acceptodds
Withdrawn from ICLR 2027

AMPLE: A Cost-Aware Adaptive Multi-Expert Framework for Antimicrobial Peptide Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Antimicrobial peptides (AMPs) are key components of innate immune defense and promising candidates for combating antimicrobial resistance, making accurate prediction of their diverse antibacterial, antifungal, antiviral, and antiparasitic activities important for identifying broad-spectrum anti-infective peptides. However, existing AMP prediction methods often rely on limited information sources or fixed strategies for integrating heterogeneous representations, limiting their ability to capture the distinct predictive characteristics of different types of biological information. To address this issue, we propose AMPLE, a cost-aware adaptive multi-expert learning framework for antimicrobial peptide activity prediction that models heterogeneous biological representations through specialized experts and selectively integrates their predictive capabilities within a unified architecture. AMPLE achieves state-of-the-art performance on the evaluated benchmark while substantially reducing unnecessary computation from expensive experts, demonstrating the effectiveness of adaptive expert acquisition for multilabel AMP activity prediction. Its design enables flexible performance–cost trade-offs across different computational budgets, providing a general framework for studying and modeling heterogeneous information in antimicrobial peptide prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antimicrobial peptides, Multilabel prediction, Multi-expert learning, Adaptive expert acquisition, Cost-aware routing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_r1y1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

AMPLE predicts five antimicrobial peptide activities using sequence, physicochemical, protein language model, and structural experts. A router selectively runs the expensive experts, and a separate gate combines their predictions.

Critical strengths
  1. AMPLE-Best matches the 76.1% mAP of the full four-expert model while reducing reported latency from 95.37 to 23.89 ms/sample.
  2. Adjustable routing thresholds allow a single trained model to support different computational budgets without retraining.
Critical weaknesses
  1. The architectural novelty modest. The framework combines existing experts with cost-aware routing and prediction fusion. The paper does not convincingly establish a methodological advance over existing approaches or demonstrate why a simpler fusion would be insufficient.
  2. Evaluation is limited to aggregate results on one benchmark, without concrete routing cases or detailed failure analysis.
Questions for authors
  1. What specific methodological contribution distinguishes AMPLE from existing active feature acquisition or dynamic modality selection approaches? Can the authors compare AMPLE with these methods under matched computational budgets?
  2. How does AMPLE handle missing modalities at inference time, such as unavailable structural information, when the corresponding expert cannot be executed?
  3. Why is the router trained without accounting for upstream structure-prediction costs? Would including ESMFold costs during training change the routing decisions?
  4. How were the Efficient and Best configurations chosen? Were they fixed before evaluating on the test set?
  5. The generalization ability of AMPLE remains unclear, as it is evaluated on only one benchmark. The authors should evaluate the model on additional AMP datasets to demonstrate its generalizability.
  6. The paper should provide representative successful and failed routing cases showing when acquiring an expensive expert helps or harms the prediction.
  7. The paper does not provide publicly available code or implementation details.
Reviewer_KkzGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AMPLE is a cost aware adaptive multi expert framework for multilabel AMP activity prediction over five labels. Four heterogeneous experts (lightweight sequence Transformer, physicochemical descriptors, frozen ESM2-650M, ESM-IF1 structure) are trained independently, frozen, and integrated at the prediction level. On top of this fixed pool sit two decoupled modules: an acquisition router that sequentially decides, at most one expert per step, whether another expert is worth its computational cost, and a label wise fusion gate over the acquired subset. The sequence and physicochemical experts are always evaluated, so routing applies only to the two expensive experts. Both modules are trained on cross fitted out of sample expert predictions, and per expert thresholds move a single trained model along a performance versus cost curve without retraining. On ESCAPE, AMPLE improves over the published baseline on both mAP and F1 at its higher operating point, while matching the predictive performance of evaluating all four experts at a fraction of the latency. A cheaper operating point of the same trained model still outperforms the standalone protein language model expert at a small fraction of its cost.

Critical strengths

The paper addresses a real and well posed question: heterogeneous biological representations differ by orders of magnitude in cost, and prior AMP work fixes the representation set at design time for every peptide. The proposed decoupling of acquisition from fusion is a sensible response to that setting, the framework is described precisely enough to be implemented, and the computational cost of every configuration is measured rather than estimated. The two main strengths are the computational cost analysis and the clarity with which the problem and the method are presented.

1. The computational cost analysis is thorough and well specified.

Table 1 gives per expert latency across eight peptide lengths with parallel execution measured rather than assumed, which preempts the obvious objection that the full computation baseline is a serial sum. The appendix then recomputes every configuration with ESMFold included, which raises the cost of structure involving configurations substantially, including the paper’s own. This level of detail is what makes the efficiency claims checkable, and it is the main reason I trust the reported numbers.

2. The problem is framed clearly and the proposed solution is explained precisely.

The distinction between deciding whether to compute an unobserved expensive representation and weighting it once observed is stated cleanly and motivated by the specific setting of independently trained, partly frozen, unevenly reliable experts. The reader always knows what is being claimed, which is also what makes the gaps below easy to identify.

Critical weaknesses

My central concern is that the experiments do not support the claims the paper makes. The questions posed in the introduction are about per sample and per label heterogeneity, but the evidence is reported almost entirely at aggregate level, the ablations that carry the main claims are sub point and untested, and the cost function the router optimizes omits the dominant cost component. The result is a framework that is plausible and cleanly presented, but whose specific contributions remain unestablished. The two main weaknesses are the absence of the analyses that would answer the paper’s own research questions, and the lack of evidence that the router and the fusion gate are responsible for the reported gains.

1. The paper’s own research questions are not answered at the resolution at which it poses them.

The introduction asks how the utility of different representations varies across samples and whether every peptide requires the same expensive representation. The evidence offered is aggregate: scalar invocation rates and leave one expert out frontiers. What is missing is the characterization that would answer the questions. There is no breakdown of which peptides take which path, by length, by label, or by composition. There are no per label results anywhere, despite the task being multilabel throughout and the fusion gate being explicitly label wise, so it is never shown that the learned label wise weights differ across labels at all. There is no incremental analysis of the sequential acquisition the method is built on: since Seq and Phys are always evaluated, the informative intermediate configurations are Seq plus Phys plus ESM2 and Seq plus Phys plus Struct, each isolating one optional expert on top of the fixed base. Table 2 reports only single experts and all four, so the reader cannot see what either acquisition step buys on its own, nor whether the two optional experts are complementary or redundant. With only two optional experts the full combination table is four rows, so there is no cost argument for omitting it.

2. The claims about the router and the fusion gate are not established.

Three distinct problems. First, the adaptive fusion claim rests on sub point differences (74.8 versus 74.4, 76.1 versus 75.6) with no significance test, no statement of what the reported standard deviations are computed over, and no sensitivity analysis over initialization, so it cannot be told whether these exceed run to run variation. Second, the learned cost aware router is never compared against the obvious non learned alternative: always run Seq and Phys, acquire ESM2 or Struct when fused uncertainty or Seq to Phys disagreement crosses a threshold. The router state in Equation 2 already contains exactly those statistics, and since the decision space is two binary choices controlled by two thresholds, such a rule has the same degrees of freedom as the learned policy. Random routing cannot stand in for it, because it discards the information the heuristic would use. Third, the cost term that trains the router is instantiated as inference latency conditional on the input representation being available, which excludes structure acquisition, the component the appendix then shows dominates. The appendix recomputes costs with the policy deliberately frozen, so the paper never shows what the router learns under the cost structure that applies at deployment. The framework is cost aware with respect to the wrong cost, and its conclusions are not shown to survive the right one.

Questions for authors

The questions below follow the two weaknesses. Items 1 to 3 ask for the per label and per subset analyses that would answer the paper’s own research questions, items 4 to 7 for the comparisons and tests that would show the router and the fusion gate are responsible for the reported gains, and the remainder are clarifications.

1. Please report per label AP for the five ESCAPE activities, and show that the learned label wise fusion weights differ across labels rather than collapsing to a shared weighting.

2. Please report the two intermediate configurations the acquisition process actually traverses, Seq plus Phys plus ESM2 and Seq plus Phys plus Struct, under both global and sample adaptive fusion, to isolate each optional expert and show whether the two carry complementary or overlapping information.

3. Across the ESCAPE sequences, what is the joint distribution of acquired expert subsets, and how does it vary by peptide length and by true label?

4. Please add a non learned confidence thresholded cascade to Figure 2a (acquire ESM2 or Struct when fused uncertainty or expert disagreement exceeds a swept threshold). How much of the gap between AMPLE and random routing does it close?

5. Are the ablation differences statistically significant? Please state what the reported standard deviations are computed over, and report sensitivity to initialization across multiple seeds for the router and fusion gate.

6. The FLOPs and invocation rates do not reconcile under uniform sampling: at the Best point, ESM2 accounts for roughly a quarter of its average cost despite a 60 percent invocation rate, and the gap is larger at the Efficient point. Does the router preferentially acquire ESM2 for short peptides because they are cheap? Please report invocation rate as a function of length and mAP stratified by length bin.

7. The router’s cost term excludes structure acquisition, which Table 4 shows dominates end to end cost. What policy does the router learn when trained with ESMFold inclusive costs, and does Struct get acquired at all?

8. What structural inputs were used for the ESCAPE peptides: experimental, predicted, or a mixture?

9. There are no experiments substituting ESM2 or ESM-IF1 with other, more recent representation models. How sensitive are the conclusions about representation utility to this particular choice of backbones?

10. At deployment, how is the operating point chosen in practice? The threshold values are not reported, and a practitioner needs to pick them for a given budget without retraining.

11. The abstract claims state of the art performance, but AMPLE-Best does not exceed the mAP of the paper’s own all expert sample adaptive fusion configuration. Would the authors reframe the claim around the performance versus cost frontier, which is what the evidence supports?

Reviewer_A4teOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper the authors propose AMPLE, a cost-aware adaptive multi-expert framework for multilabel antimicrobial peptide (AMP) prediction. The method is evaluated on the ESCAPE benchmark. The key insight is that prediction does not need to be multi-modal, that some modalities are wasted on samples, and so a dynamic method can determine based on an acquisition function which modalities to selected and then fuse for prediction among four possible ones: a lightweight sequence Transformer, a physicochemical descriptor-driven one, a frozen ESM2-650M protein language model-based one, and a structural expert (ESM-IF)-based one. An interesting feature of the proposed method is that AMPLE decouples the acquisition router from the adaptive fusion. The authors report that AMPLE achieves strong predictive performance (exemplified in 76.1% mAP) while reducing unnecessary computation compared to an evaluation that uses all four experts (and is competitive with other methods).

Critical strengths

(1) AMPLE selectively avoids expensive expert computations for a large portion of the samples. This is shown to significantly reduce average inference latency and FLOPs while maintaining competitive predictive accuracy on the ESCAPE benchmark.

(2) Decoupling the acquisition router from the post-acquisition fusion module allows AMPLE to separate the sequential choice of information gathering from the weighting of available prediction evidence.

(3) The authors evaluate AMPLE under multiple operational budgets, and the appendix also analyzes the impact of incorporating upstream structure-prediction costs (such as with ESMFold)

Critical weaknesses

(1) The authors define acquisition cost strictly as runtime inference latency and FLOPs during deployment. However, I am not convinced this is justified. For pre-trained models like ESM2-650M, the massive cost of pre-training has already been paid upfront. Extracting sequence embeddings for short peptides ought to be computationally lightweight, thus straining the premise of “expensive acquisition.”

(2) Relying on static, solution-state or predicted vacuum structures (via ESM-IF/ESMFold) for AMPs is fundamentally artificial. Antimicrobial peptides are known to be conformationally flexible and even intrinsically disordered at equilibrium in aqueous solution. Many undergo induced fit, changing their structures dynamically, for instance, upon membrane interaction. A static structure really fails to capture this functional reality.

(3) In fact, the structural expert clearly performs poorly on its own, achieving only 56.5% mAP. Because of this, the router learns to avoid it for roughly 90% of samples. So, the router is largely functioning as a noise filter for an uninformative, unreliable feature stream rather than optimizing true multi-modal utility, as claimed in the paper.

(4) The evaluation lacks systematic static baselines (such as fixed 2-of-4 or 3-of-4 combinations, like Sequence + ESM2 run uniformly across all peptides). Without comparing against a simple static multi-modal pipeline, it really remains unproven whether the complex, decoupled routing infrastructure provides any meaningful advantage over standard static integration.

(5) AMPLe introduces an intricate, decoupled routing and adaptive fusion infrastructure that seems to yield negligible performance gains over simple static models. Case in point from reported results: standalone ESM2 expert achieves 74.0% mAP, which is just 0.8% below full AMPLE-Efficient configuration (74.8% mAP). For further comparison, the lightweight sequence expert alone captures 70.9% mAP at a fraction of the computational overhead. This marginal improvement raises fundamental questions about whether the dynamic scheduling approach is addressing any real bottleneck or just adding unnecessary architectural complexity. To make the argument simpler: AMPLE seems to be an over-engineered solution in search of an actual problem.

Questions for authors

How can the use of static structure predictors (such as ESM-IF and ESMFold) be justified for antimicrobial peptides, given that AMPs lack stable equilibrium structures and alter their conformations dynamically depending on membrane composition and mechanisms of action?

How does AMPLE’s performance compare against a simple, static multi-modal pipeline (e.g., a fixed combination of the Sequence and ESM2 experts run uniformly on all peptides) that avoids the added complexity of a cross-fitted routing and adaptive fusion framework?

Given that the structure expert achieves a low standalone mAP of 56.5% and is suppressed for about 90% of peptides, does the router's utility stem from intelligent resource allocation, or is the router simply learning to discard a noisy and uninformative feature representation?

What is AMPLE really providing? For comparison, the lightweight Sequence expert alone achieves 70.9% mAP at just 2.90ms. This means ESM2 alone already captures the vast majority of the predictive performance, outperforming the full AMPLE-Efficient configuration (74.8% mAP) by only 0.8% while running on every sample.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.