Default Bias in Generative Antimicrobial Design

When a machine translation system is asked to translate "my doctor — she is pregnant" into Italian, it frequently produces the masculine form anyway. The model's statistical prior, learned from a corpus in which masculine forms dominate, overrides an explicit cue in the input. This failure mode — a pre-training distribution acting as an attractor that beats explicit conditioning — is now well documented in language systems, along with a second, quieter version of it: automated quality evaluators that systematically score minority variants lower and thereby erase them from filtered datasets. We propose that both versions recur, unrecognized, in generative protein design, and that the antibiotic resistance crisis is where they do the most damage.

Generative models for antimicrobial peptides and antibody-format binders are trained almost exclusively on natural sequence corpora, and their outputs are gated by developability evaluators — automated "watchers" that predict solubility, aggregation, and viscosity — calibrated on clinical human antibodies. Two concrete observations motivate the worry. First, AMPGAN v3, a generative adversarial model for antimicrobial peptides, was explicitly conditioned on target bacterial species and desired potency, yet its wet-lab-validated candidates lost the requested species selectivity (Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3) — precisely the pattern of a prior overpowering a prompt. Second, sequence-only developability predictors are known to systematically underestimate biophysical metrics and reject viable candidates, while screening tools like TAP rely on cutoffs calibrated against standard clinical antibodies (A Guided Design Framework for the Optimization of Therapeutic-like Antibodies) — the pattern of a biased gatekeeper. No paper in the indexed literature has connected these two failure modes, measured them in a protein-design setting, or attempted the NLP-style fix.

Core Claim

The claim is threefold and testable. First, generative protein and peptide models exhibit a conditioning compliance gap — they satisfy explicit generation constraints at a materially lower rate when the constraint conflicts with the training corpus's dominant statistics, the direct analog of the "feminine penalty" in translation. Second, this failure is diagnosable: attribution of the model's output to the conditioning tokens, the technique that exposed suppressed pronoun attention in translation models (A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation), predicts which conditions will be ignored. Third, the fix transfers too: a small set of "hard" demonstrations selected by attribution, or attribution-weighted steering during generation, should close the compliance gap without retraining — and do so more compute-efficiently than rejection sampling, because it tells you which constraints the generator can be made to satisfy rather than blindly resampling.

The application regime is antimicrobial design against drug-resistant Gram-negative pathogens, where resistance determinants now stack on single mobile plasmids — Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex found 25% of clinical carbapenem-resistant E. cloacae carrying inducible colistin resistance alongside metallo-β-lactamases, with colistin exposure driving minimum inhibitory concentrations from 4 to 128 mg/L — and where carbapenem resistance in A. baumannii has reached nearly 80% in surveillance networks (Hui Zhang et al.). Every candidate the pipeline wrongly discards or wrongly generates matters here, because the small-molecule pipeline is largely exhausted (Artificial intelligence-driven antimicrobial peptide discovery).

Research Program

The program has four parts, each adapted from a specific NLP methodology.

A compliance benchmark for protein conditioning. Translation bias was established with minimal pairs and challenge sets: unambiguous gender triggers across 20 languages, on which state-of-the-art systems fail to reach even 70% accuracy (Investigating Failures of Automatic Translation in the Case of Unambiguous Gender), and embedding-similarity networks showing that of 157 gender-neutral English words, Google Translate mapped 91 to masculine forms and only 32 to feminine ones (Identifying Gender Stereotypes and Biases in Automated Translation From English to Italian Using Similarity Networks). We build the protein analog: condition a generative model on an explicit, checkable constraint, hold everything else fixed, and flip only whether the constraint is congruent with the training prior. For a peptide generator retrained on DBAASP v3 (the database of validated antimicrobial peptide activities used by AMPGAN v3), the flip is target species and potency window — B. subtilis with a common MIC band versus carbapenem-resistant Enterobacterales with a rare band. For an antibody diffusion model of the DiffAb family, the flip is the developability condition and the antigen: standard viral antigens from SAbDab versus bacterial outer-membrane and conjugative-machinery targets. Compliance is objectively checkable (predicted activity within band, predicted hydropathy within range), so no human judgment is needed. We define the headline metric as the difference in compliance rates:

$$\Delta_C \;=\; \text{compliance}(\text{prior-congruent condition}) \;-\; \text{compliance}(\text{prior-incongruent condition})$$

This mirrors the gender bias gap $\Delta_G$ that A Tale of Pronouns reduced from 7.2 to 2.1 with four targeted examples. We also test a dose-response: as incongruence increases (rarer species, larger required deviation from corpus-average hydropathy), $\Delta_C$ should grow — the analog of the stereotype-incongruence degradation in Investigating Failures of Automatic Translation in the Case of Unambiguous Gender.

Attribution diagnosis. Using Integrated Gradients — a method that assigns each input token a score for how much it drove the output — we compute attribution of each conditioning token (species identifier, potency value, property prompt) to the generated sequence, the analog of the pronoun-to-profession attribution score $a_{\text{pron,prof}}$ in A Tale of Pronouns. The hypothesis is that noncompliant generations are preceded by suppressed attribution on the conflicting condition. This requires white-box models, which we have: AMPGAN v3, DiffAb, and p-IgGen (p-IgGen: A Paired Antibody Generative Language Model) are all open architectures; this constraint, noted for NLP in A Tale of Pronouns, simply rules out API-only models.

A watcher audit. Watching the Watchers: Exposing Gender Disparities in Machine Translation Quality Estimation showed that quality-estimation metrics err 1.5x to 8x more often on feminine referents, and that a standard score threshold retained 75% of masculine instances but only 63% of feminine ones — bias that propagates into training data. We replicate this design on developability evaluators. The test sets are matched minimal-edit groups: validated AMPs from DBAASP (known to be synthesizable and active, so a correct watcher should pass them) paired with canonical clinical-antibody-style sequences, plus charge-matched antibody binders to negatively charged bacterial surfaces versus oncology-target binders. We run the standard gatekeepers — TAP red flags, the sequence-based predictors of A Guided Design Framework, and the surface-based SurfProp evaluator (SurfProp: A surface-based property prediction framework for antibody developability and screening, which processes a full structure in roughly 270 ms) — and measure score disparities and retention at threshold. The prior here is explicit: SurfProp was trained only on antibody Fv fragments, and its high-concentration viscosity correlation already drops from $p = 0.84$ on the Apgar dataset to $p = 0.48$ on the Dai dataset, so these watchers are known to be fragile outside their training distribution. For peptide-specific formulation viscosity there is no reported result found on Lacuna for any AMP-trained evaluator, which is itself part of the gap.

Attribution-guided mitigation. Two interventions, both retraining-free. The first is hard-example demonstrations: select the four to eight conditioning cases with the most suppressed attribution and add them as in-context demonstrations or conditioning contexts, exactly the intervention that cut the translation bias gap from 7.2 to 2.1 while raising accuracy by 7.1 points in A Tale of Pronouns. The second is attribution-weighted steering: boost the logit contribution of the ignored conditioning channel during generation. Both are compared at matched generation budget against rejection sampling and against fixed property priors in the diffusion sampler.

Baselines

The "do nothing" reference for guided antibody generation is the unconditioned DiffAb baseline in Guiding diffusion models for antibody sequence and structure co-design with developability properties, which reports a predicted folding-stability change of $\Delta\Delta G = -0.99$ across 19 SAbDab CDR-H3 redesign targets; for the peptide regime it is AMPGAN v3 with its published species/MIC conditioning and PepCraft heuristic filtering, whose wet-lab validation found 2 of 5 candidates active with lead HY-P60322 at an MIC of 8 µg/mL against B. subtilis (Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3). These establish what current pipelines deliver without any new machinery.

The strongest existing competitor for constrained generation is the property-conditioned diffusion of Guiding diffusion models, which biases the diffusion prior toward hydrophilic residues (weighting $b = 0.8$) and scores intermediate denoising steps with a graph network, reaching a predicted $\Delta\Delta G$ of $-4.23$ on the same 19 targets. It is the strongest because it is the only published method that improves developability inside the sampler rather than by post-hoc filtering, and it preserved structural fidelity. We will run it with its published settings on our benchmark, including the bacterial-antigen and incongruent-condition splits it was never tested on, with the same generation budget, sampler step count, and tuning effort as our mitigation. Beating it means achieving comparable developability on the standard targets while achieving materially better compliance ($\Delta_C$ closer to zero) on prior-incongruent conditions, where its fixed hydropathy prior is not the relevant constraint. For the evaluator side, the strongest baseline is SurfProp-expt, which reached a Pearson correlation of $p = 0.86$ for hydrophobic interaction chromatography on the 137-antibody Jain clinical panel (SurfProp); the audit beats it not by a higher correlation on clinical antibodies but by demonstrating calibrated, disparity-free scoring on antimicrobial formats, which no current watcher has been shown to do.

The key ablation isolates the interpretability ingredient: attribution-selected demonstrations versus an equal number of randomly selected or easy demonstrations, and attribution-weighted steering versus uniform steering strength. If random demonstrations close $\Delta_C$ as well as attribution-selected ones, the diagnostic story collapses and the contribution reduces to "more demonstrations help," which would not be worth a paper.

Why not just…?

Why not just generate millions of candidates and filter harder? This is the rejection-sampling alternative, and it is a genuine threat: with SurfProp at roughly 270 ms per structure, filtering a large pool is feasible. Two pieces of evidence say it falls short. A Guided Design Framework shows that screening tools have limited utility for viscosity liabilities and rest on arbitrary cutoffs calibrated to clinical antibodies, so a filter that is itself biased cannot be fixed by scaling it — our watcher audit tests exactly this. And post-hoc optimization of surviving leads is known to disrupt binding, whereas conditioning inside the sampler achieved better stability without that cost (Guiding diffusion models). Still, at a large enough budget rejection sampling may match our aggregate compliance numbers; the decisive check is a head-to-head at matched generation budget, where steering should win because it spends samples only where the generator can be made to comply. If it does not win, we will say so and let the diagnosis and audit stand on their own.

Why not just fine-tune the generator on antimicrobial peptide data? Fine-tuning on DBAASP would realign the prior for covered conditions, and for those conditions a specialist model may well dominate — we acknowledge this and will run a fine-tuned control wherever data permits. But the review literature identifies data scarcity for properties beyond activity and hemolysis as the binding constraint (Artificial intelligence-driven antimicrobial peptide discovery), and the semantic-sink analysis of Identifying Gender Stereotypes suggests attractors persist after retraining: fine-tuning moves the default, it does not remove default dominance. New resistance conditions — a species outside the corpus, or the inducible resistance shift documented by Molecular antibiotic resistance mechanisms and co-transmission of mcr-9, where colistin exposure moves MICs from 4 to 128 mg/L — will always be incongruent with any fixed corpus, which is precisely the regime our compliance benchmark measures.

Why not just prompt a large instruction-tuned model? The evidence against this is direct. Large multilingual systems fail to reach 70% accuracy on explicit, unambiguous cues because priors override instructions (Investigating Failures of Automatic Translation); LLM evaluators show high tie rates, assigning identical scores to correct and incorrect outputs (Watching the Watchers); and instruction-following debiasing only emerged above 11 billion parameters in A Tale of Pronouns. That scale threshold is a real risk for our setting too, since protein design models are typically far smaller, and we will explicitly test compliance and steering efficacy as a function of model scale — if the threshold recurs, that is itself a finding that reshapes the mitigation design.

Why not just improve antibiotic stewardship instead of engineering biologics? Stewardship delays selection pressure but does not create new agents, while co-resistance now stacks carbapenem and last-resort polymyxin resistance on single conjugative plasmids (Molecular antibiotic resistance mechanisms and co-transmission of mcr-9) and carbapenem resistance nears 80% in A. baumannii (Hui Zhang et al.). The two strategies are complements; this proposal targets the manufacturability bottleneck that currently wastes the alternative-modality pipeline.

Decisive Checks

Four results decide the proposal. First, $\Delta_C$ must be nonzero and must grow with prior–condition incongruence across at least two model families; if compliance is uniform, the transfer claim is falsified and we report it. Second, attribution on conditioning tokens must predict noncompliance at the sample level, with a useful area under the curve — otherwise the diagnosis is decorative. Third, attribution-selected demonstrations or steering must close $\Delta_C$ more than matched random demonstrations and more than matched-budget rejection sampling, particularly on incongruent conditions where the strongest published baseline was never evaluated. Fourth, the watcher audit must show a retention disparity on matched sets analogous to the 75%-versus-63% split in Watching the Watchers, or the gatekeeper half of the claim fails. If all four hold, a final gate is a small synthesis batch of attribution-steered candidates measured by MIC against a carbapenem-resistant Enterobacterales panel, benchmarked against HY-P60322's 8 µg/mL, and assessed for the species selectivity that AMPGAN v3's conditioning promised but did not deliver.