Abstract Regression and machine-learning models for cyclodextrin inclusion constants are often reported with cross-validated Q2 near 0.9, yet the reproducibility of the compiled data on which they are trained has not been quantified. This work quantifies it at three levels, each defined by the uncertainty it captures. The within-experiment fitting precision of a single calorimetric study, 0.008 log units, is a lower bound on repeatability; head-to-head comparisons between laboratories and methods give 0.115 log units (95% CI: 0.079–0.210); and 128 replicated host–guest pairs in a verified extraction of the standard thermodynamic compilation give 0.157 log units (95% CI: 0.115–0.220, from a pair-level bootstrap), with a heavy tail in which 7.3% of pairwise comparisons, and an estimated 3–5% of single determinations, deviate by more than 1 log unit. Propagated through the replication structure of a 44-pair modeling set under three error models, this noise implies an out-of-sample R2 ceiling of about 0.70–0.78, and response noise accounts for roughly two-thirds to nine-tenths of the leave-one-out error of the adopted model, an error statistically compatible with a model limited by the data alone. Against this bound, a size descriptor derived from measured limiting diffusion coefficients through the Edward-corrected Stokes–Einstein relation was compared with McGowan volume, each paired with log P. The two are statistically indistinguishable (95% CI of the difference in mean squared error, −0.024 to +0.042) under leave-one-out and leave-one-guest-out validation, and both lose accuracy when whole chemical families are withheld. No packing optimum is detectable, and normalizing guest size by cavity volume degrades performance by aliasing the size term onto the host intercepts. The durable result is the bound, which should accompany any accuracy reported on compiled data.
Protein thermodynamic stability changes upon mutation profoundly affect protein structure and biological function. Recent advances in deep learning and protein language models have provided powerful tools for computational analysis of mutation-induced stability changes. However, most existing methods focus on quantitative regression of free-energy changes, which places strong dependence on precise experimentally measured ΔΔG values, although reliably determining the direction of stability change is itself an important objective in many practical applications. Here, we propose BlindDDG, a novel framework that reformulates protein mutation stability prediction from quantitative ΔΔG estimation to direction-oriented stability change classification. First, we introduce a blind-evaluation strategy that models mutation-induced changes by learning the representation difference between wild-type and mutant proteins, rather than relying solely on their individual intrinsic properties. This formulation focuses on the direction of stability changes and reduces the dependence on precise quantitative ΔΔG measurements. Second, a Siamese architecture with shared parameters is employed to encode the wild-type and mutant states and capture subtle mutation-associated differences. Third, a local structure-aware cross-attention mechanism is developed to integrate mutation-centered structural information with global protein language model representations, enabling effective multimodal feature interaction. Experimental results on the independent S669 benchmark demonstrate that BlindDDG achieves strong classification performance, obtaining an F1 score of 0.5449, an MCC of 0.3929, an accuracy of 0.7721, and a recall of 0.5417 for stabilizing mutation identification. BlindDDG outperforms all evaluated methods in F1 score, MCC, accuracy, and recall, highlighting its effectiveness in capturing mutation-induced stability changes, particularly for imbalanced stability classification.
In recent years, the use of document information to predict compound activity has gained attention. However, existing methods face challenges in applying these predictions to compounds not included in the training data because of the shortage and insufficient applicability of chemical knowledge as well as in interpreting the basis for predictions. Therefore, this study developed an approach to classifying the chemical roles of these compounds. By analyzing the descriptions of compounds registered in databases using Doc2Vec, we obtained the embeddings. Specifically, by setting the tag portion of Doc2Vec to the substructures of compounds, we acquired embeddings corresponding to the substructures and summed them to represent the compounds. We predicted the presence or absence of 10 chemical roles for the compounds; compared to predictions using chemical descriptors, our predictions had higher F1 scores for many chemical roles. Additionally, by applying Shapley additive explanations, we visualized the substructures likely to influence the prediction of the presence or absence of each role, confirming that their interpretation was consistent with the findings of the previous research. This paper suggests that the obtained embeddings, which combine the structural information and annotations of compounds, are useful descriptors for classifying the chemical roles of the compounds.
Rapid development of therapeutics against emerging viral pathogens increasingly depends on machine learning (ML), both to generate candidate binders and for structural validation. Nanobodies (Nbs), or single-domain antibodies, are promising antiviral therapeutics, but accurate in silico prediction of viral epitopes remains challenging. Here we benchmark three ML-based structure prediction models, Boltz-2, Chai-1 and AlphaFold3 (AF3), against a curated data set of viral antigen-nanobody complexes deposited after the AF3 training cutoff. AF3 outperformed the other models: success (interface DockQ > 0.23) rose from 33% with a single seed to 50% and 65% with 10 and 100 seeds-per-target, respectively, compared with 18% for Boltz-2 and 30% for Chai-1. Using an expanded data set, we evaluated confidence metrics and found that a strict ipTM cutoff of 0.8, or a combination of ipTM ≥ 0.6 with a CDR3 loop pLDDT ≥ 60, reliably excluded incorrect predictions while retaining most successful ones. We further establish best practices for two AF3-specific capabilities, the choice of antigen multimeric state and the inclusion of N-glycosylation, both of which can be decisive for epitope success depending on the target. Finally, we apply these guidelines to two design problems: ranking a library of VEEV-neutralizing nanobodies and mapping the epitope of the broadly neutralizing nanobody CC3 across homologous alphavirus antigens. These results provide a practical framework for ML-based viral nanobody epitope prediction and design.
Abstract Histatins are a family of human saliva peptides; the three main members consist of Histatin 1 (Hst1), Histatin 3 (Hst3), and Histatin 5 (Hst5). Despite their high primary-structure similarity, the peptides display varied functions and behaviors. Experimental evidence supports concentration-dependent self-association of Hst1, but the molecular mechanisms remain poorly understood. This property is of interest because it is not observed in Hst5. Here, we combine atomistic and coarse-grained molecular dynamics simulations with a decomposition-based trajectory analysis to investigate the structural origins of Hst1 self-association across multiple length and time scales. Atomistic simulations reveal dimers that remain associated over the simulated time scales, together with transient higher-order oligomers. Analysis of intermolecular residue contacts highlights aromatic residues as prominent contributors to peptide association, consistent with a multivalent interaction network. Coarse-grained simulations extend these observations to larger systems and longer time scales, demonstrating a progressive assembly into dynamic peptide clusters. By discretizing the resulting conformational landscape, we extracted representative structures to reconstruct the self-association pathway from dispersed monomers to higher-order assemblies. The simulations provide a multiscale description of Hst1 self-association and support a model in which weak, multivalent peptide–peptide contacts contribute to the formation of dynamic peptide-rich assemblies. Although the present work does not directly demonstrate liquid–liquid phase separation, it identifies molecular features associated with biomolecular assembly.
