structural bioinformatics

Protein dynamics are essential for understanding biological function, yet progress is hindered by the high computational cost of molecular dynamics simulations and the scarcity of dynamic structural data. This work provides a systematic review of artificial intelligence approaches in this domain, organized around three complementary directions: learning from structural ensembles and trajectories, learning from physical energy signals, and methods designed to accelerate molecular simulations. For the first time, it integrates AI techniques through the tripartite lens of structure, energy, and dynamics, comprehensively surveying recent advances in conformational generation, trajectory prediction, Boltzmann generators, and physics-informed adaptation. The review also catalogs representative methods, datasets, and evaluation metrics, while identifying key challenges—particularly regarding scalability and thermodynamic consistency—and outlining the current frontiers and future pathways for AI-driven protein dynamics modeling.

Current 3D molecular generation research widely relies on the GEOM-Drugs dataset for evaluation, yet its preprocessing suffers from severe chemical inaccuracies—including erroneous valence bond assignments, miscalculated bond orders, and inconsistent classical force-field parameterization misaligned with reference structures—compromising chemical validity of evaluation metrics. Method: We systematically diagnose and rectify these flaws by introducing a “chemical consistency–first” evaluation framework: (i) reconstructing valence rules per IUPAC standards; (ii) replacing empirical force fields with GFN2-xTB quantum-chemical geometry optimization and energy computation to establish chemically rigorous ground-truth benchmarks; and (iii) implementing a molecular topology validation and graph-rule modeling pipeline. Contribution/Results: Re-evaluating state-of-the-art generative models under this framework reveals substantial overestimation of prior performance metrics. We publicly release corrected data protocols and evaluation scripts to foster community-wide adoption of chemically sound 3D molecular generation assessment standards.

Experimental protein–ligand complex structures are scarce, hindering data-driven drug binding affinity prediction. Method: We propose a novel AI-based data augmentation paradigm: (i) generating synthetic complexes en masse using protein–ligand co-folding models (e.g., AlphaFold-Multimer or RoseTTAFold-All-Atom), and (ii) automatically filtering high-quality predictions via lightweight heuristic rules—based on per-residue pLDDT, interface residue confidence, and geometric plausibility—to substitute experimental structures for training machine learning scoring functions. Contribution/Results: This work is the first to systematically demonstrate that rigorously filtered AI-predicted structures can support high-accuracy affinity modeling. On standard benchmarks (e.g., PDBbind), models trained solely on filtered synthetic data achieve performance on par with—or even surpassing—that of baselines trained on experimental structures (ΔRMSE ≤ 0.2 kcal/mol), markedly reducing reliance on experimentally determined complexes.

Current structure-based scoring functions exhibit limited generalizability across diverse biomolecular systems. To address this, we propose UniScore—the first unified geometric graph learning framework for biomolecular scoring—integrating dual-scale graph modeling (atomic- and residue-level), structure-aware encoding, and hybrid biomolecular pretraining. UniScore enables joint modeling of binding affinity prediction, pose ranking, and virtual screening, while supporting zero-shot and few-shot transfer. Evaluated on 16 benchmarks, it consistently matches or surpasses 70 state-of-the-art methods: achieving over 60% reduction in affinity prediction error, 40% improvement in protein–protein interaction prediction, and >90% increase in correlation for antigen–antibody binding prediction. Critically, UniScore demonstrates markedly enhanced cross-system generalizability and multi-task compatibility, establishing a new foundation for unified, data-efficient biomolecular scoring.

This study investigates whether activity cliffs—pairs of structurally similar compounds exhibiting markedly different biological activities—are intrinsic molecular properties or artifacts induced by the geometric effects of molecular representations. To address this, the authors develop a six-step analytical pipeline to systematically evaluate the geometric characteristics, enrichment, activity gradients, persistent homology structures, and predictive performance of activity cliffs across multiple molecular embeddings (Morgan, MolFormer, MACCS, RDKit, ChemBERTa) and distance metrics. Validation is performed using matched molecular pairs and stereoisomers. The findings reveal that the definition of activity cliffs is highly representation-dependent: different embeddings capture distinct aspects of molecular recognition—Morgan with Tanimoto similarity shows optimal enrichment, MolFormer is sensitive to stereochemistry, MACCS and RDKit are most responsive to structural changes, while ChemBERTa underperforms due to embedding collapse. This work redefines activity cliffs as representation-dependent phenomena rather than absolute properties.

Current antibody affinity evaluation methods typically analyze antibody sequences or structures in isolation, lacking a unified benchmark that treats the antibody–antigen (Ab–Ag) complex as the functional unit and reflects true binding capability. To address this, we propose AbBiBench—the first function-oriented evaluation framework grounded in complex likelihood estimation, breaking from conventional single-antibody assessment paradigms. AbBiBench integrates masked language modeling, autoregressive generation, inverse folding, diffusion-based structure generation, and geometric graph neural networks, jointly scoring candidates across experimental affinity, structural integrity, and biophysical properties. We systematically evaluate 14 state-of-the-art models on a benchmark comprising 9 antigens and 156,000 antibody variants. Results show that structure-conditioned inverse folding models achieve top performance. In an H1N1 antibody design case study, AbBiBench demonstrates strong predictive validity: model-derived complex likelihood correlates significantly with experimental dissociation constants (K_D; Pearson *r* = 0.72).

Traditional reaction path visualization methods—e.g., energy–displacement plots—project high-dimensional structural evolution onto a one-dimensional trajectory, obscuring configurational details and impeding cross-algorithm comparison of optimization dynamics. To address this, we propose a two-dimensional projection framework based on permutation-corrected RMSD: the configuration plane is defined by RMSD distances from reactant and product structures; radial basis function interpolation and energy-based coloring yield a continuous, interpretable potential energy surface. This enables simultaneous trajectory visualization, convergence diagnostics, and precise saddle-point localization, facilitating intuitive comparative analysis across optimization algorithms. Validation on cycloaddition reactions shows that machine-learned interatomic potentials predict saddle points with minor geometric deviations, yet these lie within energy contour levels closely matching DFT reference results—demonstrating strong configurational–energetic consistency.

Extracting interpretable and robust descriptions of protein conformational dynamics from high-dimensional molecular dynamics trajectories remains challenging, as existing approaches often rely on a single conformational representation while overlooking its fundamental influence on outcomes. This work systematically compares multiple conformational representations and introduces a geometry-driven, rotation-aware backbone orientation feature. To facilitate efficient computation and analysis across diverse representations, the authors develop the ManiProt library. Validation on fast-folding proteins, large-scale domain motions, and protein–protein binding systems demonstrates that different representations are complementary, with no single one fully capturing the complete dynamic landscape. The study reveals that the choice of representation significantly affects inferred conformational organization, similarity metrics, and transition pathways, advocating for a representation-aware comparative framework to establish a new methodological foundation for analyzing protein dynamics.

Recent advances in de novo protein binder design have enabled increasing experimental validation, yet reported in silico metrics remain difficult to interpret or compare across studies due to non-standardized evaluation protocols. We introduce ProtDBench, a standardized and throughput-aware evaluation framework for protein binder design. ProtDBench defines unified benchmark tasks, evaluation protocols, and success criteria, enabling systematic analysis of how evaluation design influences observed performance. Using a large wet-lab annotated dataset, we analyze commonly used structure prediction models as evaluation verifiers, revealing substantial verifier-dependent bias and limited agreement under identical filtering protocols. We then benchmark representative open-source generative binder design methods across ten diverse protein targets under a fixed evaluation protocol. Beyond per-sequence success rates, ProtDBench incorporates throughput-aware metrics based on a fixed 24-hour budget, as well as cluster-level success criteria to account for structural diversity. Together, these results expose systematic differences induced by filtering rules, success definitions, and throughput-aware evaluation between computational efficiency, success rate, and structural diversity. Overall, ProtDBench provides a fair and reproducible evaluation pipeline that supports systematic and controlled comparison of protein binder design methods under realistic evaluation settings.

This study addresses the critical challenge of **generalization to unseen protein targets** for learnable protein–ligand scoring functions, revealing that standard benchmarks (e.g., PDBbind) severely overestimate real-world performance due to target overlap and data leakage. To rectify this, we propose a **rigorous unseen-target evaluation paradigm**, constructing target-level splits grounded in realistic scarcity of structural and affinity data. Methodologically, we integrate **large-scale self-supervised molecular pretraining** with **lightweight few-shot fine-tuning**, substantially enhancing cross-target extrapolation. Experiments show that state-of-the-art scoring functions suffer >40% average performance degradation on truly unseen targets—validating the misleading nature of conventional benchmarks. In contrast, our approach achieves significant generalization gains using only 1–5 target-specific samples for fine-tuning. This establishes a reliable, AI-driven scoring foundation for de novo drug discovery against novel therapeutic targets.