Presentation Information
[1ASBA-13]From Prediction to Discovery: AI as an Observatory of Physical and Evolutionary Constraints in Proteins
○Qian-Yuan Tang1 (1. Hong Kong Baptist University (Hong Kong))
Keywords:
Protein dynamics,Protein evolution,Genotype-Phenotype Mapping,Protein language models,AlphaFold,Biocomplexity
[Purpose]
Artificial Intelligence is reshaping protein science not only by improving prediction accuracy, but also by providing large-scale, structured datasets that serve as scientific objects in their own right. We propose a constraint-based perspective that treats AI-derived data—including AlphaFold-predicted structures and ESM language model embeddings—as an "observational substrate." By analyzing these data, we aim to uncover how protein folding, dynamics, and function emerge as different projections of a shared, low-dimensional organization governed by fundamental physical and evolutionary constraints.
[Method]
Our approach integrates the AlphaFold Protein Structure Database (AlphaFold DB) with sequence-based representations from protein language models (ESM). We perform comparative analyses across diverse taxonomic groups using Normal Mode Analysis (NMA), scaling analyses, and representation decomposition of embedding spaces. We further quantify these constraints through residue contact network topology, sequence segregation analysis, and by mapping statistical structural proximity against phylogenetic trees to trace parallel evolutionary trajectories.
[Results]
Our analysis reveals that higher organismal complexity correlates with proteome-wide shifts toward larger radii of gyration, higher coil fractions, and lower fractal dimensions, suggesting increased functional specialization. We uncover a robust scaling relationship between folding topology (contact order) and native dynamics (fluctuation entropy): long-range contacts that slow folding simultaneously restrict conformational flexibility. As complexity increases, residue contact networks become more assortative and proteins exhibit higher hydrophilic–hydrophobic segregation. Furthermore, statistical structural proximity across proteomes mirrors phylogenetic trends, indicating a synchronized trend of protein and organism evolution.
[Consideration]
A central question in structural biology is the genotype-phenotype mapping of proteins. Our investigations into the linear responses of native proteins demonstrate a fundamental dynamics-evolution correspondence: noise-induced protein dynamics and mutation-induced variations of native structures are quantitatively correlated. This correspondence arises because both noise- and mutation-induced deformations are restricted to a common low-dimensional subspace. This shared manifold explains the evolutionary mechanism by which proteins simultaneously gain dynamical flexibility and evolutionary structural variability. By bridging AI-derived representation spaces (ESM) with these physical dynamical landscapes, we show that the constraints governing stability and folding are intrinsically coupled to the functional and evolutionary potential of the proteome.
[Conclusion]
AI models act as data-rich interfaces that make otherwise hidden organizing principles accessible. By revealing the low-dimensional organization that unifies folding, dynamics, and evolutionary variation, this study provides new insights into how protein functional diversity increases while dynamical dimensionality reduces during evolution. This work points toward a fundamental shift in biophysics: from the prediction of individual protein properties to the constraint-based discovery of the universal principles governing the origin and evolution of life.
Artificial Intelligence is reshaping protein science not only by improving prediction accuracy, but also by providing large-scale, structured datasets that serve as scientific objects in their own right. We propose a constraint-based perspective that treats AI-derived data—including AlphaFold-predicted structures and ESM language model embeddings—as an "observational substrate." By analyzing these data, we aim to uncover how protein folding, dynamics, and function emerge as different projections of a shared, low-dimensional organization governed by fundamental physical and evolutionary constraints.
[Method]
Our approach integrates the AlphaFold Protein Structure Database (AlphaFold DB) with sequence-based representations from protein language models (ESM). We perform comparative analyses across diverse taxonomic groups using Normal Mode Analysis (NMA), scaling analyses, and representation decomposition of embedding spaces. We further quantify these constraints through residue contact network topology, sequence segregation analysis, and by mapping statistical structural proximity against phylogenetic trees to trace parallel evolutionary trajectories.
[Results]
Our analysis reveals that higher organismal complexity correlates with proteome-wide shifts toward larger radii of gyration, higher coil fractions, and lower fractal dimensions, suggesting increased functional specialization. We uncover a robust scaling relationship between folding topology (contact order) and native dynamics (fluctuation entropy): long-range contacts that slow folding simultaneously restrict conformational flexibility. As complexity increases, residue contact networks become more assortative and proteins exhibit higher hydrophilic–hydrophobic segregation. Furthermore, statistical structural proximity across proteomes mirrors phylogenetic trends, indicating a synchronized trend of protein and organism evolution.
[Consideration]
A central question in structural biology is the genotype-phenotype mapping of proteins. Our investigations into the linear responses of native proteins demonstrate a fundamental dynamics-evolution correspondence: noise-induced protein dynamics and mutation-induced variations of native structures are quantitatively correlated. This correspondence arises because both noise- and mutation-induced deformations are restricted to a common low-dimensional subspace. This shared manifold explains the evolutionary mechanism by which proteins simultaneously gain dynamical flexibility and evolutionary structural variability. By bridging AI-derived representation spaces (ESM) with these physical dynamical landscapes, we show that the constraints governing stability and folding are intrinsically coupled to the functional and evolutionary potential of the proteome.
[Conclusion]
AI models act as data-rich interfaces that make otherwise hidden organizing principles accessible. By revealing the low-dimensional organization that unifies folding, dynamics, and evolutionary variation, this study provides new insights into how protein functional diversity increases while dynamical dimensionality reduces during evolution. This work points toward a fundamental shift in biophysics: from the prediction of individual protein properties to the constraint-based discovery of the universal principles governing the origin and evolution of life.
Comment
To browse or post comments, you must log in.Log in