ProLoc-IHS: Multi-label protein subcellular localization based on immunohistochemical images and sequence information

Area of Science:

• Biotechnology
• Computational Biology
• Proteomics

Background:

• Immunohistochemistry (IHC) imaging is crucial for studying human protein subcellular localization (SCL) in tissues.
• Manual SCL annotation is labor-intensive and limits dataset size, necessitating automated computational tools.
• Existing computational models often overlook valuable protein sequence information.

Purpose of the Study:

• To develop a novel computational model, ProLoc-IHS, for predicting protein SCL using IHC images.
• To integrate both visual (IHC image) and sequential (protein sequence) data for improved prediction accuracy.
• To create a new bimodal dataset for training and evaluating the SCL prediction model.

Main Methods:

• A bimodal dataset was curated from the Human Protein Atlas (HPA) and UniProt, comprising IHC images and corresponding protein sequences.
• ProLoc-IHS employs Vision Transformer (Vit) for image embeddings and ProtT5 for protein sequence embeddings.
• Embeddings are fused via a cross-attention module, followed by a feature learning module with multi-head attention and residual connections. Binary cross entropy and Focal loss were used for multi-label classification.

Main Results:

• ProLoc-IHS demonstrated superior performance compared to existing prediction models.
• The model effectively combines visual and sequential features for enhanced SCL prediction.
• A novel dataset and the ProLoc-IHS code were made publicly available.

Conclusions:

• ProLoc-IHS offers a significant advancement in automated protein SCL prediction from IHC images.
• Integrating protein sequence data substantially improves the accuracy of computational SCL analysis.
• The developed tool and dataset facilitate large-scale analysis of protein localization in biological and pathological contexts.