Coarse composition suffices: tabular in-context learning for multi-activity antimicrobial peptide profiling

Raunak Kumar; Anuj Pal; Dhruvi Solanki; Parikshit Pareek; Juhi Singh; Jitin Singla

Source-linked AI summary

Coarse composition suffices: tabular in-context learning for multi-activity antimicrobial peptide profiling

Raunak Kumar, Anuj Pal, Dhruvi Solanki, Parikshit Pareek, Juhi Singh, Jitin Singla

arXiv:2608.30337v1Aug 31, 2026cs.LGq-bio.BM

TL;DR

AMPs commonly act across multiple pathogen classes, motivating multi-label prediction, while leading ESCAPE methods often depend on costly multimodal deep architectures. The paper instead combines interpretable sequence descriptors with TabPFN and reports state-of-the-art benchmark performance, while identifying label-structure and scope constraints.

Problem
AMP screening needs multi-activity profiles rather than single binary labels, but existing leading approaches commonly use costly structure- and language-model-conditioned architectures.
Method
The paper applies standard multi-label transformations to 330 interpretable sequence descriptors using TabPFN for single-pass in-context prediction without gradient updates or target-data hyperparameter search.
Results
77.8% mAP-5 versus 72.1% for the previous best method is achieved on ESCAPE, with gains persisting under matched single-fold evaluation and across descriptor ablations.
Takeaways & Limitations
Calibrated multi-activity probabilities support ranking which activity to assay next from partial positive evidence, while coarse sequence composition captures much of the benchmark signal.
Takeaways & Limitations
Label-powerset prediction cannot assign probability to profiles absent from its in-context samples, whereas classifier chains can assign mass to contradictory profiles.

Abstract

from arXiv · show

Antimicrobial peptides (AMPs) often act against multiple pathogen classes, making multi-label activity prediction a more realistic screening target than binary antimicrobial classification. The ESCAPE benchmark formalizes this setting, but leading approaches typically rely on multimodal, structure-conditioned deep models that are costly to train and tune. We show that a simple, sequence-only pipeline can match and surpass these methods by combining 330 interpretable sequence descriptors with TabPFN, a tabular foundation model that performs in-context prediction in a single forward pass without gradient-based training or hyperparameter search. On ESCAPE (82,359 peptides; five labels), a label-powerset TabPFN model achieves mAP-5 = 77.8%, improving on the previously best reported 72.1%. A probabilistic classifier chain is the first method to match or exceed the best published average precision on each of the five labels simultaneously. The gains persist under the prior state-of-the-art single-fold training protocol, indicating they are not a training-set-size artefact, and are largest for remote homologues (+11.2 points below 30% sequence identity). Ablations further show that predicted structure is unnecessary at inference and that performance is not driven by any single descriptor family: ten global physicochemical scalars recover 91% of full-feature performance. Finally, explicitly modelling label dependence yields targeted benefits for scarce activities and supports ranking which activity to assay next from partial positive evidence.

1 Introduction

AMP screening requires multi-activity profiles because individual peptides often affect multiple pathogen classes, while existing prediction has largely focused on binary classification. The paper tests whether a simple sequence-only TabPFN pipeline can outperform more elaborate multimodal approaches on ESCAPE.

A single AMP is frequently active against bacteria, fungi, viruses, and parasites, making multi-activity profiling more realistic than isolated binary activity prediction.
Computational AMP prediction has been dominated by binary classifiers, while newer multi-activity methods commonly use graph, transformer, or multimodal architectures.
The study combines 330 interpretable sequence descriptors with TabPFN for in-context prediction without gradient updates or target-data hyperparameter search.
77.8% mAP-5 versus 72.1% for the previous best method is achieved by label-powerset TabPFN on ESCAPE.
The gains persist under matched single-fold evaluation, are largest for remote homologues, and remain largely intact with ten global physicochemical scalars.Those scalars recover 91% of full-feature performance, and no single descriptor family explains the gains.

2 Materials and methods

The study evaluates multi-label transformations on ESCAPE using a fixed sequence-based representation, predefined split, and TabPFN backbone. Evaluation uses multiple macro-average scales, bootstrap uncertainty estimates, and a three-seed protocol without target-data hyperparameter tuning.

2.1 Dataset: ESCAPE contains 82,359 peptides with four activity labels and a deterministic Antimicrobial parent label; folds 1–2 provide 65,870 in-context samples and 16,489 peptides form the test fold.
2.1 Dataset: Antiparasitic is the rarest test label and constrains macro-averaged performance.
2.2 Evaluation metrics and statistics: The analysis reports mAP-5 across five labels, mAP-4 across four real activities, and mAP-3 across the three most populated non-parent activities.
2.2 Evaluation metrics and statistics: Uncertainty uses 2,000 bootstrap resamples, with scores averaged across three seeds and significance defined by two-sided 95% percentile intervals excluding zero.
2.3 Feature representation: Each peptide is represented by a 330-dimensional vector divided into eight interpretable descriptor families, identically applied across transforms.CTD descriptors separately encode composition, transitions between property classes, and positional distribution along the sequence.
2.4 Backbone: TabPFN uses a labelled reference set as context and returns calibrated test posteriors in one forward pass, without peptide-data weight updates or target-data hyperparameter tuning.

2.5 Multi-label transforms

The study compares standard multi-label transforms on a shared feature matrix and TabPFN backbone, differing in how they factorize activity profiles and model label dependence.

Five standard transforms were compared on an identical feature matrix, split, and backbone, differing only in multi-label target factorization.
Binary relevance fits one classifier per label and assumes conditional independence, providing the dependence-free comparison floor.BR costs L inference passes.
Label powerset treats each complete activity profile as one category, solving a single 16-class problem for the benchmark’s constrained profiles.Per-label probabilities are recovered by summing probabilities of profiles containing the label; LP uses one inference pass.
Probabilistic classifier chains predict labels in a fixed order, condition later labels on earlier ones, and evaluate all 32 possible profiles.PCC can represent profiles beyond those permitted by the deterministic Antimicrobial parent and therefore may return incoherent profiles.
Classifier chains follow one trajectory through the label sequence, reducing cost relative to PCC but allowing early prediction errors to cascade downstream.Hard context supplies binarized earlier predictions, whereas soft context supplies their raw probabilities.
Ensemble classifier chains average per-label probabilities across eight independently ordered chains, avoiding dependence on one arbitrary label ordering.ECC uses n_estimators = 8 to make the ensemble affordable.

2.6 ESCAPE baseline re-run and matched-protocol evaluation

The evaluation re-runs the ESCAPE baseline, matches its single-fold protocol, and tests structural-input dependence alongside descriptor sufficiency and conditional activity analyses.

ESCAPE baseline re-run and matched-protocol evaluation: Published ESCAPE baseline per-label values are unpaired point estimates, while released checkpoints were re-run without retraining for paired comparison.Structural-map analyses are restricted to 5,495 of 16,489 test peptides because maps are available for 33.3%.
ESCAPE baseline re-run and matched-protocol evaluation: The matched-protocol evaluation supplies each fold separately as context and averages the resulting probabilities, isolating the effect of additional in-context samples.
ESCAPE baseline re-run and matched-protocol evaluation: Structural-branch ablations replace each peptide’s structural input with zeros, noise, or another peptide’s map while holding sequence input fixed.Changes are measured using per-label average precision and absolute probability changes relative to real-map runs.
ESCAPE baseline re-run and matched-protocol evaluation: Dependence analyses exclude the deterministic Antimicrobial parent and compare oracle conditional predictability with held-out likelihood under BR and PCC.The joint-versus-product likelihood comparison tests whether activity dependence exists independently of ranking performance.
ESCAPE baseline re-run and matched-protocol evaluation: Descriptor-family analyses distinguish standalone sufficiency, performance loss when a family is unavailable, and reliance on shuffled test columns.All analyses share one stratified 4,000-peptide test subsample, whose absolute values are not comparable to Section 3.1.
ESCAPE baseline re-run and matched-protocol evaluation: Positive-unlabelled prioritization treats confirmed positives as reliable and zeros as unlabelled, ranking which remaining activity to assay next.LP and PCC joint posteriors are compared with BR using Hit@1, mean reciprocal rank, and a lower-bound pooled positive-unlabelled AP.

3 Results

Across ESCAPE evaluations, simple sequence-only TabPFN transforms outperform published methods, with gains persisting under matched protocols and increasing for remote homologues. Ablations indicate that coarse, redundant sequence composition and label dependence explain much of the performance and practical utility.

3.1 Improved average precision across all activity labels: 77.79 ± 0.23 mAP-5 for label powerset and 77.76 ± 0.12 for probabilistic classifier chains exceed the previous best 72.12.Binary relevance also reaches 77.01 ± 0.17, despite encoding no activity dependence.
3.1 Improved average precision across all activity labels: PCC is the first model to match or exceed the best published average precision on all five labels simultaneously.On Antiparasitic, PCC reaches parity at 37.68 versus 37.6, with the comparison limited by only 77 test positives.
3.2 Separating Data-Size Effects from Model Gains: LP retains a significant +3.69 [+1.04, +6.46] mAP-5 advantage under the matched single-fold protocol.The paired comparison uses faithfully reproduced ESCAPE checkpoints on the 5,495 test peptides with released structural maps.
3.3 Remote-homologue peptides show strongest gains: Below 30% sequence identity, LP gains +11.23 points over the previous state of the art, compared with +3.8 in the reported high-similarity comparison.Across the covered subset, LP’s pooled paired margin is +6.08 [+3.13, +9.10], making the remote-homologue gain close to twice the pooled margin.
3.4 Sequence-only information is sufficient: Ten global physicochemical scalars reach 73.52, recovering 91.4% of full-feature performance, while leave-one-family-out changes mAP-5 by at most 0.74 points.The results indicate broadly distributed, redundant signal rather than dependence on any single descriptor family.
3.5 Dependence helps mainly for scarce activities: Joint posteriors improve held-out label-vector log-likelihood by +0.0200 nats per peptide, but sibling-label headroom is limited and concentrated in scarce activities.The oracle lift is +9.66 for Antiparasitic versus +0.82 for Antibacterial, while PCC’s individually resolvable gain is +2.91 on Antiparasitic.

4 Discussion

The discussion argues that simple sequence descriptors with TabPFN can outperform more complex multimodal approaches while supporting activity-aware assay prioritization. Label dependence and coarse composition provide useful structure, but sparse labels, concealed-activity imbalance, and positive–unlabelled assumptions constrain interpretation.

Performance and efficiency: 77.79 mAP-5 makes label powerset the most accurate and cheapest transform across reported metrics.It requires one fit and inference pass, with 18 s fitting and 128 s inference, and is faster than competing transforms.
Performance and efficiency: Label powerset encodes activity co-occurrence as profile identity, preventing incoherent predictions because Antimicrobial is the logical OR of the four activities.This representation captures observed co-occurrence patterns exactly in the in-context samples.
What drives performance: Ten global physicochemical scalars recover 91.4% of full-feature performance, while removing structure at inference changes ensemble macro average precision by at most 0.056.For one released checkpoint, removing structure changes no predicted probability by more than 2.1 × 10^-5.
Assay prioritization: Assay prioritization treats unassayed activities as unknown and evaluates which remaining activity should be tested after confirmed positives.The evaluation reports Hit@1, mean reciprocal rank, and pooled positive-unlabelled average precision, the latter as a lower bound.
Assay prioritization: 70.89% of eligible peptides have the concealed activity ranked first at m = 0 versus 25% by chance, with ranking improving as positives are revealed.At m = 0, the result reflects sequence descriptors alone because no conditioning information is available.
Limitations: The main limitations are sparse antiparasitic evidence, split-based rather than cluster-held-out homology control, and lower-bound PU-AP under unlabelled zeros.The prioritization negative control also uses a single donor assignment rather than a permutation distribution.

5 Conclusion

A sequence-only TabPFN pipeline establishes a new state of the art on ESCAPE multi-activity AMP prediction. Its calibrated profile probabilities also support next-assay ranking, while the largest gains occur for remote homologues and inference requires no structural input.

5 Conclusion: 77.8% five-label macro AP versus 72.1% establishes a new ESCAPE state of the art using standard sequence descriptors.The method is also the first to meet or exceed the best published average precision on all five labels simultaneously.
5 Conclusion: The model requires neither task-specific training nor structural input, and gains over published methods are largest for remote homologues.Ablations suggest that benchmark signal is largely coarse composition.
5 Conclusion: Calibrated profile probabilities enable practical ranking of which activity to assay next.This extends the model’s use from per-label prediction to experimental prioritization.

S1 The multi-label transforms, with a worked example

The five transforms use the same descriptors and TabPFN backbone but differ in how the multi-label target is decomposed and whether label dependence is represented. Worked examples show how these choices affect marginals, joint profiles, and logical consistency.

The transforms differ only in how the five-label target is broken into TabPFN-solvable problems and how activity co-occurrence can be represented.
Binary relevance (BR): BR fits one independent classifier per label, recovering marginals but not the joint profile distribution.Its displayed factorization treats each label separately.
Binary relevance (BR): 0.276 versus 0.42: BR assigns the doubly-active profile probability 0.60 × 0.46 = 0.276, versus 0.42 for a joint model.Per-label average precision can remain competitive even while held-out joint log-likelihood is worse.
Label powerset (LP): LP relabels each observed profile as one multiclass class and recovers label marginals by summing profile probabilities.It reaches the example’s 0.46 antifungal marginal in one fit instead of a chain.
Label powerset (LP): LP’s output space is closed to profiles observed in context, but the benchmark’s 16 realizable profiles all occur and exclude illegal parent-label combinations.The Antimicrobial parent is computed as a deterministic OR of the four primary activities.
Classifier chains: PCC returns a full joint posterior by evaluating every profile, whereas CC follows one chain path and substitutes its own predictions at inference.CC’s committed path can discard probability mass and accumulate errors with chain depth; ECC averages across eight label orders without fixing that path commitment.
Probabilistic classifier chains (PCC): PCC can assign mass to contradictory profiles because its output space contains all 2^5 profiles, unlike LP’s constrained profile space.This mechanism produces the violation rates reported later.
Classifier chains: Hard versus soft chain context changes CC’s oracle-ceiling recovery from −40.8% to +13.7%.Hard context matches the binary labels available in the in-context samples, while soft context supplies raw probabilities unseen there.

S2 Per-seed results

The per-seed benchmark evaluates five transforms with the same TabPFN classifier, descriptors, context set, and held-out test set. It reports average-precision, threshold-based, exact-match, and timing measures with distinctions between optimistic and honest thresholding.

All five transforms use TabPFN v3 on the same 330 descriptors, with 65,870 in-context peptides and 16,489 held-out test peptides.Three seeds are reported individually with their mean ± standard deviation.
Average precision: mAP-5 macro-averages step-wise average precision over all five labels, while mAP-3 excludes Antiparasitic because the test set has only 77 positives.
Threshold metrics: Honest F1 freezes thresholds selected from five-fold out-of-fold in-context predictions, whereas Max-F1 uses test-tuned thresholds and is an optimistic upper bound.
Threshold metrics: Subset accuracy is exact-match performance for the five-label vector at a fixed 0.5 threshold and should be read as a conservative floor, especially for low-prevalence labels.
Wall-clock: TabPFN’s fit time is dominated by ingesting the context matrix, making it one to two orders of magnitude smaller than inference.Reported timings are wall-clock seconds on one GPU and include machine contention.

S3 Scaling with the number of in-context samples

Performance declines as the number of in-context samples is reduced, with similar degradation across BR, LP, and CC. The full-data configuration is reproducible within a small fixed-seed variation.

5.01 points: BR mAP-5 falls from 76.94 to 71.93 between the two largest context-size rungs.
4.71 points: LP mAP-5 falls from 77.79 to 73.07 between the two largest context-size rungs.
4.43 points: CC mAP-5 falls from 75.25 to 70.82 between the two largest context-size rungs.
The scaling sweep uses context sizes n ∈ {250, 1,250, 6,250, 31,250, 65,870} and a common held-out test set of 16,489 peptides.Reported S2 metrics are means over three seeds; mAP-5, Max-F1, and subset accuracy are in percentage points.
0.07 points: the independent 65,870-row rerun differs from the full-data configuration by at most 0.07 points at fixed seed.This bounds run-to-run variation for that configuration.

S4 Provenance of the published baselines

The published comparison retrained seven prior methods on ESCAPE and adapted them to five sigmoid outputs, but only two were originally designed for this joint multi-activity setting. Baseline results therefore require attention to each method’s original task scope.

Seven prior methods were retrained from scratch on ESCAPE’s training folds and adapted to emit five sigmoid outputs, avoiding cross-dataset leakage in the comparison.
Only two of the seven published methods were natively designed for multi-activity prediction over these labels.
A low frontier-table result should not be read as a verdict on a method’s original publication because the benchmark task may differ from its original target.
The original baselines covered different activities, with ✓ marking predicted activities and × marking activities they did not predict.Reported sample counts come from the original publications rather than the benchmark.
AB, AV, AF, AP, and AM denote antibacterial, antiviral, antifungal, antiparasitic, and antimicrobial, respectively.The original methods could address these as separate single-activity binary datasets rather than a joint multi-label head.