BioGraphX-Sol: A Structure-Free, Fully Interpretable Protein Solubility Predictor Exceeds AlphaFold3-Based Methods and Reveals Nine Universal Cross-Species Determinants

Abstract Protein solubility is a critical determinant of recombinant expression yield, yet current state-of-the-art predictors often rely on either AlphaFold3-predicted structures or large ensembles of protein language models, limiting accessibility and interpretability. We present BioGraphX-Sol, a sequence-only predictor that constructs biophysically grounded constraint graphs directly from amino acid sequences to extract 200 physicochemical features spanning graph topology, IDP and polymer physics, aggregation-hotspot geometry, and novel surface- and terminal-conditioned descriptors, without requiring 3D coordinates, multiple sequence alignments, or pretrained models. On the standard eSOL benchmark, the XGB+ESM post-hoc ensemble essentially matches ProtSATT in predictive performance ($R^2 = 0.5451$ vs.\$R^2 = 0.5450$), despite training only 1.44M neural parameters and requiring no structure prediction. The physics-only XGBoost configuration ($R^2 = 0.499$, ROC--AUC $= 0.873$) exceeds FGNNSol ($R^2 = 0.487$, ROC--AUC $= 0.867$), which combines AlphaFold3-derived 3D coordinates with ESM-C embeddings and structural descriptors across roughly 1,831 input dimensions, showing that explicit 3D structure prediction is unnecessary when the relevant physicochemical information is correctly encoded. A post-hoc $\alpha$ sweep quantifies the physicochemical-to-evolutionary information balance at approximately 59\%/41\%, providing, to our knowledge, the first quantitative decomposition of information sources for protein solubility prediction. In zero-shot transfer to \textit{S.\cerevisiae}, the physics-only XGBoost achieves the highest ROC--AUC ($0.8524$) and PR--AUC ($0.9246$) among all evaluated methods. Permutation-importance analysis identifies nine universal solubility determinants exhibiting consistent directional effects across \textit{E.\coli} and \textit{S.\cerevisiae} (9/9 agreement). The physics-only configuration requires no GPU at inference and runs efficiently on standard CPU hardware, supporting the Green AI principle that compact, interpretable representations grounded in domain knowledge can provide a viable alternative to parameter scaling.