Protein ligand pocket prediction method based on spatial gate and local feature enhancement

By combining the selective state-space model Mamba with the encoder of the local feature enhancement module, the skip connection mechanism is optimized, which solves the problem of insufficient global modeling ability of existing deep learning methods in protein ligand binding pocket prediction and achieves more efficient and accurate prediction results.

CN122245403APending Publication Date: 2026-06-19JIANGNAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
JIANGNAN UNIV
Filing Date
2026-03-23
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing deep learning methods lack global modeling capabilities and local feature extraction capabilities in protein ligand binding pocket prediction, and suffer from redundant information transmission problems, resulting in insufficient prediction accuracy.

Method used

We employ a deep learning approach based on spatial gating and local feature enhancement, combined with the selective state space model Mamba. By integrating the residual Mamba module and the local feature enhancement module into an encoder, we achieve collaborative representation of global and local features. Furthermore, we optimize the skip connection mechanism through the spatial enhancement Mamba gating module to enhance the transmission of key spatial features.

Benefits of technology

It improves the accuracy and efficiency of protein binding pocket prediction, enhances the feature reconstruction capabilities of global semantics and local details, reduces redundant information transmission, and improves prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122245403A_ABST
    Figure CN122245403A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of intelligent cell biometrics, specifically involving a protein ligand pocket prediction method based on spatial gating and local feature enhancement. The method comprises four stages: data preprocessing and initial feature extraction, multi-scale feature encoding, spatially gating cross-layer feature fusion, and segmentation prediction based on depthwise separable convolution. This method integrates a Mamba module and a local feature enhancement module in the multi-scale feature encoding stage to achieve global modeling of protein features and adaptive enhancement of local features. Simultaneously, in the spatially gating cross-layer feature fusion skip connection stage, a spatially enhanced Mamba gating module is constructed to filter redundant information and enhance the fusion of multi-scale features. This invention effectively captures the long-range semantic information of proteins and intelligently filters and enhances cross-layer features, ultimately achieving protein ligand binding pocket prediction by generating a voxel-level probability map through depthwise separable convolution.
Need to check novelty before this filing date? Find Prior Art