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
{"title":"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","authors":"Alina Bazarova, Gennady M. Verkhivker","doi":"10.1021/acs.jcim.6c03114","DOIUrl":null,"url":null,"abstract":"<p>Allosteric regulation represents a fundamental mechanism\r\nof protein\r\nfunction, yet distinguishing allosteric from orthosteric protein-binding\r\nsites remains a persistent computational challenge. While multimodal\r\nprotein foundation models offer the potential to integrate complementary\r\nbiological signals, including sequence, structure, functional annotations,\r\nand conformational dynamics, their performance determinants in allosteric\r\nbinding site detection remain poorly understood. We introduce a unified\r\ncomputational framework for profiling multimodal protein foundation\r\nmodels across distinct binding site separability regimes. Rather than\r\nevaluating models solely by predictive accuracy, the framework combines\r\nsystematic modality embedding ablations, encoder architecture comparisons,\r\nand variance decomposition to characterize how evolutionary, structural,\r\nfunctional, and dynamical information contributes to allosteric site\r\ndiscrimination. Using the OneProt multimodal model, we evaluate two\r\ncomplementary levels of multimodal integration: (a) encoder architectures\r\nthat differ in the modalities incorporated during pretraining and\r\n(b) downstream combinations of pocket, sequence, and text embeddings\r\nused for classification. To systematically probe the determinants\r\nof model performance, we benchmark these configurations across four\r\nassembled data sets of protein complexes representing a spectrum of\r\nbiological complexity and a range of structural, dynamic, and evolutionary\r\ncontexts for orthosteric and allosteric binding sites. Through comprehensive\r\nembedding ablations, encoder architecture comparisons, and variance\r\ndecomposition, we demonstrate that model performance is governed primarily\r\nby intrinsic data set properties rather than architectural complexity,\r\nwith data set identity accounting for 63.7% of explainable variance.\r\nAcross the data sets examined, we identify substantial differences\r\nnot only in overall separability but also in the information accessible\r\nfrom different pretrained representations. Single-modality controls\r\nshow that high predictive performance does not necessarily imply reliance\r\non the explicitly provided pocket representation: depending on data\r\nset composition, substantial discrimination can already be obtained\r\nfrom sequence- or text-based multimodal representations. Within-protein\r\ncomparisons nevertheless demonstrate that pocket-containing representations\r\nretain binding site-specific discriminatory information when protein\r\nidentity and the associated sequence and text representations are\r\nheld constant. At the encoder level, structural and molecular dynamics\r\nencoders provide the greatest benefit in intermediate and high separability\r\nsettings. Example-level analysis further shows that combining data\r\nsets can reshape which binding sites are preferentially recognized.\r\nThese findings establish that the success of multimodal foundation\r\nmodels depends critically on alignment between available modalities\r\nand the biological signatures that distinguish regulatory classes\r\nin each data set.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 21","pages":"13913–13937"},"PeriodicalIF":6.4000,"publicationDate":"2026-11-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Chemical Information and Modeling ","FirstCategoryId":"92","ListUrlMain":"https://pubs.acs.org/doi/10.1021/acs.jcim.6c03114","RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2026/10/26 0:00:00","PubModel":"Epub","JCR":"Q1","JCRName":"CHEMISTRY, MEDICINAL","Score":null,"Total":0}
引用次数: 0
Abstract
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.
期刊介绍:
The Journal of Chemical Information and Modeling publishes papers reporting new methodology and/or important applications in the fields of chemical informatics and molecular modeling. Specific topics include the representation and computer-based searching of chemical databases, molecular modeling, computer-aided molecular design of new materials, catalysts, or ligands, development of new computational methods or efficient algorithms for chemical software, and biopharmaceutical chemistry including analyses of biological activity and other issues related to drug discovery.
Astute chemists, computer scientists, and information specialists look to this monthly’s insightful research studies, programming innovations, and software reviews to keep current with advances in this integral, multidisciplinary field.
As a subscriber you’ll stay abreast of database search systems, use of graph theory in chemical problems, substructure search systems, pattern recognition and clustering, analysis of chemical and physical data, molecular modeling, graphics and natural language interfaces, bibliometric and citation analysis, and synthesis design and reactions databases.