MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202641093588 A) filed by National Institute Of Technology Karnataka on August 02, 2026, for Structure-Guided Cross-Attention And Multi-Level Structural Feature Integration Framework For Protein Subcellular Localization Prediction.
Inventors include Prajna Bhat; and Nagamma Patil.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: Title: Structure-Guided Cross-Attention and Multi-Level Structural Feature Integration Framework for Protein Subcellular Localization Prediction The disclosed invention relates to a framework for predicting protein subcellular localization using protein sequence information and three-dimensional structural information without relying on external biological databases. The framework generates residue-level sequence embeddings using a protein language model, extracts residue-level structural descriptors and whole-protein structural descriptors (230) and integrates the sequence-derived and structural representations through a structure-guided unidirectional cross-attention mechanism (250) wherein the structural descriptors selectively identify contextually relevant sequence features. The resulting structure-informed residue-level representations are aggregated using complementary mean pooling and maximum pooling operations (260) to generate a protein-level representation, which is subsequently fused (280) with whole-protein structural descriptors comprising geometric and electrostatic characteristics, including radius of gyration, globularity, and rotation-invariant descriptors derived from whole-protein electrostatic potential distributions. A classification module (290) processes the fused representation to generate single-label or multi-label protein subcellular localization predictions. The framework enables accurate localization prediction for known and previously uncharacterized proteins using only protein sequence and structural information.
Disclaimer: Curated by HT Syndication.
