Many antibiotics are chiral molecules: they come in mirror-image and near-mirror variants, collectively called stereoisomers, that share the exact same atom-by-atom connectivity, molecular formula, and mass, and differ only in the three-dimensional orientation at a handful of "handed" atoms called stereocenters. Because every biopolymer in a living cell is itself handed, these seemingly trivial variants can interact with biology in completely different ways — one stereoisomer of a Pseudomonas quorum-sensing analog fully activates the LasR virulence regulator while its sibling stereoisomers show near-zero agonism or act as competitive antagonists Stereoisomers of P. aeruginosa Autoinducer Analog to Probe the Regulator Binding Site. Meanwhile, the antibiotic resistance crisis keeps worsening: resistance genes such as $bla_{NDM-1}$ and $mcr-9$ now co-travel on single conjugative plasmids, simultaneously neutralizing carbapenems and colistin, the last-resort drugs 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. Crucially, bacteria resist drugs largely through molecular recognition: β-lactamase enzymes recognize a drug's three-dimensional shape and cleave it, and efflux pumps recognize chemical shape and charge and eject it Microbial Characteristics and Genomic Analysis of an ST11 Carbapenem-Resistant Klebsiella pneumoniae Strain Carrying blaKPC−2 Conjugative Drug-Resistant Plasmid. These machines are, functionally, classifiers.
This proposal treats those resistance machines as adversarial classifiers and stereocenter inversion as the adversarial perturbation. In computer vision, "stealth" adversarial attacks modify an input enough to fool a detector while staying on the manifold of natural data, so no human inspector flags the change — for example, searching the latent space of a face generator so that a perturbation manifests as ordinary makeup CLIP2Protect: Protecting Facial Privacy using Text-Guided Makeup via Adversarial Latent Search. Chemistry has an exact analog that no one has exploited: for a molecule with $k$ stereocenters, the set of all $2^k$ stereo-configurations — its stereoisomer lattice — is a search space in which every point is, by construction, a legal, synthesizable molecule with unchanged two-dimensional graph, formula, mass, and gross physicochemical profile, yet three-dimensional recognition can flip dramatically. We will formalize antibiotic-resistance evasion as a constrained search over this lattice: find the configuration the resistance enzyme no longer recognizes, subject to the bacterial target still binding it.
Core Claim
Stereocenter inversion is a well-defined, chemically valid, discrete adversarial perturbation class — the molecular counterpart of manifold-constrained perceptual stealth — and searching it with chirality-aware surrogate models can produce antibiotic variants that evade resistance recognition while retaining target activity. This requires two things that current models cannot do. First, the surrogates must actually distinguish stereoisomers: standard graph neural networks treat neighboring atoms as an unordered bag and are mathematically incapable of separating enantiomers ChiENN: Embracing Molecular Chirality with Graph Neural Networks. Second, the search must be adversarial: it must optimize against the failure modes of resistance machinery, not just for target affinity, the way adversarial red-teaming agents in TRIAGE: An AI Scientist for Adversarial Target Falsification outperformed passive generative filtering by actively hunting for counter-evidence.
Why the Intersection Is Empty
Three bodies of work come close but leave this exact gap open, and the research file's own synthesis states plainly that no paper unifies them. The chirality-modeling line — ChiENN, ChIRo, and ChiDeK — built encoders that separate mirror images, but evaluated them only on synthetic benchmarks (R/S labels, optical rotation, docking) and never on pathogen or resistance data. The antibiotic-screening line — CL-MFAP and the ESCAPE peptide benchmark — predicts minimum inhibitory concentration from SMILES strings and 2D graphs and is functionally stereo-blind; CL-MFAP itself fell to fifth place on bioavailability precisely because it omits 3D conformation. The adversarial line has entered biology only through one-dimensional sequence mutations: Structural Memorization in AlphaFold probes protein structure predictors with point mutations and deletions, never with stereochemical perturbations of ligands.
The single closest work is PepMirror, which designs protease-resistant all-D peptides by injecting axial vectors into an equivariant diffusion model. It differs from this proposal in three load-bearing ways. It flips the entire molecule to its mirror image rather than searching partial inversions (epimers); it is motivated by protease stability, not formulated as an adversarial optimization against a recognition model; and it addresses only peptide backbones against a eukaryotic target, never small-molecule scaffolds or resistance enzymes. Notably, PepMirror's own wet-lab result undercuts the whole-mirror strategy as a general recipe: the L-enantiomer of its hit D-1412 also bound CD38, showing stereoselectivity is interface-dependent — which is exactly why one must search the lattice rather than jump to its mirror endpoint.
Research Program
The formal object is the following. Let $m$ be an antibiotic scaffold with $k$ stereocenters, $c \in \{R,S\}^k$ a stereo-configuration, and $c_0$ the parent configuration. Let $R(m,c)$ be a surrogate for recognition by a resistance determinant (probability a β-lactamase hydrolyzes it, a protease cleaves it, or an efflux pump exports it), and $A(m,c)$ a surrogate for activity at the therapeutic target. We search
$$\min_{c \in \{R,S\}^k}\; R(m,c) \quad \text{subject to} \quad A(m,c) \;\ge\; \theta\, A(m,c_0),$$
with the stealth constraint structural rather than parametric: the two-dimensional graph is fixed, so every candidate is synthesizable and nearly identical in gross properties by construction. The number of inverted centers, $|c \oplus c_0|$, acts as an explicit inversion budget, letting us trace how recognition decays per inversion.
Three components get built. First, chirality-aware surrogates: encoders built from signed-determinant local geometry and chiral cross-attention as in ChiDeK (whose reference implementation is public under an MIT license, Meteor-han/ChiDeK), and dual-edge chiral message passing as in ChiENN, trained as ensembles over multiple generated conformers. The recognition surrogate for β-lactamases is trained on substrate and inhibitor bioactivity from the ChEMBL-derived screening corpus used by CL-MFAP; the activity surrogate on the same MIC panels plus the chirality-aware binding-affinity setup where ChiENN layers previously cut mean absolute error from 0.330 to 0.258 ChiENN. For peptides, the lattice generalizes naturally: each residue's L/D choice is one stereocenter, activity surrogates train on the 82,359-sequence ESCAPE collection, and protease susceptibility is the recognition surrogate, with PepMirror's axial-vector trick available when a folded target needs explicit reflection handling. Second, the search itself: branch-and-bound over the lattice ordered by the evasion surrogate, with bounds from monotonicity estimates of recognition decay — this scales where brute-force enumeration of $2^k$ cannot, and for a 12-mer peptide ($k = 12$, 4096 configurations) enumeration is already at the edge of docking feasibility while a 30-mer is hopeless. Third, an audit layer: because the surrogates lean on structure prediction for resistance-enzyme pockets, we probe those predictions with enantiomeric pairs, extending the adversarial-probing methodology of the AlphaFold memorization study from sequence mutations to stereochemistry.
Proposed Experiments
The program runs five experiments, all on benchmarks present in the research file. Experiment 1 verifies the enabler: our stereo-lattice encoder must reproduce chiral discrimination at published levels — R/S classification on the PubChem3D subset where ChiDeK reports 99.8%, ChiENN 98.9%, and ChIRo 98.5% accuracy ChiDeK; ChiENN; ChIRo — and enantiomer-sensitive docking on the PDB:4JBV pocket benchmark where ChIRo reports 72.0% ranking accuracy. Experiment 2 trains and calibrates the two surrogates, reporting binding-affinity MAE against the GPS-with-ChiENN figure of 0.258 and MIC ROC-AUC against CL-MFAP's 0.854 on E. coli (with MoLFormer's 0.71 as the stereo-blind reference point) CL-MFAP. We additionally construct an enantiomer-paired MIC subset from the ChEMBL corpus — molecules where both stereoisomers have measurements — and demonstrate that 2D models are structurally tied on such pairs; no reported result found on Lacuna for enantiomer-paired MIC prediction, so this subset is itself a contribution. Experiment 3 runs the evasion search on two case studies with ground truth: the LasR autoinducer stereoisomers, where published agonism/antagonism labels let us test whether the search recovers known activity-divergent lattice points LasR stereoisomers, and β-lactamase evasion for carbapenem-scaffold epimers against the KPC-2 and NDM-1 enzyme contexts documented in the plasmid-genomics studies. Experiment 4 is the AlphaFold audit: we dock enantiomeric and epimeric probe pairs into AlphaFold 3-predicted pockets of resistance enzymes and ask whether interface scores and confidence distinguish stereoisomers at all. The memorization study predicts trouble — confidence scores selected the best structure only ~25% of the time and interfaces degraded fastest under perturbation Structural Memorization in AlphaFold — so this experiment decides whether the pipeline may lean on AlphaFold or needs ESMFold-plus-physics rescoring. Experiment 5 runs the peptide regime: L/D lattice search over ESCAPE-derived candidates scored for activity retention and protease-evasion, benchmarked against whole-mirror inversion.
Baselines
The do-nothing reference point is the parent configuration $c_0$ itself: under enzyme pressure it is, by hypothesis, recognized and neutralized, and on the screening benchmarks this corresponds to the stereo-blind status quo — CL-MFAP's multimodal 2D model, which reports a ROC-AUC of 0.854 on E. coli MIC classification with MoLFormer at 0.71 CL-MFAP. A second naive reference is a uniformly random lattice point, which calibrates the base rate of accidental evasion at a given retention threshold $\theta$; beating this is trivial and stated only for calibration. The strongest existing competitor differs by regime. For small-molecule screening it is CL-MFAP, retrained on our splits with the same ChEMBL data and the same hyperparameter search budget and compute as our surrogates; because its architecture cannot distinguish enantiomers, the decisive comparison is the enantiomer-paired subset, where any stereo-blind model is structurally tied regardless of tuning, and on the full benchmark our model must match or exceed its retrained ROC-AUC to claim the chiral encoder costs nothing in accuracy. For the evasion task and the peptide regime, the strongest existing method is PepMirror's whole-mirror design, which achieved nearly 100% chirality consistency and one wet-lab binder (D-1412, $K_D \approx 10$ µM) out of twelve synthesized candidates PepMirror; we run its full-inversion output as a single lattice point under our surrogates, with the same target-retention threshold applied, and count beating it as finding a partial-inversion point with higher predicted target activity at equal or better evasion. The key ablation swaps the chiral encoder for an E(3)-invariant or purely 2D encoder while leaving the objectives, data, and search untouched. The file's evidence predicts collapse: E(3)-invariant models sat at 50–60% (chance) on R/S classification, and axial-chirality baselines hovered near 50% on optical rotation where ChiDeK reached 69.2% ChIRo; ChiDeK. If the ablated search still finds evading configurations, the chiral encoder is not the key ingredient and the claim must be weakened. A second ablation removes the evasion objective, keeping only target-affinity maximization over the lattice, to isolate the adversarial formulation from mere stereochemical lead optimization.
Why not just…?
Why not just use the huge 2D foundation models already trained on billions of molecules? They cannot represent the search space, not merely perform worse in it. ChiENN proved that permutation-invariant aggregation makes enantiomers indistinguishable to graph models ChiENN, so a SMILES- or 2D-graph-based surrogate returns identical recognition and activity scores for every lattice point, collapsing the search to noise. Scale does not rescue this: MoLFormer, trained on over a billion molecules, reached only a 0.71 ROC-AUC on E. coli MIC screening versus 0.854 for the domain-specific CL-MFAP CL-MFAP. We will still run these models as baselines on the paired subset, where the tie is a structural property we expect to demonstrate empirically.
Why not just flip the whole molecule into its mirror image? PepMirror shows whole-mirror inversion works when the goal is protease stability, but it is one point of the lattice, chosen blind to the target's stereospecificity. The LasR stereoisomer study shows sibling configurations range from full agonists to competitive antagonists at the same regulator LasR, and PepMirror's own observation that the L-form of D-1412 also bound CD38 confirms that stereoselectivity varies interface by interface. Wholesale inversion therefore sometimes destroys target activity and sometimes is unnecessary; a search with an explicit retention constraint handles both cases. The full-mirror baseline is run in Experiment 5, so this is settled empirically rather than by argument.
Why not just enumerate all $2^k$ stereoisomers and dock each one? For small-molecule cases with $k \le 4$ this is a fair and cheap baseline, and we will run it; if it recovers the same top candidates our search returns, we will report that honestly, and the proposal's value would then rest on the evasion objective and the large-$k$ peptide regime. Two things keep it from being sufficient in general. Docking scores target binding only — it has no counterpart for whether a β-lactamase recognizes and hydrolyzes the molecule or an efflux pump exports it, which is half the objective. And structure-prediction-based docking on resistance enzymes is the weakest link in the chain: AlphaFold 3 memorizes templates rather than modeling perturbed biophysics, retaining global folds at TM-score $\ge$ 0.5 under 40% mutation while its confidence scores select the best structure only ~25% of the time Structural Memorization in AlphaFold. For a 30-mer peptide, $2^{30}$ dockings are also simply infeasible.
Why not just do traditional medicinal chemistry — add bulky groups to block the enzyme? This is the rebranding critique, and it deserves a direct answer: static steric blocking has been tried at scale against β-lactamases and has not kept pace, because resistance enzymes diversify and spread horizontally on mobile plasmids faster than scaffolds iterate blaKPC-2 plasmid study; mcr-9 co-transmission study. The adversarial formulation differs operationally, not just terminologically: it optimizes explicitly against a learned model of the recognizer's failure modes — the lesson TRIAGE established when adversarial agents outperformed passive generative filtering TRIAGE — and it changes the perturbation vocabulary from adding atoms (which alters every physicochemical property and restarts safety profiling) to inverting orientations (which preserves them). The clinical precedent that isolated stereoisomers can improve therapeutic index, as with S-ketamine lowering dose below the hallucination threshold S-ketamine RCT, suggests the vocabulary is underused.
Decisive Checks and Risks
Three checks decide the proposal. The paired-MIC experiment must show stereo-blind models are tied on enantiomer pairs while our surrogate separates them; the LasR experiment must recover known agonist/antagonist lattice points from 2D structure alone; and the ablation must show the search degenerates without the chiral encoder. The main acknowledged risks are the conformer problem — cheap generated conformers systematically mismatch quantum-mechanical ground truth, which the ChiDeK authors flag as their limiting factor ChiDeK, mitigated by multi-conformer ensembles since ChIRo held ~79% optical-rotation accuracy on single conformers where SphereNet dropped from 65% to 55% ChIRo — and the surrogate-to-wet-lab gap, which PepMirror's one-in-twelve hit rate makes vivid PepMirror. This proposal therefore claims validated computational methodology, not new drugs: a well-defined adversarial perturbation class, the first chiral surrogates of resistance recognition, and an audit of whether current structure prediction can be trusted to score stereoisomeric antibiotics at all.
References
- Stereoisomers of P. aeruginosa Autoinducer Analog to Probe the Regulator Binding Site
- 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
- Microbial Characteristics and Genomic Analysis of an ST11 Carbapenem-Resistant Klebsiella pneumoniae Strain Carrying blaKPC−2 Conjugative Drug-Resistant Plasmid
- CLIP2Protect: Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search
- ChiENN: Embracing Molecular Chirality with Graph Neural Networks
- TRIAGE: An AI Scientist for Adversarial Target Falsification
- Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations
- Learning Molecular Chirality via Chiral Determinant Kernels
- CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening
- A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification
- Structural Memorization in AlphaFold: Adversarial Mutations Reveal Template Reliance, Confidence Failures, and Implications for Protein Design
- Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design
- S-ketamine promotes postoperative recovery of gastrointestinal function and reduces postoperative pain in gynecological abdominal surgery patients: a randomized controlled trial