed by a team including Heming Zhang, Di Huang, and Fuhai Li, GALAX aims to provide a more comprehensive and interpretable approach to target and pathway discovery in precision medicine. The core idea behind GALAX is to integrate diverse data types—quantitative multi-omic features (like genomics, transcriptomics, and proteomics), topological context from biological networks, and vast textual biological knowledge—into a unified reasoning process.
Existing approaches have limitations. For instance, models focusing solely on numerical omics data might miss the crucial topological relationships within biological networks. Text-based Large Language Models (LLMs), while powerful in understanding language, often struggle with quantitative reasoning and can sometimes “hallucinate” information not grounded in facts. Graph-only models, on the other hand, might not fully utilize the rich semantic information associated with nodes (like genes or proteins) or the generalization capabilities of LLMs.
GALAX addresses these issues by combining the strengths of pretrained Graph Neural Networks (GNNs) with Large Language Models (LLMs) through a sophisticated reinforcement learning mechanism. This mechanism is guided by a Graph Process Reward Model (GPRM). Unlike previous Process Reward Models (PRMs) that can be limited by vague step definitions and unreliable intermediate evaluations, GALAX uses the GNN as a step-wise supervisor. This means the GNN evaluates intermediate subgraphs (which represent partial signaling pathways) for their biological plausibility and relevance to the disease, providing fine-grained, graph-based rewards without needing explicit manual annotations for each reasoning step.
The process begins with an LLM proposing candidate targets based on multi-omic profiles and existing knowledge graphs. Then, a reinforcement learning-based graph generator iteratively builds disease-relevant subgraphs. These subgraphs are essentially simplified, yet mechanistically grounded, patient-specific networks that highlight potential therapeutic targets. This step-wise construction, evaluated by the pretrained GNN, translates the LLM’s language-based reasoning into an interpretable graph structure, offering clear insights into the disease mechanism.
To evaluate GALAX, the researchers also developed a new benchmark dataset called Target-QA. This dataset combines CRISPR-identified targets, multi-omic profiles, and biomedical graph knowledge across various cancer cell lines. Target-QA facilitates the pretraining of GNNs for supervising graph construction and supports complex reasoning over text-numeric graphs, making it a valuable resource for the field. Both GALAX and Target-QA are publicly available on Huggingface and GitHub, encouraging further research and development.
The experimental results show that GALAX significantly outperforms existing baselines across various evaluation metrics, including precision, recall, F1-score, and Hit@K (Hit@5 and Hit@10). For example, it achieved an overall precision of 0.5472 and recall of 0.5332, surpassing the strongest baseline. This robust performance was observed across different cancer types, such as lung adenocarcinoma (LUAD) and breast cancer (BRCA), demonstrating its generalizability. The generated subgraphs also provide explainable insights, with functional enrichment analysis confirming their relevance to known cancer-associated signaling pathways, including those involving established therapeutic targets like EGFR.
Also Read:
- Small Language Models Drive Accurate Genomics Research with Nano Bio-Agents
- New AI Model Predicts Oligopeptide-Infectious Disease Links
In essence, GALAX offers a scalable and biologically grounded framework for explainable, reinforcement-guided subgraph reasoning. This innovation promises to lead to more reliable and interpretable discovery of therapeutic targets and disease mechanisms in precision medicine. For more technical details, you can refer to the full research paper.
