Chapter Four · failure evidence
What Structural Docking & Molecular Mechanics got wrong, from 41 dissertations
Computational molecular docking and structural scoring methods frequently encounter failures stemming from rigid approximations, inaccurate scoring functions, and flawed receptor conformations. These records highlight how simulated poses often drift away from binding pockets, while machine learning scoring functions frequently fail to correlate with experimental affinity or lose to simpler baselines. These records come from PhD theses at 19 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Rigid receptor and rigid backbone approximations cause steric clashes and fail to capture conformational adaptation
Rigid-body docking setups and static receptor models frequently generated severe steric clashes, negative fitness scores, and distorted or unfolded peptide poses. Researchers regularly abandoned rigid docking in favor of flexible-residue or ensemble methods to properly accommodate ligand binding and transient pocket rearrangements.
Considered and rejected
Considered and rejected: Single static docking snapshots / docking binding free energies rejected due to inability to capture protein induced fit, binding kinetics, and solvent entropy.
Considered and rejected
Considered and rejected: AutoDock Vina and HADDOCK docking were rejected for modeling the LptH-thanatin complex due to lack of backbone flexibility and generating unfolded/unrealistic peptide poses
Considered and rejected
Considered and rejected: Rejected rigid-body Fourier transform-based docking models due to erratic and ineffective results, adopting flexible docking with GOLD instead.
Use of Computational Methods To Understand The Pattern Of Antimicrobial Resistance · University of Nottingham Repository
Considered and rejected
Considered and rejected: Rejected rigid-body docking in favor of flexible docking with HDOCK/HADDOCK to allow atomic side-chain rotations and mutual conformational adaptation at large protein-protein interfaces.
El rol estructural de la subunidad β3 como regulador de hNav1.7 · Repositorio Institucional BUAP
Considered and rejected
Considered and rejected: Rigid receptor docking was rejected in favor of flexible-residue docking within a 5 Å active-site radius to correctly capture aromatic ligand stabilization.
Considered and rejected
Considered and rejected: Rejected rigid receptor docking in GOLD due to extreme steric clashes and negative fitness scores.
Experimental and computational study of latex clearing protein LcpK30 for rubber degradation · University of Nottingham Repository
Considered and rejected
Considered and rejected: AutoDock Vina rigid-body docking rejected for variants containing active-site F87/bulky residues due to severe steric exclusion of productive poses; ADFR flexible-receptor docking used instead.
Enantioselective remote functionalisation of cyclic amines by engineered P450BM3 variants · Oxford
Considered and rejected
Considered and rejected: Constraining rotational bonds of the 90-residue EREG and TGFα substrate segments during AutoDock Vina docking to treat them as rigid bodies due to structural complexity.
Considered and rejected
Considered and rejected: Rigid-receptor docking purely against the apo crystal structure was rejected because it failed to account for the transient subpocket opened by His253 rotation.
Structure-Based Discovery of Lipoteichoic Acid Synthase Inhibitors. · Cambridge
Docking scores fail to correlate with experimental binding affinity and functional activity
Predicted docking energies and scoring ranks showed weak or nonexistent correlation with biophysical binding assays, positive controls, and cellular potency measurements. Scoring algorithms also failed to distinguish partial from full agonists and retained non-zero baseline correlations even at massive structural docking errors.
Tried and failed
molecular docking virtual screening compared to biophysical assays applied to fragment library screening against RNA targets. Outcome: no signal. Reason: computational docking hits showed zero overlap with surface plasmon resonance experimental binding hits
Tried and failed
structure-based molecular docking score ranking applied to target enzyme ligand screening. Outcome: no signal. Reason: docking scores showed no correlation with known positive and negative control performance
Tried and failed
3D CNN scoring of molecular docking poses applied to protein kinase ligand binding prediction. Outcome: did not generalise. Reason: experimental assays failed to confirm predicted inhibitor binding to homology-modeled targets
Leveraging diverse data modalities to study kinase inhibitor polypharmacology · Harvard
Tried and failed
Voronoi-based molecular docking with force-field optimization applied to zeolite structure-directing agent binding ranking. Outcome: did not generalise. Reason: Failed to rank experimentally validated host-guest pairs favorably due to scoring/sampling limitations in rigid cavities
Lost to a baseline
Molecular docking scores showed a weak correlation with functional potency (R2 < 0.1), failing to accurately predict cellular IC50 ranking.
Design, Synthesis, and Development of Novel Intracellular Allosteric Modulators of CXC Chemokine Receptor 1 and 2 · University of Nottingham Repository
Lost to a baseline
On CASF 2016 docking RMSD bins (e.g. 0-1Å and 1-2Å), all MLBSFs and physics-based scoring plateaued at ~10Å docking error and retained non-zero baseline correlation even at extreme docking errors (25-30Å), matching non-structural ligand bias behavior.
Towards addressing structural data limitations in machine learning for small molecule drug discovery · Oxford
Considered and rejected
Considered and rejected: Scoring solely based on docking energy calculations was rejected because it fails to distinguish partial agonists from full or dual agonists.
Defining Novel Clusters of PPAR gamma Partial Agonists for Virtual Screening · Virginia Tech
Considered and rejected
Considered and rejected: Molecular docking scoring (AutoDock4) was rejected as a quantitative alternative/replacement for experimental G-quadruplex DNA binding affinity measurements due to inability to preserve true rank order and errors exceeding method uncertainty (2.5 kcal/mol)
Docking against inaccurate homology models or incorrect oligomeric states creates clashes and pocket collapse
Using unliganded or homology-modeled targets led to docking failures because closed conformations lacked essential binding pockets or substrate tunnels. Docking into isolated monomers instead of correct oligomeric assemblies also caused severe steric clashes by placing contacts at internal subunit interfaces.
Tried and failed
molecular docking on homology-predicted protein structures applied to small molecule ligand-target binding identification. Outcome: no signal. Reason: predicted closed-conformation model failed to capture appropriate binding pockets for the ligand
Understanding responses to chemical and environmental perturbations in blood stage Babesia parasites · Harvard
Tried and failed
rigid-body protein-protein macromolecular docking applied to heteromeric enzyme complex modeling. Outcome: did not converge. Reason: produced over 400 highly variable models without a consensus interface
Structural investigations of adenosylcobalamin-dependent enzyme maturation · MIT
Tried and failed
molecular docking of cofactor into active site applied to predicted oxidoreductase enzyme structures. Reason: severe steric clashes and absence of a substrate-binding tunnel prevented docking
Expanding the Tree of Life and altering models of eukaryogenesis · UT Austin
Considered and rejected
Considered and rejected: Rejected docking studies against homology models based on E. coli or A. thaliana IspD crystal structures due to large unstructured insertions in PfIspD and lack of cross-species potency.
Novel Antimalarial Compounds from the Optimization of the Malaria Box · Virginia Tech
Considered and rejected
Considered and rejected: Direct receptor-ligand docking of homotrimers was rejected because it produces an artificial dimeric state instead of a trimeric assembly, requiring symmetry docking instead.
Considered and rejected
Considered and rejected: Rejected performing molecular docking between collagen and the monomeric NTHiENO model because contact residues mapped to internal dimer interfaces causing steric clashes, requiring a dimeric homology model.
Evaluación de la enolasa recombinante de Haemophilus influenzae como un factor de virulencia y su potencial uso para el desarrollo de una vacuna · Repositorio Institucional BUAP
Considered and rejected
Considered and rejected: Rejected building CXCR1 homology models in Modeller without the template-bound ligand, as unliganded models collapsed the binding cavity due to rotamer shifts of I64(1.57).
Design, Synthesis, and Development of Novel Intracellular Allosteric Modulators of CXC Chemokine Receptor 1 and 2 · University of Nottingham Repository
Docked ligands suffer from instability and unconstrained drift away from target binding sites
Docked complexes frequently exhibited instability during simulation or refinement, with ligands drifting away from active catalytic pockets to off-target locations. Unconstrained sampling or lack of pre-refinement also yielded distorted geometries and geometric deviations that degraded downstream scoring accuracy.
Tried and failed
graph theoretical energy scoring on docked poses applied to allosteric ligand affinity ranking. Reason: scoring accuracy degraded significantly due to docking pose inaccuracies and geometric deviations
Allostery and signalling pathways: methods and applications through graph theory · Imperial
Tried and failed
molecular docking with score thresholding applied to RNA-fragment binding site prediction. Reason: High-scoring conformers docked to off-target residues outside the targeted binding site
Tried and failed
molecular dynamics refinement of docked ligand poses applied to protein-ligand complex binding pose validation. Outcome: unstable. Reason: docked ligand poses drifted away from the binding site during simulation
Structure-based discovery of lipoteichoic acid synthase inhibitors. · Imperial
Tried and failed
constrained docking followed by unconstrained validation docking applied to enzyme active site mutant design. Outcome: did not generalise. Reason: predicted intermediate poses were unstable and drifted away from target catalytic sites when geometric constraints were removed
Computational and structural investigations of class I diterpene synthases · Iowa State
Tried and failed
linear interaction energy for binding affinity estimation applied to protein-ligand complex molecular dynamics simulations. Outcome: no signal. Reason: ligand stabilized at an off-target site outside the catalytic pocket during simulation
Tracing the Evolution of Substrate Specificity in Low-Molecular-Weight Protein Tyrosine Phosphatases and Arsenate reductases · Georgia Tech
Tried and failed
direct flexible backbone docking with non-canonical amino acids applied to peptide-protein complex modeling. Outcome: unstable. Reason: produced distorted bond angles and lengths without initial canonical amino acid backbone pre-refinement
Machine learning scoring and generative models overfit to oracles and fail to generalize
Deep learning scoring functions trained on specific active sites failed to generalize to allosteric binding modes or were outperformed by standard docking baselines. Reinforcement learning driven purely by docking score rewards exploited the oracle to generate overly large, lipophilic compounds that lacked drug-likeness or failed experimental validation.
Tried and failed
unsupervised word2vec molecular embeddings as sole features applied to drug-target docking score prediction. Outcome: worse than baseline. Reason: mol2vec features alone lacked sufficient structural and chemical specificity compared to concatenated domain-specific fingerprints
Advancing Personalized Medicine Through Generative Artificial Intelligence · Georgia Tech
Tried and failed
3D-CNN scoring of docked protein-ligand poses applied to allosteric inhibitor binding prediction. Outcome: did not generalise. Reason: Training solely on orthosteric active-site docked structures failed to generalize to allosteric binding modes
Leveraging diverse data modalities to study kinase inhibitor polypharmacology · Harvard
Tried and failed
reinforcement learning on docking score objective alone applied to de novo small molecule generation. Outcome: overfit. Reason: model exploited the scoring oracle, generating overly large, highly lipophilic molecules with poor drug-likeness
Language-based Sample-efficient and Synthesizable Molecular Generative Design · EPFL
Lost to a baseline
DOVE (CNN docking tool) was beaten by native ZDOCK on reranking antibody model docking poses.
Lost to a baseline
On Smina redocking, PoseTriager trained on Augmented2020 (MCC 0.34) was beaten by PoseTriager trained on standard Redocked2020 (MCC 0.44 ± 0.02).
Towards addressing structural data limitations in machine learning for small molecule drug discovery · Oxford
Tried and failed
geometric deep learning for de novo ligand design applied to protein-ligand binding generation. Outcome: no signal. Reason: computationally designed compound failed to show measurable binding affinity in experimental assays
Computational Design of Structure-Guided Biomolecular Interactions · EPFL
Voronoi tessellation and geometric surface search algorithms underperform relative to Monte Carlo and rigid baselines
Voronoi-based docking algorithms struggled to generate stable complexes for bulky molecules in constrained pores when optimal positions fell outside Voronoi nodes. On standard protein-protein benchmarks, surface-based search methods were beaten in top-ranked solution rates by traditional rigid docking baselines.
Tried and failed
Voronoi tessellation and Monte Carlo rigid docking applied to zeolite framework guest molecule docking. Outcome: unstable. Reason: Failed to generate stable docked complexes for bulky template molecules in constrained pore frameworks
Lost to a baseline
Monte Carlo outperforms Voronoi docking when optimal positions cannot be reached by placing the center of mass on Voronoi nodes.
Lost to a baseline
Monte Carlo docking generated stable complex structures for 113 pairs, whereas Voronoi docking generated stable structures for 107 pairs out of 120.
First-principles control of zeolite synthesis, transformations, and intergrowth · MIT
Lost to a baseline
In oracle iRMSD on the rigid docking test set, FRODock achieved a median of 1.9 Å compared to DiffMaSIF's 3.5 Å
Predicting protein interactions using geometric deep learning on protein surfaces · EPFL
Lost to a baseline
On the holo/bound large-scale docking benchmark (100 targets), ZDock+ZRank2 solved 45 complexes in Top 1 (and 63 in Top 10) compared to MaSIF-search (2000 decoys) solving 43 in Top 1 (and 56 in Top 10)
Predicting protein interactions using geometric deep learning on protein surfaces · EPFL
Lost to a baseline
On unbound (apo) docking benchmark top-10, ZDock (5 solved) and ZDock+ZRank2 (5 solved) outperformed MaSIF-search (2 solved)
Predicting protein interactions using geometric deep learning on protein surfaces · EPFL
Left open by the authors
Problems the authors named and did not get to.
Left open
Simulate membrane deformation and endocytosis initiation driven by localized multivalent receptor-ligand accumulation in fluid membranes. Blocker: Lacks specific model parameters, coupling mechanism to deformable membrane models, and benchmark targets
Left open
Evaluate Graph GA using molecular docking software augmented with synthetic noise sampled from an assumed error distribution. Blocker: None
Molecular Property Predictors for Downstream De Novo Generation · Harvard
Left open
Decouple chemical space search from conformational search by incorporating molecular dynamics simulations or docking tools for binding pocket conformations. Blocker: None
Scalable Fragment-Based 3D Molecular Design with Reinforcement Learning · Harvard
Left open
Integrate propensity optimised paths with ensemble docking to evaluate multiple ligand poses in fragment-based virtual screening for allosteric drug design. Blocker: None
Allostery and signalling pathways: methods and applications through graph theory · Imperial
Left open
Integrate ligand pose scoring and drug-like property predictors into the LDDM generative pipeline to constrain hit identification. Blocker: None
Computational Design of Structure-Guided Biomolecular Interactions · EPFL
Left open
Test the iterative batch Bayesian Optimization and D-MPNN-guided genetic algorithm on multi-objective optimization combining synthesizability, off-target effects, and docking scores. Blocker: None
Molecular Property Predictors for Downstream De Novo Generation · Harvard
Left open
Dock diverse pathogen-derived lipid structures into primate CD1a structural models using molecular docking tools to test binding affinity. Blocker: None
Adaptive Evolution in Primate Immune Receptors · Scholars' Bank
Left open
Experimentally validate the computational docking and molecular dynamics model of the proteusin precursor peptide interacting with peptide brominase SrpI. Blocker: Requires wet-lab experimental validation (biochemical binding/activity assays)
Biosynthesis Guided the Discovery of Proteusin RiPP Natural Products · Georgia Tech
Left open
Finetune GeoMol on molecular docking pose generation or 4D QSAR descriptor prediction datasets. Blocker: None
Towards Automated Reaction Kinetics with Message Passing Neural Networks · MIT
Left open
Benchmark RoseTTAFold All-Atom on predicting ligand-induced protein conformational changes using public apo/holo paired datasets. Blocker: None
Explainable Machine Learning and Applications in Protein-Ligand Complex Structure Prediction · ResearchWorks
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.