Antimicrobial peptides — short amino-acid chains that kill bacteria, often by disrupting the membrane — are one of the few remaining bright spots against drug-resistant infections, because bacteria evolve resistance to them more slowly than to single-enzyme-targeted small molecules (A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection). But machine learning has overpromised here. A recent audit, QMAP, shows that peptide-potency models have been graded on random data splits that let them memorize close sequence relatives; under honest splits that hold out non-similar peptides, deep models and simple linear ones differ by almost nothing, and on the peptides that matter most — the potent ones, with minimum inhibitory concentration (MIC, the lowest concentration that stops bacterial growth) below 10 µM — performance collapses toward chance (QMAP: A Benchmark for Standardized Evaluation of Antimicrobial Peptide MIC and Hemolytic Activity Regression). Meanwhile the wet lab cannot keep up: generative models emit thousands of candidates, and only a trickle ever gets synthesized and tested.
This proposal argues that the missing piece is an intermediate, mechanistic prediction target. Instead of mapping a peptide's sequence directly to a potency number, the model should first predict how the bacterium will respond — which stress pathways get perturbed — and only then predict potency. Functional enrichment analysis, the statistical machinery that aggregates noisy per-gene expression changes into coordinated pathway-level signals, provides exactly that target. The payoff is twofold: mechanistically similar peptides should share pathway fingerprints even when their sequences have diverged beyond what sequence models can exploit, and the intermediate prediction is cheap to check in the wet lab (one RNA-seq run per peptide) long before the expensive MIC and toxicity assays.
Core Claim
Direct sequence-to-MIC regression fails on non-homologous, high-potency peptides because the label carries no mechanistic structure: two membrane-disrupting peptides with 20% sequence identity look unrelated to a sequence model, yet they trigger the same envelope-stress response in the cell. I claim that a two-stage model — predict the bacterial pathway-response signature first, then the potency — will substantially outperform direct regression on QMAP's homology-aware splits, specifically on the MIC < 10 µM subset where existing models approach random performance, and that its predicted signatures will be verifiable against measured transcriptomes at low cost.
Formally, let $f$ be a peptide encoder (a transformer pretrained on peptide sequences, in the spirit of deepAMP's masked-language-model pretraining). Head $g$ predicts a signature vector $s \in \mathbb{R}^R$, one normalized enrichment score per regulon (a regulon is the set of genes a bacterial transcription factor switches on together), and head $h$ predicts $\log \mathrm{MIC}$:
$$\mathcal{L} = \big\| g(f(x)) - s \big\|_2^2 \;+\; \lambda \big( h(f(x)) - \log \mathrm{MIC} \big)^2$$
The enrichment head acts as a structured bottleneck: it forces the shared representation to encode mode of action, which is conserved across divergent sequences, rather than sequence surface features that do not transfer. Critically, at test time the model needs only the peptide — no transcriptome — so it can be evaluated on the full QMAP benchmark without any data the baselines lack.
Why This Is Not Already Done
Three lines of work come close and each stops one step short. Generative and predictive peptide models with real wet-lab validation — deepAMP (A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection) and AMPGAN v3 (Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3) — work purely on 1D sequence grammar, and AMPGAN v3 documents the failure this causes: conditioning on the target bacterial species did not produce species-selective killing in vitro. Conversely, transcriptome-conditioned screening models — MoAble's structure-to-signature coembedding (Predicting mechanism of action of novel compounds using compound structure and transcriptomic signature coembedding), CellVS-Net's cell-level network rewiring (Cell-Level Virtual Screening), and CURE's perturbation-conditioned diffusion (Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design) — prove the value of pathway-level targets but run exclusively on human cancer cell lines, evaluate only retrospectively, and never touch a bacterium. ARG-BERT classifies resistance mechanisms from protein sequences (Prediction of antibiotic resistance mechanisms using a protein language model) but is diagnostic, not predictive of new compounds. No existing system predicts a pathogen's pathway response to a candidate antimicrobial and uses it to rank or validate that candidate.
A second obstacle is infrastructure: standard enrichment knowledge bases were built for eukaryotes and cover prokaryotic stress networks sparsely, a limitation the enrichment-methods literature itself flags (Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings). So the proposal must build the gene-set library too.
Research Program
The program has three components, ordered so each de-risks the next. First, construct a prokaryotic regulon library: a curated set of roughly 50–100 bacterial stress-response gene sets — envelope stress, oxidative stress, stringent response, iron homeostasis, efflux regulons — assembled from the bacterial-regulation literature and encoded in a machine-readable form compatible with an enrichment pipeline such as funkea (funkea: Functional Enrichment Analysis in Python). Second, generate the paired dataset that currently does not exist: select 300–500 peptides from DBAASP v3 with measured MICs, spanning known mode-of-action classes, treat three organisms (a laboratory E. coli, K. pneumoniae, and a clinical carbapenem-resistant A. baumannii — carbapenem resistance in this pathogen approaches 80% in surveillance networks (Hui Zhang et al.)) at sub-inhibitory concentrations, and collect bulk RNA-seq per peptide, converting each profile into a regulon-enrichment signature. Third, train the two-stage model on the paired data plus DBAASP potency labels, and validate along a staged ladder that directly attacks the in silico validation gap: predicted signatures are checked against RNA-seq of held-out peptides (cheap, mechanistic), and only the top-ranked novel candidates proceed to synthesis, broth-dilution MIC assays against multidrug-resistant isolates — including strains carrying stacked resistance plasmids such as mcr-9 co-localized with metallo-β-lactamase genes (Jie Li) — and hemolytic-toxicity ($HC_{50}$) testing. The staged design means a wrong model fails visibly at the RNA-seq step instead of wasting synthesis and assay budgets.
Proposed Experiments
Experiment 1 is retrospective and decisive for the generalization claim: evaluate on QMAP's five homology-aware splits (≤ 60% sequence identity, enforced by the benchmark's Leiden-based partitioning via the qmap-benchmark package), reporting MIC regression error overall and stratified on the MIC < 10 µM subset, plus $HC_{50}$ regression. Success means materially lower error than every sequence-only baseline on the low-potency subset — where QMAP reports existing models "approach random performance" — with no degradation overall. Experiment 2 tests signature learnability: for held-out peptides, measure per-regulon correlation between predicted and measured enrichment; MoAble's human-cell analogue achieved a median gene-set-enrichment pathway recovery $p$-value of 0.03 versus 0.381 for a raw-expression baseline (Predicting mechanism of action of novel compounds using compound structure and transcriptomic signature coembedding), which sets the reference point for what a useful signature-recovery signal looks like, though our bacterial setting is harder and not directly comparable. Experiment 3 is prospective: synthesize 15–30 novel candidates ranked by the model among sequences non-homologous to training data, and measure confirmed-active rate and MIC. Existing prospective reference points are AMPGAN v3's two of five synthesized candidates active, best MIC 8 µg/mL against B. subtilis (Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3) — noting that organism is non-pathogenic and the sample tiny — and deepAMP, where over 90% of designed peptides beat the penetratin template, the lead T2-9 matched FDA-approved antibiotics, and three peptides held stable efficacy with no resistance emergence over 25 days in S. aureus (A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection). I will include the serial-passage resistance assay so results are comparable to that 25-day stability benchmark.
Baselines
The naive reference point is a constant predictor that outputs the training-set mean MIC on each QMAP split; on the hardest subset this is nearly the state of the art, since QMAP reports the strongest existing models approach random performance there. The strongest competitor is sequence-only deep regression as configured in QMAP — BERT-AmPEP60 and the models of J. Witten et al., all evaluated under identical homology-aware splits. No reported result found on Lacuna for their exact numeric QMAP metric values; the benchmark paper reports qualitatively that deep and linear models show "only modest improvements over each other" and near-random error on MIC < 10 µM peptides. I will re-run these through the qmap-benchmark package on the same five splits with the same tuning budget (nested search over the same hyperparameter ranges and equal compute), and also run deepAMP's prediction classifier (its SVM trained on GRAMPA) and the ESCAPE dual-branch transformer (A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification) adapted to regression, since its multimodal sequence-plus-structure design reported a 2.56% relative mean-average-precision gain over the second-best method on its own benchmark and represents the strongest published peptide representation. For the prospective stage, the comparator is a sequence-only prioritizer of identical architecture and generation budget, so any hit-rate difference isolates the enrichment signal. Finally, the key ablation: the identical two-stage architecture with the enrichment target replaced by an equal-sized library of random gene sets (and, separately, by a generic reconstruction loss of matched dimension). Beating the baselines means at least halving low-subset MIC error relative to the best sequence-only model, achieving positive per-regulon correlation between predicted and measured signatures, and a higher confirmed-active rate than sequence-only ranking at a fixed synthesis budget. If the random-gene-set ablation matches the curated-regulon model, the claim is falsified regardless of headline numbers.
Why not just…?
Why not just scale up the sequence model? More parameters and more sequences do not add mechanistic information, and the evidence says capacity is not the binding constraint: QMAP found deep transformers and ridge regression nearly indistinguishable under honest splits, with degradation concentrated exactly where extra capacity should help most — the high-potency, low-homology peptides (QMAP). The matched-capacity ablation above settles this directly: if the gain survives replacing pathway targets with random gene sets of the same size, then it was never about the mechanism, and the proposal's claim dies by its own experiment.
Why not just retrieve nearest neighbors from DBAASP? Similarity search works only while sequences stay recognizable. ARG-BERT showed alignment-based matching already slips at 40% identity even for conserved resistance proteins (BLAST 0.917 accuracy versus the language model's 0.927, Prediction of antibiotic resistance mechanisms using a protein language model), and QMAP's splits are constructed precisely to remove those neighbors, so retrieval degenerates toward random there. More damning, AMPGAN v3 conditioned generation on target species and still failed to achieve selective killing in vitro (Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3) — sequence similarity carries no mechanistic guarantee, which is the gap the pathway signature is designed to close.
Why not just rank the top differentially expressed genes instead of doing enrichment? Single genes are noisy and threshold-dependent — the enrichment literature calls this the curse of the gene list, since coordinated but individually subtle shifts get discarded by per-gene cutoffs (Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings) — and structure demonstrably beats raw expression as a representation: CellVS-Net's gene-dependency networks outperformed molecular and expression-snapshot baselines on drug-disease retrieval (Hits@1 of 0.1250 versus 0.0357 for fingerprints and 0.0179 for random) and cut network prediction error by 12–47% across perturbation types (Cell-Level Virtual Screening). A raw-DEG variant of my model will be run as an additional ablation, so this alternative is answered empirically rather than by argument.
Decisive Checks
Three results would each independently falsify the core claim, and the program is built to surface them early. If the matched-capacity ablation with random gene sets matches the curated regulon library, the improvement is an artifact of extra supervision, not pathway structure. If predicted signatures fail to correlate with measured RNA-seq on held-out peptides, the intermediate target is unlearnable at this data scale, and the paired dataset should be expanded before any synthesis is attempted. If QMAP gains do not carry into the prospective hit rate, then the benchmark's low-homology subset is not representative of real screening, and the field's problem is the evaluation, not the models — itself a publishable and consequential finding.
References
- A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection
- QMAP: A Benchmark for Standardized Evaluation of Antimicrobial Peptide MIC and Hemolytic Activity Regression
- Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3
- Predicting mechanism of action of novel compounds using compound structure and transcriptomic signature coembedding
- Cell-Level Virtual Screening
- Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design
- Prediction of antibiotic resistance mechanisms using a protein language model
- Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings
- funkea: Functional Enrichment Analysis in Python
- Antimicrobial Activity of Colistin Against Contemporary (2015 – 2017) P. aeruginosa and A. baumannii Isolates From a Chinese Surveillance Program
- Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex
- A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification