Journal of Chemical Theory and Computation

Coarse-grained protein force fields enable simulations of biomolecular systems at length and time scales that are difficult to access with atomistic models, but achieving transferability across folded, intrinsically disordered, and multidomain proteins remains challenging. A central difficulty is that one-bead-per-residue models must represent chemically specific residue interactions while also absorbing solvent-mediated and many-body effects into a simplified energy function. Here, we present MOFF2, a transferable coarse-grained protein force field that combines residue-pair-specific interactions with a density-dependent many-body potential. MOFF2 is optimized using a two-stage strategy: bottom-up parameter learning from heterogeneous reference ensembles followed by refinement against experimental conformational observables. The resulting model provides balanced performance across ordered proteins, intrinsically disordered proteins, and multidomain proteins, and predicts condensate saturation-concentration trends for A1-LCD variant systems. Analysis of the learned parameters reveals chemically interpretable interaction patterns and density-dependent effects that explain the models improved transferability. These results demonstrate that combining a generalized coarse-grained energy function with data-driven optimization can produce a practical and interpretable force field for protein conformational and condensate simulations.

Wild-type T4 lysozyme (T4L) is used as a benchmark to evaluate conformational sampling across generative AI, AI-accelerated molecular simulation (AMS), and physics-based enhanced molecular dynamics (EMD). A four-state model: exposed/open, exposed/closed, buried/open, and buried/closed; is defined using physically meaningful collective variables. While generative AI methods (AF-cluster, MSA subsampling of AlphaFold2, ConforFold, AlphaFlow, ESMFlow, ConfRover, BioEmu) largely sample only the exposed/open state, AMS integrating generative ensembles with iterative molecular dynamics, recovering all states and reproducing equilibrium populations similar to EMD and experimental smFRET signatures.

All-atom molecular dynamics has become an indispensable tool in development of nanopore sensors of biological information. In a typical nanopore experiment, measurements of ionic current flowing through a nanopore report on the chemical structure of biomolecules that pass through the nanopore. Such experiments alone are often insufficient to relate the structure of the biomolecules to the ionic current modulations. The molecular dynamics method can establish such a relationship directly through a brute force simulation under applied electric field. Here, we examine the ability of molecular dynamics force fields to reproduce experimentally measured nanopore blockade currents produced by single-stranded DNA. Our simulations show that none of the "off the shelf" force fields (CHARMM36, AMBER Parmbsc1 and DES-AMBER) is capable of reproducing experimental data with the desired level of accuracy. To improve the accuracy, we examined and refined interactions between ions, protein nanopores and DNA, guided by experiments designed specifically for this purpose. Ultimately, the introduction of surgical corrections to non-bonded interactions within the CHARMM36 force field produced a favorable agreement between simulation and experiment. This refined parameterization, initially developed for nanopore sensing simulations, may have broader applications in computational studies of DNA-protein systems.

Designing peptide sequences that remain stable and selective across heterogeneous environments remains a central challenge in biomolecular modeling. Here we introduce an interpretable, physics-based Hamiltonian for environment-conditioned design of -helical peptide sequences. The model integrates helix propensities, pairwise interactions, electrostatics, anisotropic solvent exposure, and interfacial geometry into a unified energy function. To enable comparison across sequence lengths and environments, all contributions are rescaled and expressed as Z-scores relative to random sequence ensembles, yielding a normalized design landscape with balanced physical terms. This formulation defines a structured optimization problem that can be explored using exact, heuristic, and hybrid quantum- classical approaches without modification of the underlying model. The Hamiltonian recovers polar and apolar limits, discriminates experimentally characterized water-soluble and transmembrane -helical peptide sequences, and captures the preferential stabilization of membrane-active sequences at anionic interfaces over non-functional controls. It further enables multi-objective and selective design, generating candidate sequences with tunable environmental specificity.

The molecular mechanics-generalized Born surface area method (MMGBSA) is one of the most commonly used end state approaches used for the calculation of the binding free energy towards computational drug design and screening studies. It is customary to break up the free energy into van der Waals, electrostatic, polar solvation (GB), and nonpolar solvation (SA) terms and then either correlate these terms with experiment or assign physical meaning to each term. Here, we demonstrate that this assumption of independent fitting coefficients for decomposed energy terms could be invalid. Through analytic derivation and large-scale molecular dynamics simulations, we show that (i) the protein and ligand Coulomb interaction energy and the GB solvation correction are almost perfectly collinear (R2[≥]0.99) reflecting their designed role as vacuum electrostatics plus solvent screening, and (ii) the van der Waals interaction and SA term likewise exhibit strong correlation, as both depend primarily on buried surface area. Interaction entropy and C2 entropy corrections are also found to be strongly dependent on underlying electrostatic fluctuations, further reinforcing redundancy. These findings hold both at the level of instantaneous trajectory fluctuations and when averaged across a diverse set of 139 protein-protein complexes and persist in both single-trajectory and three trajectory MMGBSA protocols. Our results caution against using decomposed MMGBSA terms as independent predictors in regression models and suggest instead combining correlated terms into effective polar, nonpolar, and entropic contributions. Our study provides a systematic diagnosis of collinearity in MMGBSA and highlights pathways toward more interpretable and statistically robust predictive modeling.

Coarse-grained structure-based models (CG-SBMs; or G[o] models) are simplified potential energy functions of biomolecules or biomolecular complexes that encode their structure. Molecular dynamics simulations of such SBMs have been successfully used to study long time-scale dynamics such as protein and RNA folding, and large conformational transitions of biomolecular complexes. SBMs have several advantages: (1) Their MD simulations are computationally inexpensive, making extensive sampling easily accessible to many researchers. (2) They are easy to modify and can be adapted for the specific biomolecular problem that needs to be investigated. However, the force-fields of SBMs are not usually included in commonly used biomolecular simulation packages resulting in a barrier to their use. Here, we present SuBMIT (Structure Based Models Input Toolkit; https://github.com/sglabncbs/submit), a toolkit for generating coarse-grained SBM input files for performing MD simulations with GROMACS and OpenMM/OpenSMOG. Simulations whose input files can be generated using the different flavors of CG-SBMs present in SuBMIT include the folding and conformational ensembles of proteins with intrinsically disordered regions, 3D-domain-swapping in proteins and the dynamics of RNA-protein assemblies (e.g., simple RNA viruses).

All-atom simulations of RNA using molecular dynamics have the promise of modeling conformational preferences, folding thermodynamics, conformational change kinetics, and binding affinities of small molecule therapeutics. These simulations rely on a force field, a set of equations and parameters that model the potential energy as a function of conformation using classical mechanics. One popular force field for RNA is Amber OL3, with the most recent iteration derived in 1999 and with subsequent updates to backbone dihedral parameters. The Amber force field, while frequently used, is known to have limitations; for example, it does not properly stabilize native structures against alternative structures. Here, we provide a new approach to fitting the non-bonded parameters for the force field, specifically atom-centered point charges for electrostatics and the Lennard-Jones parameters. The parameters are fit to quantum mechanics (QM) interaction energies calculated with symmetry-adapted perturbation theory (SAPT), including embedded point charges to represent the electrostatic field from solvent and adjacent nucleotides. In this pilot study with a limited set of fitting data, we use the Amber ff99 equations and atom types unchanged. With the revised parameters, we observe improvement in the stability of native structures relative to alternative structures. Native tetraloop conformations, which unfold with the Amber OL3 force field, are stable on the microsecond timescale with our new force field parameters. We also see improvement in the conformational preferences of tetramers. Crucially, A-form helices are still well-modeled, but we observe additional flexibility in an internal loop that is not consistent with NMR data. Overall, we provide evidence that this new approach to fitting RNA force field parameters to SAPT interaction energies with native-structure context represented as embedded point charges is promising. It offers a flexible solution for revising the equations in future work or for extension to other molecules that interact with RNA, such as proteins and small molecules. We call this new set of force field parameters Amber RNA.ROC26.

Excipient effects on protein stability are critical for biological formulations, yet their selection remains largely empirical. Here, we use molecular dynamics simulations to define unifying metrics of protein-excipient interactions at atomistic resolution. Enhanced sampling simulations of fast-folding miniproteins, including Trpzip, WAAAH-helix (an alanine-rich -helix), and Trp-Cage, were performed to capture folding transitions across diverse excipient conditions. We identified a general stabilization mechanism based on shape complementarity between protein networks and surrounding solvent networks. Stabilizing excipients were found to form solvent structures that preferentially complement each protein, as well as residues central to known folding pathways. This framework enables a unifying approach to mechanism-based excipient selection across diverse protein and solvent chemistries. More broadly, by treating protein and solvent as dynamically coupled partners, it provides a transferable strategy for understanding solvent-mediated effects in complex molecular systems.

Characterizing conformational transitions between distinct structural states is essential for understanding protein function but remains challenging due to the timescale limitations of atomistic molecular dynamics. While coarse-grained models like Martini accelerate sampling, classical elastic-network or G[o]-like restraints often trap proteins in a single energy basin, precluding the study of transition pathways between distinct functional states. Here, we present CTGoMartini, a comprehensive Python package designed to simulate protein conformational transitions using G[o]-Martini models in explicit membranes. CTGoMartini addresses key methodological limitations of existing approaches by redefining native contacts as a dedicated interaction type, thereby eliminating spurious protein aggregation artifacts in multi-copy simulations. The package implements both switching and multiple-basin approaches (Exponential and Hamiltonian mixing) to sample transitions between experimentally defined states. Furthermore, it integrates Hamiltonian replica exchange molecular dynamics (HREMD) with PyMBAR analysis, enabling efficient optimization of mixing parameters that govern barrier heights and relative state stabilities. We demonstrate the power of CTGoMartini through two biologically significant membrane protein systems: (1) capturing the inward-open to outward-open transition of the lipid transporter SPNS2, revealing the molecular mechanism of S1P translocation; and (2) elucidating how membrane surface tension and anionic lipids (POPA, PIP2) modulate the conformational equilibrium of the mechanosensitive ion channel TREK1. By streamlining model construction, simulation, and analysis, CTGoMartini offers an easy-to-use platform that connects static structural snapshots with their underlying dynamic functional mechanisms. TOC Graphic O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=118 SRC="FIGDIR/small/721921v1_ufig1.gif" ALT="Figure 1"> View larger version (26K): org.highwire.dtl.DTLVardef@75eb26org.highwire.dtl.DTLVardef@1a12accorg.highwire.dtl.DTLVardef@e927org.highwire.dtl.DTLVardef@1cb0dcd_HPS_FORMAT_FIGEXP M_FIG C_FIG

Membrane association is governed by the thermodynamics of amino acid partitioning between water and the lipid bilayer. Here, we quantified amino acid side-chain insertion energetics in a 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) bilayer using unbiased molecular dynamics simulations. Equilibrium depth distributions of 28 analogs, including multiple protonation states, were converted into potentials of mean force (PMFs) by Boltzmann inversion. The resulting PMFs reproduced the main features of bilayer partitioning. Hydrophobic analogs favored the bilayer core, aromatic analogs were stabilized in interfacial regions, and polar or charged analogs remained unfavorable in the hydrophobic interior. A diglycine analog representing the peptide backbone behaved similarly to uncharged polar residues. Depth-dependent pKa profiles and orientational analyses further showed how protonation equilibria and aromatic-ring alignment influence insertion energetics. Agreement with experimental hydrophobicity scales supports the robustness of the approach. These results provide an efficient and internally consistent framework for characterizing bilayer insertion energetics and establish a reference for future studies in more complex lipid environments. O_FIG O_LINKSMALLFIG WIDTH=198 HEIGHT=200 SRC="FIGDIR/small/723583v1_ufig1.gif" ALT="Figure 1"> View larger version (79K): org.highwire.dtl.DTLVardef@127b12org.highwire.dtl.DTLVardef@14de924org.highwire.dtl.DTLVardef@53b27org.highwire.dtl.DTLVardef@16e8ee4_HPS_FORMAT_FIGEXP M_FIG C_FIG SIGNIFICANCEMembrane-associated proteins represent a large fraction of the proteome and include many major drug targets, yet quantitative understanding of their interactions with lipid bilayers remains limited. Here, we present an unbiased molecular dynamics framework for systematically determining amino acid side-chain insertion free energies in a model bilayer. By deriving potentials of mean force directly from equilibrium depth distributions, this approach enables internally consistent comparisons across residue classes and protonation states without biasing restraints. The resulting free-energy profiles reproduce established hydrophobicity trends and show how protonation equilibria and aromatic-ring orientation modulate bilayer partitioning. This scalable strategy provides a quantitative reference for residue-level membrane thermodynamics and establishes a foundation for extending insertion energetics to more diverse lipid compositions and more complex membrane-associated systems.

The SMN2 exon 7 5' splice-site/U1 snRNA duplex contains an A$_{-1}$ bulge that weakens splice-site recognition and represents a therapeutically relevant RNA connectivity defect, yet its conformational landscape and coupling to solvation remain poorly understood. Here, we performed enhanced-sampling Hamiltonian replica-exchange molecular dynamics simulations of the SMN2 splice-site duplex using four explicit-solvent models (OPC, TIP4P-Ew, TIP3P, and SPC/E) and characterized the sampled ensemble using linear and machine-learned latent representations. Across representations, the A$_{-1}$ defect consistently populated three metastable conformational states distinguished by local duplex geometry, base stacking, hydrogen-bonding patterns, and solvent exposure. The relative populations of these states, together with first-shell hydration and Na$^+$ distributions around the defect, varied substantially across water models, demonstrating that hydration and ion organization actively shape the equilibrium between locally accommodated and solvent-exposed conformations of the SMN2 splice-site bulge. Our results shed light on the conformational components of this therapeutic RNA target and highlight the impact of solvation model as an important consideration for molecular simulations of RNA splice-site recognition and small-molecule repair.

Accurate calculation of energy changes upon mutation is a key requirement for the effective use of computational methods in protein design. In this study, we applied free energy perturbation (FEP) calculations to predict the effects of mutations on the binding free energy between the immunoglobulin subtype G (IgG) antibody fragment-crystallizable (Fc) region and the neonatal Fc receptor (FcRn), an interaction that is primarily responsible for antibody half-life. We assembled an extensive experimental dataset of Fc-FcRn binding affinities for wild-type (wt) and mutant complexes, including values from literature and from newly measured results. Starting from a crystal structure of the M252Y/S254T/T256E ("YTE") Fc variant bound to FcRn, we prepared all-atom models of human IgG1-subtype wt and YTE variant Fc-FcRn complexes, adding explicit hydrogens and assigning protonation states for key ionizable residues. Initial results using standard FEP protocols to compute relative binding free energies were promising but exhibited multiple outliers. By accounting for coupling effects for FEP mutations near key histidine residues, we improved the results for several outliers, suggesting such coupling as an important approach for pH-sensitive systems. Further, upon determining new crystal structures of four Fc variants at multiple pH values, we observed subtle conformational changes in unbound Fc; by accounting for these conformational changes in FEP calculations, we additionally improved agreement with experiment. The detailed structural and energetic analyses of the Fc-FcRn system we present here thus provide an accurate energy-calculation framework to enable rational in silico design of novel Fc variants. SignificanceThe ability to determine changes in binding affinity upon mutation is critical to structure-based protein design. In this study, we demonstrate a successful computational approach using free energy perturbation (FEP) calculations on the antibody Fc-FcRn complex, a medically relevant system with implications for both therapeutic and prophylactic antibody use. Our successful calculation of accurate binding energies across a wide range of cases speaks to the power of the FEP methodology in navigating the free energy landscapes of dynamic molecular complexes. Furthermore, we show that accurate Fc-FcRn affinity calculations required careful consideration of conformational flexibility between bound and unbound states, contributing to our functional understanding of a system that will be important for future rational antibody-design efforts. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=75 SRC="FIGDIR/small/721095v1_ufig1.gif" ALT="Figure 1"> View larger version (26K): org.highwire.dtl.DTLVardef@f754d8org.highwire.dtl.DTLVardef@1e366beorg.highwire.dtl.DTLVardef@6e67caorg.highwire.dtl.DTLVardef@602a14_HPS_FORMAT_FIGEXP M_FIG C_FIG

The integration of coarse-grained (CG) approaches into docking workflows offers a powerful strategy for modelling large biomolecular assemblies with reduced computational costs. We present here the implementation of the MARTINI2 coarse-grained force field into the HADDOCK3 integrative modelling platform. This development enables the use of the CG representations and parameters within HADDOCK3 for efficient sampling and scoring of large protein-protein complexes. The implementation takes advantage of the modular and flexible architecture of HADDOCK3, allowing a seamless combination of MARTINI2 representation with the various modules. Conversion from and to all-atom models is integrated into the coarse-grained modelling workflow. The performance of the protocol is first assessed on protein-protein and protein-DNA benchmarks and then illustrated on a few representative large-scale systems, demonstrating a significant reduction in computational costs while maintaining biologically relevant accuracy.

Binding kinetics are crucial for antibody function, shaping pharmacokinetics and in vivo efficacy beyond what equilibrium affinity captures. We present "Metadynamics-Anchored Hybrid Learning for Engineering off-Rates (MAHLER)", a fully open-source machine learning/physics hybrid method that predicts relative antibody-antigen residence times at scale. Incorporating inverse-folding models into molecular dynamics simulations, MAHLER shows first-in-class screening-grade accuracy in calculating relative antibody-antigen dissociation kinetics across a family of point mutants. After initial antigen-specific setup, each prediction takes only 4 minutes on a single NVIDIA A100 GPU, compared to days even with already enhanced molecular dynamics simulations. This provides practical kinetics-aware complement to current computational design approaches that focus primarily on binding affinity for antibody-antigen complexes.

Deep learning approaches have revolutionized protein structure prediction. These tools are trained using experimental data and recapitulate reported conformations, but there is great interest in predicting conformations that may be functionally relevant although experimentally underrepresented. Since many modern structure prediction tools use generative artificial intelligence diffusion models, we reframe the search for alternative molecular conformations as that of sampling from a diffusion distribution conditioned using any arbitrary Bayesian likelihood. We implement a twisted diffusion sampler in Boltz-2 to sample this conditioned distribution and demonstrate the utility of this approach, which does not require any additional training of the neural network, by implementing a diffusion analog of steered molecular dynamics simulations applied to mechanical systems. We can reproduce predicted stretched states of fragments of DNA, the muscle protein titin, and the inner-ear protocadherin-15 protein, as well as open states of the MscL ion channel consistent with experimental results. We expect that steered structure predictions will help sample underrepresented and non-equilibrium conformations for many macromolecular systems.

The ability of small molecules to cross the blood-brain barrier (BBB) remains a major bottleneck in neurotherapeutic development. While experimental assays and machine learning approaches provide approximate permeability estimates, they lack atomistic insight into the underlying transport mechanisms. Here, we employ all-atom molecular dynamics simulations of a compositionally realistic BBB lipid bilayer to characterize the passive permeation of two bioactive propolis-derived compounds, Caffeic Acid Phenethyl Ester (CAPE) and Artepillin-C (ARC). Using steered molecular dynamics and umbrella sampling, we computed free energy profiles, diffusion coefficients, and permeability metrics across the membrane. CAPE encounters a modest barrier at the lipid headgroup region but minimal resistance within the hydrophobic core, resulting in a low free energy barrier ([~]2-3 kcal/mol) and favorable permeability (logP_eff {approx} 0.28). In contrast, ARC exhibits a substantial energetic barrier within the membrane core, leading to high resistivity and strongly unfavorable permeability (logP_eff {approx} -10.91). The heterogeneous lipid model reproduces experimentally consistent membrane properties and reveals how lipid composition modulates transport energetics. These findings provide mechanistic insight into BBB permeability and demonstrate the utility of atomistic simulations for guiding the design of neuroactive therapeutics.

Hydrogen-bond networks are central to protein function, but most network analyses rely on static representations that neglect how interactions evolve in time. Here, we introduce a framework that combines instantaneous and temporal graph analysis of hydrogen-bond networks derived from molecular dynamics trajectories to quantify ligand-directed hydrogen-bond connectivity. We apply the method to E. coli dihydrofolate reductase (DHFR) and its L28R mutant, computing shortest hydrogen-bond paths from all residues to the substrate dihydrofolate (DHF). The instantaneous analysis reveals that DHF-directed connectivity is organized through a sparse set of preferred routes, with D27 and T113 acting as prominent hubs in the wild-type enzyme. Temporal analysis highlights residues that preferentially support time-ordered DHF-directed connectivity. Comparison with L28R shows that the mutation preserves the main substrate-contacting architecture and the overall communication scaffold but redistributes pathway usage, persistence, and temporal convergence. The network-derived hotspots partially overlap with independent coevolution signals, most strongly in the K109-I115 region, while overlap with cryptic-site predictors is more limited. This pattern indicates that the hydrogen-bond network captures evolutionarily supported communication regions in DHFR that are not fully recovered by static structural approaches. The framework is broadly applicable to ligand-binding proteins and provides a route to identify persistent, delayed, and mutation-sensitive signaling pathways directly from time-ordered simulation data.

The conformational ensembles of intrinsically disordered proteins (IDPs) respond sensitively to ionic strength, with the direction and magnitude of response varying widely across sequence classes from polyelectrolyte contraction to polyampholyte swelling. Recent sequence-conditioned models provide rapid access to full ensembles or ensemble-averaged properties at specified solution conditions, but do not directly identify which low-dimensional polymer physics descriptors organize salt response behavior. Here, we construct a 511-sequence library spanning controlled{kappa} -variants, NCPR series, IDRome-stratified natural IDRs, and low-FCR IDRome sequences, and perform 2,555 CALVADOS-2 simulations across five monovalent salt concentrations (50 mM to 500 mM). For each sequence, we extract the salt-response slope, dRg/d[salt], and assign one of four regimes: polyelectrolyte contraction, polyampholyte swelling, non-monotonic response, or salt-insensitive behavior. Using eight theory-motivated sequence descriptors, we find that sequence charge decoration weighted by chain length, SCD xN, is the dominant coordinate organizing salt response, accounting for [~]40% of total SHAP attribution and exceeding the next feature by more than twice. Ridge regression explains substantial in-distribution variance (R2 = 0.83 under random cross-validation), whereas gradient-boosted trees improve in-distribution performance (R2 = 0.97) and retain predictive power under the more stringent leave-one-subset-out validation test (R2 = 0.60), indicating that salt response contains transferable but nonlinear sequence-encoded structure. Regime classification robustly recovers the direction of salt response, with no polyelectrolyte-polyampholyte confusion, whereas non-monotonic and salt-insensitive sequences remain harder to distinguish from static sequence features alone. Together, these results establish SCD xN as a compact, interpretable organizing coordinate for CALVADOS-2-derived IDP salt response and provide a polymer-physics, feature-level complement to ensemble-level generative models.

We present Pathway Representation via Intrinsic Structural Medoids (PRISM), a state-aware framework for clustering pathways from molecular dynamics simulations of biomolecular transitions. In PRISM, each pathway is mapped to a small set of structural medoids obtained via a deterministic k-means clustering scheme. Pairwise pathway dissimilarities are computed using a weighted average Hausdorff distance between these representative sets, effectively capturing mean nearest-neighbor structural deviations while reducing sensitivity to outliers. Hierarchical agglomerative clustering of the resulting dissimilarity matrix defines pathway families. We evaluate PRISM across three biomolecular transitions of increasing complexity: alanine dipeptide C7eq [->] C7ax isomerization, adenylate kinase opening, and HIF-2 PAS-B ligand unbinding. PRISM consistently yields robust cluster assignments, with medoids faithfully representing distinct conformational states. By combining a state-based description with robust geometric dissimilarities, PRISM provides a scalable framework for organizing complex transition pathways.

Industrial biocatalysis increasingly requires enzymes capable of operating under extreme physicochemical conditions, yet most natural sequence data reflect adaptation to mild environments, leading conventional predictive models to suffer from regression-to-the-mean effects in extremophilic regimes. Here we present EnzOracle, a classification-guided mixture-of-experts framework that enables distribution-aware prediction of enzyme melting temperature (Tm), optimal catalytic temperature (Topt), and optimal pH (pHopt) directly from sequence. EnzOracle demonstrated robust predictive accuracy across diverse benchmarks, achieving RMSE of 5.245{degrees}C for Tm, 11.458{degrees}C for Topt, and 0.781 for pHopt. Beyond predictive accuracy, we introduce a trait-resolved molecular simulation strategy to evaluate whether EnzOracle-derived attribution patterns correspond to independent physical mechanisms. Across representative systems, attention hotspots mapped onto rigidity-conferring interaction networks for Tm, dynamically preorganized active-site ensembles for Topt, and pH-dependent electrostatic and hydration networks for pHopt. These orthogonal validations indicate that EnzOracle captures transferable biophysical principles of enzyme environmental adaptation rather than merely exploiting dataset-specific correlations, positioning sequence-based learning as a mechanism-aware framework for discovering stability and activity determinants across diverse catalytic landscapes.