Wenfei Li

Protein language models (pLMs) learn rich sequence semantics from vast volume of protein sequences, yet the information encoded in their high-dimensional embeddings remains largely a black box, posing an interpretability challenge. Here we show that semantic features learned by pLMs can be decoded to reveal biologically meaningful signals using coevolutionary analysis without downstream supervised-training. Building on this insight, we developed MaskNGly, a zero-shot framework that identifies protein N‑glycosylation sites directly from PLM-embeddings while achieving performance comparable to state-of-the-art supervised predictors. MaskNGly is computationally efficient and intrinsically transferable, enabling omics-scale characterization of protein glycosylation landscape. Applied across proteomes, it reveals that canonical N-X-S/T motifs gain higher glycosylation propensities with increasing organismal complexity and that glycosylation preferentially occurs at structured, solvent‑exposed regions. Mutations altering high-scoring motifs are more frequently disease-associated. The framework also captures signatures of glycosylation‑mediated protein allostery. These results establish an interpretable and scalable route to study glycobiology from pLM-embeddings.

Biomolecular condensates play essential roles in cellular organization and are implicated in diverse pathological processes. Their formation is driven by liquid–liquid phase separation (LLPS), a process that requires coordinated multistep desolvation of biomolecular chains and multivalent inter-chain interactions. Although coarse-grained (CG) models with implicit solvent are widely used to probe LLPS thermodynamics and kinetics, they typically neglect water-mediated desolvation effects, limiting their accuracy and mechanistic interpretability. Here, guided by all-atom simulations and experimental measurements, we develop a desolvation-aware implicit-solvent CG model by incorporating residue-level desolvation terms directly into the pairwise energy function, and apply it to investigate LLPS of intrinsically disordered proteins. Incorporating these desolvation interactions reshapes the phase diagram, alleviating dense-phase overcompaction. Notably, we observe an approximately linear correlation between the temperature gap (simulation temperature relative to the critical point) and the extent of conformational expansion accompanying the dilute-to-dense phase transition, a result further supported by theoretical analysis. We also find that desolvation barriers slow early density-fluctuation growth and shorten transient kinetic arrest, whereas solvent-separated contact interactions exert the opposite effects. Both terms further modulate chain mobility within mature condensates through competing packing and energy-landscape effects. Together, this framework enables an efficient representation of desolvation in CG simulations and reveals how desolvation energetics shape both the thermodynamic landscape and kinetic properties of biomolecular LLPS.

Biomolecular condensates play essential roles in cellular organization and are implicated in diverse pathological processes. Their formation is driven by liquid-liquid phase separation (LLPS), a process that requires coordinated multistep desolvation of biomolecular chains and multivalent inter-chain interactions. Although coarse-grained (CG) models with implicit solvent are widely used to probe LLPS thermodynamics and kinetics, they typically neglect water-mediated desolvation effects, limiting their accuracy and mechanistic interpretability. Here, guided by all-atom simulations and experimental measurements, we develop a desolvation-aware implicit-solvent CG model by incorporating residue-level desolvation terms directly into the pairwise energy function and apply it to investigate LLPS of intrinsically disordered proteins. Incorporating these desolvation interactions reshapes the phase diagram, alleviating dense-phase overcompaction. Notably, we observe an approximately linear correlation between the temperature gap (simulation temperature relative to the critical point) and the extent of conformational expansion accompanying the dilute-to-dense phase transition, a result further supported by theoretical analysis. We also find that desolvation barriers slow early density-fluctuation growth and shorten transient kinetic arrest, whereas solvent-separated contact interactions exert the opposite effects. Both terms further modulate chain mobility within mature condensates through competing packing and energy-landscape effects. Together, this framework enables an efficient representation of desolvation in CG simulations and reveals how desolvation energetics shape both the thermodynamic landscape and kinetic properties of biomolecular LLPS.

Enzymes within cells often function in crowded environments due to the high concentration of macromolecules and the formation of biomolecular condensates. However, most enzyme kinetics insights stem from dilute solution studies. In this study, we examined how molecular crowding impacts enzyme catalysis by constructing a residue-resolved dynamic energy landscape model that explicitly accounts for the crowding agent. By performing molecular simulations for the full enzymatic cycle of a model enzyme adenylate kinase, we revealed a crowding-induced allosteric regulation mechanism of enzymatic catalysis. The crowding agent modifies the conformational distribution of the enzyme, which in turn alters the enzymatic dynamics. Additional simulations of AdK variants with various intrinsic conformational propensities showed that the effect of molecular crowding depends on the rate-limiting step in the enzymatic cycle. Molecular crowding increases enzymatic activity if the rate-limiting step involves the formation of the catalytically competent enzyme-substrate complex and decreases the activity if the rate-limiting step involves the product release. The results revealed in this work not only shed insights into the general biophysical principle of protein dynamics under a crowded cellular environment but also provide a computational framework to understand the diversity of existing experimental observations of the molecular crowding effects on enzyme catalysis.

Biomolecular condensates play essential roles in cellular organization and are implicated in diverse pathological processes. Their formation is driven by liquid-liquid phase separation (LLPS), a process that requires coordinated multistep desolvation of biomolecular chains and multivalent inter-chain interactions. Although coarse-grained (CG) models with implicit solvent are widely used to probe LLPS thermodynamics and kinetics, they typically neglect explicit desolvation energetics, limiting their accuracy and mechanistic interpretability. Here, guided by all-atom simulations and experimental measurements, we develop a CG model that incorporates residue-level desolvation terms directly into the energy function and apply it to investigate LLPS of intrinsically disordered proteins. Incorporating explicit desolvation reshapes the phase diagram, yielding improved predictions of dense-phase packing density. Strikingly, we uncover a linear relationship between the temperature gap (simulation temperature relative to the critical point) and the extent of conformational compaction accompanying the dilute-to-dense phase transition, a result further supported by theoretical analysis. We also find that desolvation barriers accelerate early-stage coarsening dynamics while slowing chain mobility within mature condensates, whereas solvent-separated contact interactions exert the opposite effects. Together, this framework enables efficient and explicit treatment of desolvation in CG simulations and reveals how desolvation energetics shape both the thermodynamic landscape and kinetic property of biomolecular LLPS.

The AI revolution in natural-language and image processing has catalyzed new concepts and research paradigms in protein science. Researchers in protein science have first developed protein language models via self-supervised pre-training on large-scale protein sequences and have subsequently integrated multiple data sources or modalities into a single model to leverage vast data and the complementary information across modalities. These sources include, but are not limited to, protein sequences, structures, dynamics derived from molecular-dynamics simulations, functional annotations from resources such as the Gene Ontology, experimental data, and even natural-language descriptions. Such integration boosts performance, enables cross-talking between distinct data modalities, and endows models with new capabilities. In this review, we discuss recent studies that fuse diverse modalities, including foundation models designed for downstream tasks and representative works that integrate heterogeneous data for specific applications.

Viruses rely on the precise packaging of their genomes within a capsid to execute essential life-cycle events, yet the principles governing genome structural organization in this confined environment remain elusive. Here, we reveal that hepatitis B virus (HBV) pregenomic RNA (pgRNA) exploits liquid-liquid phase separation (LLPS) inside the capsid to sculpt its architecture. Multiscale molecular dynamics (MD) simulations, supplemented by biochemical assays, show that pgRNA coalesces into a hollow, shell-like condensate along the inner capsid surface, with coexisting low- and high-density regions. Electrostatic interactions between pgRNA and the disordered C-terminal domain of capsid protein primarily govern condensate formation. LLPS drives the establishment of microphases composed of nematically aligned RNA hairpin arrays interspersed by domains rich in flexible single-stranded RNA linkers, achieving an optimal balance between structural order and dynamic flexibility. Intriguingly, although the ensemble-averaged pgRNA density exhibits icosahedral symmetry, individual simulation snapshots display pronounced heterogeneity, indicating symmetry breaking at the single-particle level. In addition, LLPS-induced hollow-shell architecture of pgRNA genome promotes long-range RNA base-pairing and enhances polymerase mobility, which may facilitate the functional dynamics of polymerase during reverse transcription. Our findings uncover a capsid-confined LLPS mechanism that orchestrates viral genome structure and dynamics, offering new targets for antiviral intervention.

The interplay between protein folding and native dynamics remains a central question in biophysics. Analyzing an extensive set of AlphaFold-predicted structures, we uncover a robust relationship between folding topology (contact order) and native dynamics (fluctuation entropy), showing that long-range contacts that slow folding also restrict conformational flexibility across protein sizes and taxonomic groups. Scaling analysis reveals that this relationship, together with its chain-length dependence, is consistent with power-law-like trends, reflecting common organizing constraints of protein architecture. Across species, increasing organismal complexity is associated with proteome-wide shifts toward lower contact order and higher fluctuation entropy. Together, evidence from folding, stability, and functional dynamics converges on unifying constraints, revealing an intrinsic physical organizing principle captured by AI models.