When a bacterium develops resistance to an antibiotic, the drug's binding pocket on the bacterial target usually does not disappear — it reshapes. A single amino-acid substitution in a carbapenemase enzyme rotates a sidechain, shifts an active-site loop, or opens a new sub-pocket, and the drug no longer fits. Most computational drug design cannot see this, because it treats the target protein as a frozen geometric mold. A newer line of work, induced fit modeling, instead learns to generate the drug and the protein's shape changes together: the pocket's sidechains and backbone move as the small molecule binds. Meanwhile, the antibiotic resistance literature documents exactly which clinical mutations erode last-resort drugs, but its machine-learning models classify resistance genes from static structure or predict affinity changes from rigidly docked complexes. This proposal argues these two communities are holding complementary halves of one problem, and builds the missing piece: a generative model that designs molecules jointly conditioned on a wild-type target and its clinical escape mutants, where each mutant pocket is allowed to relax in its own induced-fit way.
The concrete ambition is an antibiotic candidate that is not merely potent against today's enzyme, but robust across tomorrow's mutants — a worst-case-over-variants design objective rather than an average-case one. No work in the surveyed literature does this: the closest attempts either pair two distinct rigid proteins (Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design) or measure mutation-induced affinity loss on rigidly docked structures (MdrDB: Mutation-induced drug resistance DataBase).
Core Claim
The claim is that resistance-robust antibiotic design requires both ingredients at once, and that dropping either one fails for documented reasons. Dropping induced fit (i.e., treating mutant pockets as static) fails because escape mutations remap sidechain packing and clash geometry that rigid docking cannot resolve — APObind: A Dataset of Ligand Unbound Protein Conformations for Machine Learning Applications in De Novo Drug Design shows rigid docking success collapsing from 57.94% on bound structures to 13.74% on unbound ones when the receptor's true conformation differs from the template, and FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling shows rigid-target generators producing 6.43 average steric clashes versus 1.39 when sidechain and backbone motion are co-generated. Dropping the multi-mutant conditioning (i.e., designing flexibly for the wild type alone) fails because a ligand optimized into one pocket conformation can be precisely the ligand a single mutation displaces. The proposal therefore couples a flexible flow-matching generator — a model that learns to gradually denoise random noise into a bound protein–ligand complex, in the family of FlexSBDD and Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows — to an explicit ensemble of target variants, optimized for the worst case:
$$ \max_{x} \; \min_{v \in \mathcal{V}(P)} \; \text{Affinity}\big(x, R_v(x)\big) $$
where $x$ is the generated ligand, $\mathcal{V}(P)$ is the set containing the wild-type target and its documented escape mutants, and $R_v(x)$ is variant $v$'s receptor conformation after its own induced-fit relaxation around $x$. The min says: the molecule is only as good as its affinity for the mutant it binds worst. This robustness objective is the conceptual departure from all surveyed work; the induced-fit machinery is the means of making it tractable, since the generator itself supplies the relaxed conformations $R_v(x)$ instead of requiring a separate simulation per variant.
Research Program
The program has three parts, ordered so that each de-risks the next.
First, we rebuild mutation–affinity data with flexible conformations. MdrDB already standardizes 100,537 mutation–drug pairs as binding free-energy changes ($\Delta\Delta G$), but the authors themselves acknowledge that every mutant complex lacking a crystal structure was produced with smina, a rigid-receptor docking tool, so induced-fit adjustments after mutation are systematically missing. We re-generate those complexes with a flexible docking model in the style of Fast and Accurate Blind Flexible Docking, which regresses full complex coordinates in about 0.49 seconds per pair — fast enough to cover a dataset of this size, and physically more faithful than generative sampling per pair. Where experimental mutant holo structures exist among MdrDB's 95,971 PDB-derived samples, they serve as ground truth for pose recovery; where they do not, the rebuilt complexes at least remove the known rigid-docking artifact. Critically, we also extend coverage beyond MdrDB's oncology skew toward bacterial resistance targets documented in the surveillance literature — carbapenemases circulating on conjugative plasmids (Lingyi Zeng et al.), colistin–metallo-β-lactamase co-resistance plasmids (Kewang Hu et al., Jie Li), and ribosomal mutations conferring linezolid resistance in intensive care (Jiachang Cai et al.). No reported result found on Lacuna for flexible generative docking or generation on β-lactamase or ribosomal resistance targets, so this extension is genuinely unmapped territory.
Second, we build the generator, which we call EscapeFit. It is a conditional flow-matching model whose conditioning input is not one pocket but a variant set: the wild-type structure plus $k$ mutant structures, each represented at the two scales FlexSBDD showed matter — rigid backbone frames and torus-parameterized sidechain torsions, with sidechain flexibility weighted most heavily since it is the primary mechanism for opening sub-pockets and locking ligands in place. During generation the model emits the ligand and, for each variant, its relaxed pocket; a per-variant affinity head in the spirit of FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction scores each (ligand, relaxed-variant) pair, and training penalizes the worst variant, implementing the min above. To keep the molecule chemically identical across variants, the model generates one ligand and $k$ pocket deformations rather than $k$ complexes — the cross-variant consistency that dual-target models like Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design achieve by aligning two pockets around a shared ligand, but here with both pockets flexible and the pair being wild-type versus its own escape mutant rather than two unrelated proteins.
Third, we evaluate on three axes: $\Delta\Delta G$ prediction quality on the rebuilt dataset (does flexible geometry improve regression over MdrDB's reported ExtraTrees result of Pearson 0.607 and RMSE 0.656 kcal/mol on its TKI test set, from A mutation-induced drug resistance database (MdrDB)); pose recovery on true unbound mutant structures under the APObind protocol; and generation quality under a new worst-variant metric on the bacterial target panels.
Baselines
The naive reference point is the status quo itself: rigid smina docking of each drug into a static mutant structure, exactly how MdrDB constructed its complexes, with affinity predicted from the resulting geometry and its 146 precomputed features. Its documented failure signature is the holo-to-apo collapse in APObind — 57.94% top-1 success on bound structures falling to 13.74% on unbound ones — which is the same distribution shift a mutation-induced conformational change produces. Any proposal must at minimum not inherit this.
The strongest existing competitor is DUALDIFF, the only model in the surveyed literature that generates a single ligand against two protein conditions; it reported a Dual High Affinity score of 36.3% on its DrugCombDB-derived benchmark, against 29.0% for a single-target TargetDiff baseline. We will reprogram DUALDIFF on our wild-type/mutant pairs, giving it the same training structures, the same sampling budget per molecule, the same Vina-based scoring, and the same tuning effort as EscapeFit, so the comparison isolates modeling choices rather than compute. Beating it means a higher fraction of generated molecules achieving high affinity for every variant in the panel — the worst-variant rate — not merely for both on average, plus fewer steric clashes, since DUALDIFF's frozen pockets cannot relieve clashes that a bulky escape substitution introduces.
Two further comparisons complete the picture. One is FlexSBDD applied independently to each variant — fully flexible, but with no cross-variant coupling; it reported an average Vina score of -9.12 on CrossDocked versus -8.39 for its strongest baseline, and reduced average clashes to 1.39 — which isolates the value of the robustness objective as opposed to flexibility alone. The other is the key ablation: EscapeFit with receptor flexibility switched off, so pockets are frozen at their input conformations while the multi-variant conditioning and worst-case objective remain. If the flexible and frozen versions perform comparably on mutant panels, induced fit is not earning its place and the simpler model wins; if the frozen version collapses specifically on mutants that introduce clashes or cryptic sub-pockets, that is direct evidence for the proposal's central claim.
Why not just…?
Why not just dock existing antibiotics into AlphaFold-modeled mutants with rigid docking? This is the cheapest possible experiment and worth doing as part of the dataset rebuild, but the literature quantifies its ceiling: rigid docking into structures whose conformation differs from the bound state drops from 57.94% to 13.74% success in APObind, and MdrDB explicitly concedes its own smina-built mutant complexes miss induced fit. It also only evaluates existing molecules; it cannot propose one that fits a remapped pocket.
Why not just run DUALDIFF on wild-type/mutant pairs? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. DUALDIFF's pockets are frozen, so a mutant that rotates a sidechain into the binding site produces clashes its generator cannot relieve — the failure mode FlexSBDD documents at 6.43 average clashes for rigid-target baselines. Its Dual High Affinity metric is also average-oriented, which can hide exactly the worst-mutant failure that matters clinically. Still, it may win on speed and simplicity; the head-to-head under matched conditions, plus our frozen-receptor ablation, is designed to settle this empirically rather than by assertion.
Why not just run molecular dynamics on each mutant and dock into the snapshot ensemble? MD captures the transitions we care about, and we will use short MD runs on a validation subset precisely because it is honest ground truth. But as a generation-time mechanism it does not scale: DynamicFlow makes the cost argument directly, and PIGEON: Pocket-Inferred Geometric Ensemble Flexible Docking produces full flexible docking ensembles in about 1.5 seconds where MD equilibration runs to microsecond-or-longer timescales. Designing one molecule against $k$ mutants multiplies that cost by $k$ per optimization step, and MD still does not output a molecule — it only reshapes the pocket.
Why not just use sequence- or structure-based resistance classifiers like ARGGNN? Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks is genuinely useful for what it does — its AlphaFold-based graphs lift Macro F1 from 59.7% to 63.78% over the best sequence baseline on COALA — but it classifies genes into resistance classes from static structures. It outputs no affinity, no molecule, and no conformational change. TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction comes closer by adding dynamics, gaining +2.2 AUROC over sequence-plus-structure to reach 0.897 on ProteinGym, but its dynamics are harmonic springs around a single structure, it handles only single substitutions with binary fitness labels, and it designs nothing. Both motivate the gap rather than fill it.
Decisive Checks
Four results would make or break the proposal. First, on the rebuilt mutation dataset, flexible re-docked geometry must improve $\Delta\Delta G$ regression over the ExtraTrees-on-MdrDB reference (Pearson 0.607, RMSE 0.656 kcal/mol on its TKI test set, from A mutation-induced drug resistance database (MdrDB)); we will report this both on MdrDB's original split and on the new bacterial subset, where no reported result found on Lacuna for any flexible docking method means we establish the first reference numbers ourselves. Second, pose recovery on unbound mutant structures must substantially exceed the 13.74% apo reference from APObind, using its paired apo/holo protocol. Third, on generation, the median worst-variant Vina score across each mutant panel must beat the DUALDIFF baseline run under matched conditions, with average clash counts in the range FlexSBDD achieved (1.39), confirming physical plausibility. Fourth — the check that isolates the core claim — the frozen-receptor ablation must degrade specifically on mutants whose resistance mechanism is conformational (loop shifts, sidechain rotations, cryptic sub-pocket opening), and not on purely electrostatic mutations; this differential failure pattern, if present, is the strongest evidence that induced fit is the load-bearing ingredient in mutation-robust design rather than a decorative one.
References
- Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design
- MdrDB: Mutation-induced drug resistance DataBase
- APObind: A Dataset of Ligand Unbound Protein Conformations for Machine Learning Applications in De Novo Drug Design
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows
- Fast and Accurate Blind Flexible Docking
- Microbial Characteristics and Genomic Analysis of an ST11 Carbapenem-Resistant Klebsiella pneumoniae Strain Carrying blaKPC−2 Conjugative Drug-Resistant Plasmid
- Molecular characterization of NDM-1-producing carbapenem-resistant E. cloacae complex from a tertiary hospital in Chongqing, China
- Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex
- Linezolid-resistant clinical isolates of meticillin-resistant coagulase-negative staphylococci and Enterococcus faecium from China
- FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction
- A mutation-induced drug resistance database (MdrDB)
- PIGEON: Pocket-Inferred Geometric Ensemble Flexible Docking
- Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks
- TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction