GeoEPred: A multimodal structure-aware geometric deep learning framework for Gram-negative bacterial secreted effector prediction with sequence semantics.

GeoEPred achieves over 10% higher accuracy than existing models in predicting Gram-negative bacterial effector proteins by integrating sequence and structural data.

• Why it matters: Accurate effector protein prediction is crucial for understanding bacterial pathogenicity and developing targeted anti-infective therapies, but current methods struggle to fully capture complex conformational signals.
• What they did: The framework combines pretrained protein language model embeddings with 3D structural predictions from ESMFold, employing geometric vector perceptrons and a cross-modal attention module to model sequence and structure synergistically.
• The result: GeoEPred outperforms existing methods across multiple benchmark tasks, demonstrating strong generalization, stability in remote homolog recognition, and significant potential for large-scale effector discovery.