Systematic Profiling of Encoder Architectures and Multimodal Embeddings in Protein Foundation Models Reveals Data Set-Driven Separability Regimes in Prediction of Allosteric and Orthosteric Binding Sites

Allosteric regulation represents a fundamental mechanism of protein function, yet distinguishing allosteric from orthosteric protein-binding sites remains a persistent computational challenge. While multimodal protein foundation models offer the potential to integrate complementary biological signals, including sequence, structure, functional annotations, and conformational dynamics, their performance determinants in allosteric binding site detection remain poorly understood. We introduce a unified computational framework for profiling multimodal protein foundation models across distinct binding site separability regimes. Rather than evaluating models solely by predictive accuracy, the framework combines systematic modality embedding ablations, encoder architecture comparisons, and variance decomposition to characterize how evolutionary, structural, functional, and dynamical information contributes to allosteric site discrimination. Using the OneProt multimodal model, we evaluate two complementary levels of multimodal integration: (a) encoder architectures that differ in the modalities incorporated during pretraining and (b) downstream combinations of pocket, sequence, and text embeddings used for classification. To systematically probe the determinants of model performance, we benchmark these configurations across four assembled data sets of protein complexes representing a spectrum of biological complexity and a range of structural, dynamic, and evolutionary contexts for orthosteric and allosteric binding sites. Through comprehensive embedding ablations, encoder architecture comparisons, and variance decomposition, we demonstrate that model performance is governed primarily by intrinsic data set properties rather than architectural complexity, with data set identity accounting for 63.7% of explainable variance. Across the data sets examined, we identify substantial differences not only in overall separability but also in the information accessible from different pretrained representations. Single-modality controls show that high predictive performance does not necessarily imply reliance on the explicitly provided pocket representation: depending on data set composition, substantial discrimination can already be obtained from sequence- or text-based multimodal representations. Within-protein comparisons nevertheless demonstrate that pocket-containing representations retain binding site-specific discriminatory information when protein identity and the associated sequence and text representations are held constant. At the encoder level, structural and molecular dynamics encoders provide the greatest benefit in intermediate and high separability settings. Example-level analysis further shows that combining data sets can reshape which binding sites are preferentially recognized. These findings establish that the success of multimodal foundation models depends critically on alignment between available modalities and the biological signatures that distinguish regulatory classes in each data set.