Guy W. Dayhoff

Genetic variation in cervid PRNP influences chronic wasting disease (CWD) susceptibility and progression, but it remains unclear whether disease-associated substitutions produce localized differences in sequence-derived properties relevant to prion protein (PrPC) condensation. We compared a white-tailed deer (WTD) reference sequence with 10 modeled variants on the same mature PrP scaffold (residues 24–233). For each sequence, we calculated intrinsic disorder, FuzDrop per-residue droplet-promoting propensity (pDP) and binding-mode entropy (Sbind), predicted residue-level pKa values, and a discrete-charge summary at pH 7.0. The overall profiles preserve a long N-terminal droplet-promoting region, whereas variant-associated differences are concentrated in selected intervals. In the FuzDrop outputs, residues 126–140 show allele-dependent changes in pDP near the droplet-promoting threshold, and residues 220–230 show variant-specific changes in Sbind, including lower Sbind for S225F/Y and higher Sbind for elk-like Q226E. These differences occur without large changes in overall intrinsic disorder. KaML pKa predictions identify near-neutral titratable sites that provide electrostatic context for these local predictor differences. Independent atomistic simulations of monomeric PrP further detected variant-sensitive backbone behavior and intramolecular contacts involving these intervals. Together, the analysis identifies residues 126–140 and 220–230 as computationally defined candidate intervals that overlap regions implicated independently in prion conversion or CWD fibril structure, motivating direct tests of cervid PrP condensation and maturation.

Ionization states are crucial to protein function, yet predicting their equilibrium constants remains challenging. Here we present KaML-ESMs, sequence-based neural networks built on the ESM protein language models and trained on the GAINES-augmented PKAD-3r dataset. KaML-ESM2/ESMC outperform alternative approaches across four benchmarks, achieving root-mean-square errors of about 0.5 units for Asp/Glu/His/Cys/Tyr/Lys residues in native proteins. Performance degrades on 89 buried engineered mutations lacking evolutionary support, revealing a limitation that supervised training may help overcome. These results suggest that protein sequence, through its evolutionary context, encodes not only structure and function but also, indirectly, electrostatic characteristics. Applying KaML-ESM2 to the entire human proteome, we show that predicted p K _a values can identify functional sites and infer catalytic mechanisms. We offer KaML, a sequence-based p K _a prediction platform supporting biological exploration, protein engineering, and biomolecular simulation. With further development, GAINES could provide a general framework for addressing data scarcity in machine learning.

We present a machine-augmented molecular dynamics (MAMD) framework in which particle velocities are updated using predictions from stacked long short-term memory networks trained on historical velocity and coordinate data. MAMD propagates coordinate trajectories in time without access to forces or energies during training or inference. Applied to isolated harmonic diatomics, MAMD conserves total energy, preserves molecular structure, and reproduces velocity autocorrelation functions. Small integration errors can accumulate over long trajectories, but we show that molecular dynamics stability can be recovered through periodic, but infrequent injections of velocity updates computed from Hamiltonian forces (frequency ≤0.01). We also find that the optimal history length for each diatom closely matches the first inflection point of its velocity autocorrelation function, suggesting a link between model architecture and statistical mechanics. These results establish MAMD as a proof-of-concept integration strategy for the isolated harmonic diatomic systems studied here, in which finite velocity memory supports short-horizon predictions and sparse Hamiltonian check-ins provide long-horizon stability. In particular, we show that force-free, velocity-based machine learning updates can be embedded directly into conventional molecular dynamics algorithms while retaining essential physical invariants, providing a physically interpretable basis for hybrid MD/ML integration. Systematic extensions to more complex systems and quantitative performance comparisons in regimes with expensive force evaluations (e.g., ab initio molecular dynamics) are left for future work.

We present a machine-augmented molecular dynamics (MAMD) framework in which particle velocities are updated using predictions from stacked long short-term memory (LSTM) networks trained on historical velocity and coordinate data. MAMD propagates coordinate trajectories in time without access to forces or energies during training or inference. Applied to isolated harmonic diatomics, MAMD conserves total energy, preserves molecular structure, and reproduces velocity autocorrelation functions. Small integration errors can accumulate over long trajectories, but we show that molecular dynamics stability can be recovered through periodic, but infrequent injections of velocity updates computed from Hamiltonian forces (frequency ≤ 0.01). We also find that the optimal history length for each diatom closely matches the first inflection point of its velocity autocorrelation function, suggesting a link between model architecture and statistical mechanics. These results demonstrate the feasibility of MAMD as a proof-of-concept integration strategy, in which finite velocity memory supports short-horizon predictions and sparse Hamiltonian check-ins provide long-horizon stability. Specifically, we show that force-free, velocity-based machine learning updates can be embedded directly into conventional molecular dynamics algorithms while retaining essential physical invariants, providing a physically interpretable basis for hybrid MD/ML integration.

Creating a ligandable atlas for the proteome would transform our understanding of protein functions and accelerate therapeutic discovery; however, proteomic approaches are constrained by insufficient proteome coverage and data heterogeneity, while existing machine learning (ML) models have limited oiwer due to structural dependencies and heterogeneous experimental labels. Here we developed AiPP, a multimodal AI platform that predicts and characterizes ligand interaction sites directly from protein sequence. AiPP is powered by the evolutionary-scale protein large language models (LLMs) and leverages two harmonized ML training sets derived from the new databases comprising cysteine ligandability from activity-based protein profiling (ABPP) studies and reversible binding evidenced from co-crystal structures. We developed a LLM representation based clustering framework to interrogate, reconcile, and augment experimental labels in both databases. Two complementary protocols were implemented to iteratively expand the training data while improving model performance. Although trained exclusively on ABPP data, AiPP recovers 80% (Top-1) of cysteine liganding events from co-crystal structures, with 84% AUPRC and 89% AUROC. AiPP recapitulates consistently and heterogeneously liganded cysteines across cancer cell lines and reliably identifies dynamic, ligandable pockets in “undruggable” transcription factors. Remarkably, AiPP accurately predicts active-site and allosteric cysteines in protein tyrosine phosphatases that were undetected by ABPP. Finally, we applied AiPP to the entire human proteome, identifying ligandable sites in proteins that were undetected or unliganded by ABPP, including an allosteric site in MC3R, which is a therapeutic target for treatment of eating disorder and obesity. This proteomewide covalent ligandability atlas (version 1.0) is anticipated to guide future development of chemical probes and pharmaceutical modulators, particularly for understudied proteins and currently undruggable targets. The LLM-based approach to interrogate large-scale heterogeneous data is broadly applicable to protein research and development of proteomics-derived ML models for diverse applications.

Ionization states play crucial roles in protein functions. Yet despite decades of research, predicting protein p K _a values remains a formidable challenge. Here, we present KaML-ESMs, neural networks built on the ESM protein language models and trained on a synthetically augmented experimental dataset. The KaML-ESMs significantly outperform structure-based approaches for protein p K _a prediction, achieving accuracies approaching experimental resolution across six ionizable amino acids in native proteins. On the most challenging test set of engineered OBTRUDEs (iOnizable suBsTitutions foR bUrieD rEsidues), KaML-ESM2 emerges as the top model, achieving RMSE of 1.36. We applied KaML-ESM2 to the entire human proteome and demonstrated that the predicted p K _a ’s can identify functional sites and infer catalytic mechanisms. Our work challenges the structure-based paradigm and suggests a view that protein electrostatic properties are encoded in sequence, potentially co-optimized with structure and function through evolution. We offer KaML, a sequence-based end-to-end platform to support applications ranging from biological exploration to drug design, protein engineering, and biomolecular simulations. GAINES, a latent space sampling approach introduced in this work, provides a general framework for addressing data scarcity bottle-necks in machine learning applications.

Many proteins lack stable 3D structures. These intrinsically disordered proteins (IDPs) or hybrid proteins containing ordered domains with intrinsically disordered protein regions (IDPRs) often carry out regulatory functions related to molecular recognition and signal transduction. IDPs/IDPRs constitute a substantial portion of the human proteome and are termed “the unfoldome”. Herein, we probe the human breast cancer unfoldome and investigate relations between IDPs and key disease genes and pathways. We utilized bottom-up proteomics, MudPIT (Multidimensional Protein Identification Technology), to profile differentially expressed IDPs in human normal (MCF-10A) and breast cancer (BT-549) cell lines. Overall, we identified 2271 protein groups in the unfoldome of normal and cancer proteomes, with 148 IDPs found to be significantly differentially expressed in cancer cells. Further analysis produced annotations of 140 IDPs, which were then classified to GO (Gene Ontology) categories and pathways. In total, 65% (91 of 140) IDPs were related to various diseases, and 20% (28 of 140) mapped to cancer terms. A substantial portion of the differentially expressed IDPs contained disordered regions, confirmed by in silico characterization. Overall, our analyses suggest high levels of interactivity in the human cancer unfoldome and a prevalence of moderately and highly disordered proteins in the network.