Do Antimicrobial Peptide Predictors Have a Pathogen Fairness Problem?

Antimicrobial peptides (AMPs) — short immune-defense sequences that kill bacteria by physically disrupting their membranes rather than by blocking a single enzyme — are one of the few credible answers to the antibiotic-resistance crisis, which is projected to cause up to 39 million deaths by 2050 A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification. Deep-learning predictors now screen and generate candidate peptides at scale, but they are trained on databases whose activity labels are overwhelmingly concentrated on a few laboratory strains of E. coli and S. aureus, with rare pathogens and rare functional classes reduced to tiny, data-sparse slivers. The field's implicit assumption is that more pooled data helps everyone.

A second, apparently unrelated literature has already run this experiment — on voices instead of peptides. Speech models trained on male-dominated broadcast archives developed quantified, mechanistically characterized performance gaps for female, child, and non-binary speakers. Three regularities emerged: a "performance lottery" in which the advantaged subgroup flips unpredictably across languages and architectures, with non-binary speakers facing error gaps up to 52% in some languages Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps; an "unbalanced mixture pitfall" in which naively pooling majority and minority data makes the minority worse than a dedicated minority-only model, moving children's word error rate from 14.2% to 20.6% Speaker adaptive acoustic modeling with mixture of adult and children's speech; and the counterintuitive finding that fairness improves when internal representations keep subgroup information separable — a strong negative correlation ($\rho = -0.65$) between how well a linear probe can read gender off the model's hidden states and how large the gender error gap is Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps. All three are, mathematically, statements about what empirical risk minimization does to sparse regions of a latent space — nothing in them is acoustic.

Core Claim

I propose to test whether these regularities are domain-general by treating pathogen species as the "demographic subgroups" of antimicrobial peptide prediction. The claim is twofold. First, protein language models (pLMs) trained on pathogen-skewed AMP data should exhibit a pathogen performance lottery, a mixture pitfall on rare pathogens, and an extractability–error anti-correlation, because all three are consequences of sparse versus dense latent regions under skewed training pools, not of audio. Second, if that holds, the repair is not more data but subgroup-aware representations: an auxiliary training objective that keeps pathogen identity linearly decodable from the hidden states, plus per-pathogen affine normalization of embeddings — the direct analogue of the vocal-tract-length and constrained-MLLR transformations that turned the speech pitfall around Speaker adaptive acoustic modeling with mixture of adult and children's speech. This is a representation-level intervention, deliberately different from the data-level fixes that dominate AMP work.

The closest prior work sits on both sides of the divide and stops short of this. The speech study that motivates everything here Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps never leaves audio; it does not ask whether pLMs like the ProtBERT backbone of BERT-AmPEP60 or the ESM-2 discriminator of A Multi‐Property Optimizing Generative Adversarial Network for de novo Antimicrobial Peptide Design have the same latent pathology on rare classes. On the peptide side, A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification documents that performance collapses on the rare antiparasitic class and partially recovers with a dual-branch sequence–structure model, but never probes hidden states to ask why, and Controllable Generation of Pathogen‐Specific Antimicrobial Peptides Through Knowledge‐Aware Prompt Diffusion Model solves pathogen imbalance with knowledge-graph conditioning of a generator — a data-side conditioning mechanism whose effect on the shared representation geometry it does not measure, and whose analogue in the speech world is known to sometimes be a placebo Cultural Conditioning or Placebo? On the Effectiveness of Socio-Demographic Prompting. Nobody has audited peptide models with the probing protocol, nor coupled the audit to a representation-level repair. That coupling is the proposal.

Research Program

The program has three phases, each with a falsifiable output. Phase one is the audit. I will train a ProtBERT- or ESM-2-backed regression head on minimum inhibitory concentration (MIC, the standard lab measure of how much peptide is needed to stop growth) using DBAASP v3, filtered exactly as in the two existing MIC papers: the 3,929-entry E. coli set of MBC-Attention, the 4,042 E. coli and 3,275 S. aureus entries of BERT-AmPEP60, and pathogen-stratified splits over the 56 standardized gcPathogen types used by KPPepGen, plus the ESCAPE multilabel benchmark of 21,409 AMPs and 60,950 UniProt-derived negatives A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification. Following the speech protocol layer by layer, I will fit linear probes that predict the pathogen label from the hidden states and measure the correlation between probing accuracy and the per-pathogen prediction error gap. The hypothesis, imported from the $\rho = -0.65$ result, is a similarly strong negative correlation; the null — that peptide representations show no such structure — would itself be an informative boundary condition on how far the fairness results generalize.

Phase two is the controlled pitfall experiment. For each rare pathogen with enough data to split, I will train models on mixtures that vary the ratio of majority (E. coli/S. aureus) to minority-pathogen data, and compare minority-pathogen error against a dedicated minority-only model at each ratio. The speech literature predicts a U-shaped or monotone-worse curve — the unbalanced mixture pitfall — contradicting the peptide-field assumption that pooled data always helps. I will also test the "lottery": whether which pathogen is best predicted flips across backbones (ProtBERT vs. ESM-2) and tasks (MIC regression vs. multilabel classification), as the advantaged gender flips across Whisper and SeamlessM4T Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps.

Phase three is the repair. The model trains on the standard MIC or multilabel objective plus an auxiliary loss that keeps pathogen identity linearly separable in the hidden representation, plus a learned per-pathogen affine transformation of the embedding (mean-and-covariance normalization, the direct analogue of the constrained-MLLR speaker adaptation that helped most in speech). Concretely, for hidden state $h$, task loss $L_{\text{task}}$, and a small linear probe $g$ predicting the pathogen label $y_p$:

$$L = L_{\text{task}}(f(A_{y_p} h)) + \lambda \, L_{\text{CE}}(g(h), y_p)$$

where $A_{y_p}$ is the affine map for the true pathogen and $\lambda$ is a single tuned weight. The point of the auxiliary term is not to predict pathogens for their own sake but to force the representation to spend capacity separating subgroups — exactly what the speech evidence says fair models do internally, and what "gender-blind" training destroys.

Baselines

The do-nothing reference point is the pooled model that current practice already builds: a pre-trained pLM backbone with a regression or classification head trained on all pathogens mixed together, the design of BERT-AmPEP60. This is the status quo the pitfall experiment is expected to indict, so it anchors everything: it gets the same backbone, splits, and tuning budget as every other arm, and its per-pathogen error profile is the gap we try to close.

The strongest existing method for MIC regression is MBC-Attention, the multi-branch CNN with local and cross-branch attention of A deep learning method for predicting the minimum inhibitory concentration of antimicrobial peptides against <i>Escherichia coli</i> using Multi-Branch-CNN and Attention, which reports a Pearson correlation of 0.775 and an RMSE of 0.533 log µM on the 3,929-entry DBAASP E. coli test set — the best single-dataset number in the literature I have. For the multilabel setting, the strongest competitor is the dual-branch sequence–structure transformer of A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification, which reports a 2.56% relative mean-average-precision improvement over the second-best of eight classifiers and a 35.7% relative average-precision gain on the rare antiparasitic class. And for pathogen-conditioned generation, KPPepGen reports top performance on at least 49 of 56 pathogen tasks and more than 10% improvement on data-scarce pathogens. To compare fairly, I will retrain or reimplement each of these on my identical splits, sequence lengths, and compute budget, with the same hyperparameter search width, because the original evaluations use different data slices; where my numbers diverge from the published ones, I will report both. Beating the strongest baseline means improving worst-pathogen Pearson correlation or average precision by a margin that survives seed-level significance tests, without degrading majority-pathogen performance below the pooled reference — closing the gap, not trading one subgroup for another.

The key ablation is the same architecture with $\lambda = 0$: the auxiliary pathogen-probe loss removed but the per-pathogen affine maps kept, and vice versa. This isolates which ingredient carries any gain. Two further arms serve as baselines rather than ablations: naive duplication of minority-pathogen training examples to parity, and SMOTE-style interpolation in the latent space, which leveled male/female word error rates in Portuguese ASR where duplication failed The Balancing Act: Unmasking and Alleviating ASR Biases in Portuguese. Both get the same tuning budget, so that if the simpler data-level fix works as well, it wins — as it should.

Why not just…?

Why not just condition on pathogen prompts, like KPPepGen already does? This is the real threat, and I will not argue around it: KPPepGen already handles 56-pathogen imbalance and reports gains on data-scarce pathogens. But it changes the conditioning of a generator, not the geometry of the shared representation, and its evaluation reports aggregate per-task performance rather than worst-group gaps or probing structure — precisely the diagnostics the speech work says matter. There is also direct evidence in the neighbouring literature that demographic-style conditioning can be a placebo Cultural Conditioning or Placebo? On the Effectiveness of Socio-Demographic Prompting. So the proposal includes a KPPepGen-style prompting arm trained on the same splits: if pathogen prompting closes the worst-pathogen gap as well as extractability regularization, that result settles the question in its favor and I will say so.

Why not just train a separate model per rare pathogen? Because the speech data quantifies why this loses: the dedicated child model reached 14.2% WER, but the adaptive joint model reached 10.2% on children while also improving adults to 8.2% Speaker adaptive acoustic modeling with mixture of adult and children's speech — shared signal is real and throwing it away is costly, especially when some of the 56 pathogen types have too few sequences to train anything alone. The peptide side reinforces this: transfer from large pre-trained models was shown to be critical precisely when peptide training sets are small BERT-AmPEP60. I will nonetheless include per-pathogen heads as a baseline arm, since where data exists they may be surprisingly strong.

Why not just oversample the minority (duplication or SMOTE)? Naive duplication is the cheapest fix a skeptic will name, and the speech evidence is directly against it: duplication failed or increased variance, while SMOTE in feature space worked The Balancing Act: Unmasking and Alleviating ASR Biases in Portuguese. But SMOTE-in-latent-space is a genuine competitor to my auxiliary loss, and it is included as a baseline with the same budget rather than dismissed. There is also a peptide-specific caution: interpolated latent points may decode to sequences violating biochemical validity, the same worry that led OmegAMP to filter its 100,000 synthetic negatives with hard classifiers rather than trusting interpolation OmegAMP: Targeted AMP Discovery via Biologically Informed Generation, and synthetic data is known to inherit and compound the biases of its generator in the speech setting Corpus Generation for Voice Command in Smart Home and the Effect of Speech Synthesis on End-to-End SLU. The experiments decide: if latent SMOTE matches extractability regularization on worst-pathogen error, the simpler method wins.

Proposed Experiments and Decisive Checks

The decisive checks are, in order. First, the replication: does the probing-extractability versus per-pathogen-error correlation appear at all in pLMs, with the layer-wise "twist, hump, plateau" profile, or does peptide representation behave differently? No published result found on Lacuna for probing protein language models on pathogen subgroups — that absence is the point. Second, the pitfall curve: does pooled training underperform a dedicated model on at least some rare pathogens, contra the field's assumption? Third, the repair: does the auxiliary-loss model reduce worst-pathogen error relative to the pooled BERT-AmPEP60-style baseline and to MBC-Attention (published reference: PCC 0.775 on the E. coli test set), and does the correlation between extractability and gap tighten — the mechanistic signature that the fix works through the hypothesized channel rather than by luck? Success thresholds are pre-registered per pathogen: a statistically significant worst-group improvement with no significant majority-group regression, across at least five seeds and both backbones.

Two honest risks shape the design. DBAASP's MIC values are noisy across assay conditions — differing media, temperature, and salt concentrations — as both MIC papers acknowledge BERT-AmPEP60, so I will condition on or stratify by assay metadata where it exists and report sensitivity analyses without it. And pathogen labels themselves are incomplete, the biological twin of the metadata scarcity that plagues demographic fairness work FaiST: A Benchmark Dataset for Fairness in Speech Technology; unlabeled rare-pathogen entries will be tracked as a separate evaluation stratum rather than silently pooled. If the audit phase returns nulls across the board, the deliverable becomes a clean boundary condition: the fairness regularities of speech do not transfer to sequence models, which would sharpen both literatures. But if they transfer even partially, antimicrobial peptide prediction gains a diagnostic vocabulary and a repair kit for exactly the rare-pathogen, rare-class failures that currently limit its usefulness against resistant infection — and the speech-fairness community gains evidence that its core findings are about imbalanced representation learning, not about sound.