Shomaji, S. | Search bioRxiv

bioRxiv2025

A Concept-Driven Disentanglement Framework for Interpretable Graph Neural Networks in Structure-Function Coupling

Graph Neural Networks (GNNs) achieve state-of-the-art performance in predicting brain functional connectivity (FC) from structural connectivity (SC), yet their "black-box" nature limits interpretability and scientific utility. We present a concept-driven disentanglement framework that builds inherently interpretable GNNs for quantitative hypothesis testing of structure-function relationships. The framework employs an ensemble of GNN branches, each architecturally biased to learn from a predefined structural concept (e.g., strong vs. weak connections) by processing a filtered version of the SC graph. This design enforces verifiably disentangled node embeddings, ensuring each branch captures a distinct structural feature. Using SHAP (Shapley additive explanations), we quantify the predictive contribution of each concept and assess its statistical significance against null distributions. Our framework demonstrates high predictive accuracy for FC, achieving a group-level correlation coefficient of 0.91 on a public human connectome dataset, while simultaneously yielding interpretable neuroscientific insights. This interpretable-by-design methodology bridges the gap between predictive power and scientific transparency, enabling deep learning models to provide mechanistic insights into the brains network organization.

neuroscience↗

bioRxiv2023

MetaLLM: Residue-wise Metal ion Prediction UsingDeep Transformer Model

Proteins bind to metals such as copper, zinc, magnesium, etc., serving various purposes such as importing, exporting, or transporting metal in other parts of the cell as ligands and maintaining stable protein structure to function properly. A metal binding site indicates the single amino acid position where a protein binds a metal ion. Manually identifying metal binding sites is expensive, laborious, and time-consuming. A tiny fraction of the millions of proteins in UniProtKB - the most comprehensive protein database - are annotated with metal binding sites, leaving many millions of proteins waiting for metal binding site annotation. Developing a computational pipeline is thus essential to keep pace with the growing number of proteins. A significant shortcoming of the existing computational methods is the consideration of the long-term dependency of the residues. Other weaknesses include low accuracy, absence of positional information, hand-engineered features, and a pre-determined set of residues and metal ions. In this paper, we propose MetaLLM, a metal binding site prediction technique, by leveraging the recent progress in self-supervised attention-based (e.g. Transformer) large language models (LLMs) and a considerable amount of protein sequences publicly available. LLMs are capable of modelling long residual dependency in a sequence. The proposed MetaLLM uses a transformer pre-trained on an extensive database of protein sequences and later fine-tuned on metal-binding proteins for multi-label metal ions prediction. A 10-fold cross-validation shows more than 90% precision for the most prevalent metal ions.

bioinformatics↗