Protein flexibility is essential for enzyme function, as catalysis often requires coordinated conformational changes that position catalytic residues, exclude solvent, and stabilize transition states. Rather than acting as rigid scaffolds, enzymes sample ensembles of conformations, and ligand binding frequently shifts this equilibrium through induced fit or conformational selection. This intrinsic mobility can extend to bound ligands, particularly in spacious or partially solvent-exposed active sites, resulting in alternative ligand conformations or partial disorder in experimental density maps. A well-known example is HIV-1 protease, where both the enzyme flaps and bound inhibitors adopt multiple conformations that manifest as split electron density. Such ligand disorder provides direct structural evidence of the dynamic nature of protein-ligand complexes.

X-ray crystallography and Cryo-EM remain the principal techniques for determining three-dimensional structures of protein-ligand and protein-protein complexes and are foundational to structure-based drug design (SBDD). However, conventional refinement methods rely heavily on stereochemical restraints, which can limit accuracy, particularly for flexible ligands and active sites. To address these limitations, we previously incorporated advanced QM and QM/MM ONIOM formalisms into crystallographic refinement workflows, demonstrating improved structural accuracy and a critical impact on binding affinity prediction [1-3]. Despite these advances, accurate treatment of species exhibiting alternate conformations at the QM level remains a major challenge and is not supported by most computational tools used in industry. QuantumBio’s Discovery Suite features a real-space refinement engine that integrates energy and gradients from QM, MM, or hybrid QM/MM Hamiltonians with electron-density gradients derived from X-ray or Cryo-EM data, enabling robust refinement of coordinates, occupancies, and B-factors. We have now developed and implemented a holistic, high-performance methodology for fully consistent QM/MM treatment of disordered ligands and residues during refinement. This capability is seamlessly integrated into a user-friendly, automated workflow suitable for high-throughput SBDD applications.

In this work, we present large-scale QM/MM refinements of protein-ligand complexes with disordered ligands drawn from the Iridium, CASF, and PDBBind datasets. Although many of these benchmark sets are curated, this curation typically removes alternate ligand and protein conformations from binding sites. We investigate how retaining and accurately refining such disorder at the QM/MM level affects predicted binding affinities and ligand protonation-state scoring using QuantumBio’s patented XmodeScore methodology [4]. Our results highlight the importance of explicitly modeling conformational heterogeneity for improving the reliability of structure-based drug discovery predictions.

[1] Borbulevych, O. Y., Plumley, J. A., Martin, R. I., Merz K. M. & Westerhoff, L. M. (2014). Acta Cryst., D70, 1233.

[2] Borbulevych, O. Y., Martin, R. I. & Westerhoff, L. M. (2018). Acta Cryst., D74, 1063.

[3] Borbulevych, O. Y., Martin, R. I. & Westerhoff, L. M. (2021). J. Comput. Aided Mol. Des., 35, 433.

[4] Borbulevych, O. Y., Martin, R. I. & Westerhoff, L. M. (2016). Acta Cryst., D72, 586.

Structure-guided drug discovery increasingly depends on experimental structures that contain intrinsic disorder, partial occupancy, and heterogeneous ligand binding, yet many refinement workflows still collapse these signals into single-conformer models. This not only creates ambiguity at key medicinal chemistry decision points (binding hypothesis, pose selection, and chemical state assignment), but also limits the quality of structural datasets used to train and validate artificial intelligence and machine learning (AI/ML) models, which are often underexposed to disordered loops and conformationally heterogeneous regions. These challenges are amplified for modern therapeutic modalities—including macrocycles, cyclic peptides, proteolysis-targeting chimeras (PROTACs), covalent inhibitors, and metal-containing ligands—where flexibility, correlated protein-ligand rearrangements, and multiple coexisting states are often central to mechanism.

We present a density-driven quantum mechanics/molecular mechanics (QM/MM) ensemble refinement workflow that explicitly represents disorder using group- and state-based models with partial occupancies. Candidate protein conformers are generated through backbone φ/ψ sampling and rotamer enumeration, while candidate ligand hypotheses are generated through pose sampling and chemical-state enumeration, including alternate tautomers and protonation states. Candidates are filtered and ranked using real-space Z-score of the difference density (ZDD) scoring to maximize consistency with experimental density for both X-ray crystallography and cryogenic electron microscopy (Cryo-EM). Refinement then applies occupancy-aware QM/MM energetics in which intra-state chemistry is preserved while inter-state competition is captured through occupancy-weighted interactions, including explicit presence/absence states and symmetry-aware contacts. The refinement cycles coordinate updates to atomic coordinates, atomic displacement parameters, and occupancies to convergence, enabling practical ensemble modeling rather than post hoc interpretation.

Using case studies drawn from pharmaceutically relevant targets, we show how disorder-aware refinement improves interpretability for structure-guided drug discovery by reducing artificial ligand strain, clarifying protein-ligand interaction networks, and distinguishing competing hypotheses such as alternate poses, fragment mixtures, and heterogeneous ligands occupying the same site. By retaining conformational heterogeneity instead of forcing a single averaged structure, the resulting ensembles provide improved starting points for molecular dynamics sampling, replica exchange, metadynamics, and clustering workflows, and provide higher-quality inputs for AI/ML applications that depend on accurate representations of alternative states. This talk will emphasize validation metrics and practical decision points in drug discovery, while highlighting the refinement-method advances required to make disorder and heterogeneous binding scientifically rigorous and operationally useful.