BioGraphX: bridging the sequence-structure gap via physicochemical graph encoding for interpretable subcellular...

Computational protein subcellular localization prediction is vital for understanding cellular mechanisms and disease treatments. However, current methods lack interpretability: they predict where a protein localizes but fail to explain why. Moreover, understanding protein behaviour requires costly, time-consuming three-dimensional structures.

Here, we propose BioGraphX, a novel encoding framework that constructs protein interaction graphs directly from sequences using biochemical rules, providing a constraint-based structural proxy. Building upon this, BioGraphX-Net demonstrates superior performance on the DeepLoc 2.0 benchmark by integrating ESM-2 (Evolutionary Scale Modeling) embeddings with the proposed features via a gating mechanism. Gating analysis shows that while ESM-2 embeddings contribute strongly, BioGraphX features function as high-precision filters. SHAP (SHapley Additive exPlanations) analysis reveals feature importance patterns consistent with a sophisticated biophysical logic: sequence signals act as universal exclusion filters, while organelle-specific biophysical combinations enable precise compartment discrimination. Notably, Frustration features resolve targeting ambiguities in complex compartments, reflecting evolutionary constraints while preventing mislocalization from sequence mimicry. Cross-dataset validation on a protein solubility prediction task confirms the structural proxy captures genuine biophysical signal. Additionally, BioGraphX promotes Green AI in bioinformatics, matching state-of-the-art performance with a minimal parameter count of 13.46 million. In summary, BioGraphX provides accurate predictions and new insights into the language of life.

Source code is available at https://github.com/Abubakar-Saeed/BioGraphX.