acceptodds
Under review as a conference paper at ICLR 2027

AlloMIT: Boosting Allosteric Virtual Screening via Multi-Instance Parameter-Efficient Tuning

Abstract

Allosteric modulation offers opportunities to alleviate target crowding and broaden the scope of drug discovery by targeting sites beyond canonical active sites. However, mainstream virtual screening methods largely focus on orthosteric screening, where a binding pocket is specified in advance. In contrast, allosteric sites may be unknown and require ligand-induced conformational changes, making allosteric screening more challenging. Moreover, the limited availability of allosteric activity data poses an additional challenge for training effective screening models. To this end, we introduce AlloMIT, a multi-instance parameter-efficient tuning strategy for allosteric hit discovery under limited activity data. AlloMIT integrates with pre-trained virtual screening alignment models in a plug-and-play manner, adapting them with only a small number of trainable parameters. Rather than relying on a single pre-defined binding pocket, AlloMIT represents each protein structure as a set of candidate pockets and learns target-ligand interactions in a weakly supervised, multi-instance manner, without requiring pocket-level annotations. To account for ligand-dependent pocket contributions, its Contextual Sparse Pocket Aggregation (CSPA) module considers potential pocket-ligand interactions within proteins and outputs the final pocket score via sparse aggregation. Additionally, we construct AlloVSBench to support parameter-efficient model adaptation and benchmark allosteric hit discovery performance across screening methods and parameter-efficient tuning methods. Ultimately, AlloMIT achieves the highest allosteric BEDROC score among all evaluated methods. Compared with adapted HypSeek, AlloMIT improves BEDROC from 17.09% to 23.25% and lifts EF₀.₅% from 20.79 to 41.70.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
allosteric virtual screening, multiple-instance learning, parameter-efficient fine-tuning, protein-ligand interaction, binding-site uncertainty

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_d4SbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose ALLOMIT, which adapts pretrained virtual-screening methods on allosteric activity data. It first uses FPocket to identify several pockets and encodes each with a pocket encoder. Then it uses the CSPA module to sparsely aggregate the pocket embeddings to obtain a final score. The authors also construct AlloVSBench for training and evaluation. After fine-tuning, AlloMIT outperforms both the pretrained HypSeek and fine-tuned HypSeek.

Critical strengths

The experiments include backbone comparisons including LigUnityLigUnity, DrugCLIPDrugCLIP, HypSeekHypSeek, HypSeekpocHypSeek_{poc}, HypSeekseqHypSeek_{seq}, and their fine-tuned versions. It gives fair comparisons and proves that AlloMIT is better at leveraging limited activity data than the baselines. They also provide detailed ablations on the architectural choices.

Critical weaknesses
  1. The supervision does not identify the correct pocket and an incorrect pocket assignment can satisfy the training objective. The authors did not prove that successful retrieval depends on the experimentally supported pocket. The authors did not explain the motivation behind the designs behind section 2.3. The authors report sequence-only results, but do not establish why adding potentially unreliable pocket candidates improves over a strong ESM-based sequence-only model.
  2. The allosteric test set is extremely small with only 25 allosteric positive pairs across 19 targets. Also, separating exact UniProt identities does not establish that the test pockets are structurally unfamiliar. Since reference structures and sites are available, the authors should measure local pocket sequence/structural similarity and ligand similarity against both adaptation data and relevant supervised backbone training data.
Questions for authors
  1. Does the method depend on the correct site?
  2. How were checkpoints selected, and how stable are the results?
  3. Could the authors provide a complete reproducibility package?
  4. Could the authors perform more case studies as in 4.5. For example, 8QYQ (chain A, H), 8U6B (chain A, C)?
5 comments
AuthorsOpenReview ↗

Response to Reviewer d4Sb

Thank you for your insightful comments. We have clarified the role of candidate-pocket coverage and checkpoint selection, provided stability analyses and reproducibility information, and extended the case studies to the suggested targets. To make our revisions and supporting data easier to inspect, we provide an anonymous repository here.


Q1. Does the method depend on the correct site?

A1. AlloMIT does not require the correct allosteric site to be specified in advance. It generates a bag of candidate pockets and scores each ligand against these candidates. Moreover, the table below suggests that screening performance is more closely related to target coverage than to precise pocket–site matching.

KTarget coverage (%)BEDROC (%)\mathrm{EF}_{0.5\\%}
532.117.2910.47
1050.017.8220.79
2075.023.0941.70
30 (Default)82.123.2541.70
4085.722.4041.70

Q2. How were checkpoints selected, and how stable are the results?

A2. Given the scarcity of allosteric adaptation data and the absence of experimentally verified inactive compounds in the training set, we did not hold out a separate validation set and instead selected AlloMIT checkpoints using training loss. As shown in Figure 1, we selected the checkpoint with the lowest training loss independently for each stage. Test-set performance was strictly excluded from both checkpoint selection decisions.

We assessed result stability across random seeds and sensitivity to the inclusion of individual test targets. Across three random seeds, allosteric BEDROC was 23.19% ± 0.11%. LOO denotes leave-one-target-out analysis, where each test target is removed in turn and the mean gain is recomputed on the remaining targets. As shown in the table below, AlloMIT retained positive mean gains over HypSeek for both evaluation panels and both metrics.

PanelMetricMean gainMinimum gain after LOO
Allosteric-hitBEDROC+8.31 pp+4.97 pp
Allosteric-hit\mathrm{EF}_{0.5\\%}+20.80+10.91
OverallBEDROC+2.41 pp+1.41 pp
Overall\mathrm{EF}_{0.5\\%}+1.47+0.44

Q3. Could the authors provide a complete reproducibility package?

A3. To facilitate reproducibility, we provide an anonymous repository containing the code and usage instructions. The training data will be publicly released upon acceptance of the paper.


Q4. Could the authors perform more case studies as in 4.5. For example, 8QYQ (chain A, H), 8U6B (chain A, C)?

A4. Thank you for suggesting these cases. We extended the case study to 8QYQ and 8U6B and report them alongside AKT1. None of these three cases was included in the training set. We evaluated these targets using the screening library from Section 4.5 and a commercial screening library containing 20 million compound records. Across all three targets and both screening libraries, AlloMIT consistently ranks the allosteric ligands ahead of HypSeek.

Screening library from Section 4.5

TargetHypSeek rank (top %)AlloMIT rank (top %)
AKT11,490 (8.30%)294 (1.64%)
8QYQ3,822 (21.28%)740 (4.12%)
8U6B66 (0.368%)42 (0.234%)

Commercial screening library

TargetHypSeek rank (top %)AlloMIT rank (top %)
AKT189,901 (0.4442%)654 (0.003232%)
8QYQ1,633,510–1,633,512 (8.072%)42,720 (0.2111%)
8U6B617 (0.003049%)33 (0.0001631%)

Q1 (follow-up). The answer restates that the allosteric site is not an input; that was not the question. We asked whether the pocket that drives a ligand's score corresponds to the experimentally supported site, and whether performance survives when it does not. The K-versus-coverage table cannot resolve this, because every bag contains both correct and incorrect pockets, so coverage and pocket correctness are confounded. Please report, per test target, whether the top-scoring pocket or the pocket producing the best ligand rank overlaps the annotated site, and compare performance when scoring is restricted to the annotated pocket versus randomly selected pockets of matched size.

Q4 (follow-up). We did not ask for library-scale screening on these two structures. 8QYQ and 8U6B were suggested because each contains a defined allosteric ligand bound at a defined site. Please test that pair directly. Report (i) the score and rank of that ligand across the candidate pockets, (ii) the CSPA weight per pocket in the style of Figure 9, and (iii) whether the highest-weighted pocket corresponds to the crystallographic binding site, with contact residues compared. Decoys should come from a fixed, size-matched set rather than a 20-million-compound library.

Reviewer_LNoROpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses allosteric virtual screening in the setting where the binding pocket is unknown and allosteric activity data are limited. It adapts a pretrained protein–ligand screening model (primarily HypSeek) by representing each target as a set of candidate pockets rather than a single predefined pocket. Pocket–ligand matching scores are combined by a Contextual Sparse Pocket Aggregation (CSPA) module that produces sparse, ligand-dependent contributions, and only a small number of parameters are updated via a two-stage (warm-up then joint) tuning procedure. The authors introduce AlloVSBench, a target-disjoint benchmark constructed from allosteric activity data paired with apo or AlphaFold structures. On this benchmark AlloMIT reports higher early enrichment of allosteric hits than the tuned baselines and also improves general screening performance relative to the adapted HypSeek backbone.

Critical strengths
  1. The problem is well motivated. Allosteric sites are frequently unknown and may differ from canonical orthosteric sites, so methods that require a single predefined pocket face a practical limitation.
  2. The multi-instance formulation is natural. Treating candidate pockets as competing hypotheses and aggregating their ligand-specific evidence (rather than committing early to one uncertain site) is a sensible response to the uncertainty of pocket identification; the connection to multi-instance learning is clearly articulated.
  3. The adaptation is lightweight and the method is tested with more than one pretrained backbone. Ablations further show that first training the newly introduced modules (residual MLP and CSPA) before joint tuning of selected encoder parameters is important for the final performance.
Critical weaknesses
  1. The model-selection protocol is unclear. For AlloMIT the paper states that the checkpoint with the lowest training loss is retained after each of the two stages; it is not clear whether the best PEFT configuration reported for each baseline was selected on held-out validation data or after inspection of test results. This ambiguity can render the comparisons optimistic or inconsistent.
  2. The reported gain is not cleanly separated from extra training and capacity. AlloMIT introduces candidate-pocket representations, a residual calibration module, CSPA, and a two-stage schedule. The baselines receive various PEFT methods, but they do not appear to be matched on total optimization steps, wall-clock training time, or number of trainable parameters. Consequently part of the improvement may stem from a larger optimization budget rather than from the proposed aggregation mechanism itself; a matched control is needed before the new components can be credited with the gain.
  3. The reported averages lack sufficient target-level uncertainty, and the sensitivity of the results to the candidate-pocket set is not established. Early-enrichment metrics can vary substantially when a target has few active compounds. Macro-averages over the 19 test targets are given, but neither per-target results nor confidence intervals appear in the main text, making it difficult to judge whether the gains are broadly distributed or driven by a small number of targets. Sensitivity to the number or generation of candidate pockets is likewise not quantified.
Questions for authors
  1. How are checkpoints and the best baseline tuning configuration selected, and is test performance excluded from these decisions?
  2. How much of the improvement remains when all methods are matched on training time, number of optimization steps, and trainable parameters?
  3. What are the per-target results and confidence intervals, and how much do the conclusions change with the number of candidate pockets?
1 comment
AuthorsOpenReview ↗

Response to Reviewer LNoR

Thank you for your insightful comments. We have carefully considered your concerns and revised our responses and supporting analyses accordingly. To make our revisions and supporting data easier to inspect, we provide an anonymous repository here.


Q1. How are checkpoints and the best baseline tuning configuration selected, and is test performance excluded from these decisions?

A1. Given the scarcity of allosteric adaptation data and the absence of experimentally verified inactive compounds in the training set, we used training loss to select AlloMIT checkpoints. As shown in Figure 1, we selected the checkpoint with the lowest training loss independently for each stage. Test-set performance was strictly excluded from both checkpoint selection decisions.

For our competitors, we compared multiple tuning configurations to avoid overlooking stronger baselines. We report the configuration with the best test performance (Appendix D). This gives the baselines the benefit of their best observed test performance, while AlloMIT checkpoints are selected solely by training loss.


Q2. How much of the improvement remains when all methods are matched on training time, number of optimization steps, and trainable parameters?

A2. As shown in the table below, AlloMIT retains its advantage over the HypSeek variants when training time and optimization updates are matched. Furthermore, Figure 7 in the manuscript shows that All-Module Tuning achieves an allosteric BEDROC of 19.73%, below AlloMIT's 23.25%, despite simultaneously updating 288.0K parameters. These comparisons support the view that the gains arise from AlloMIT's methodological design rather than simply increased optimization or model capacity.

MethodOverall BEDROC (%)Overall \mathrm{EF}_{0.5\\%}Allosteric BEDROC (%)Allosteric \mathrm{EF}_{0.5\\%}
HypSeek39.3124.8812.5720.79
HypSeek_poc30.8817.5310.6920.73
HypSeek_seq40.0725.8410.8010.47
AlloMIT51.2231.7823.2541.70

Q3. What are the per-target results and confidence intervals, and how much do the conclusions change with the number of candidate pockets?

A3. We provide per-target mean hit ranks and conditional confidence intervals in Figure 2. Clearly, AlloMIT improves mean hit ranks on most targets, with asterisks indicating statistically significant differences after Bonferroni correction.

AlloMIT's screening performance is stable across K = 20, 30, and 40, where the candidate-pocket bags provide broad target coverage. Smaller bags (K = 5 and 10) show both lower coverage and reduced enrichment, while increasing K beyond 30 provides only a modest coverage gain without improving enrichment.

KTarget coverage (%)BEDROC (%)\mathrm{EF}_{0.5\\%}
532.117.2910.47
1050.017.8220.79
2075.023.0941.70
30 (Default)82.123.2541.70
4085.722.4041.70
Reviewer_UkKQOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents AlloMIT, a multi-instance parameter-efficient fine-tuning framework for allosteric virtual screening, to address the challenges of unknown allosteric binding sites and limited activity data. The method represents a protein as multiple candidate pockets, integrates pocket–ligand matching information via Contextual Sparse Pocket Aggregation (CSPA), and adopts a two-stage fine-tuning strategy to adapt existing virtual screening models. The authors also construct AlloVSBench, which integrates allosteric activity data and conventional virtual screening datasets to evaluate the capability of allosteric molecule discovery. On this benchmark, AlloMIT improves HypSeek’s Allosteric BEDROC from 17.09% to 23.25% and [email protected]% from 20.79 to 41.70, and demonstrates similar performance gains on DrugCLIP and LigUnity. Ablation experiments further validate the effectiveness of multi-pocket aggregation and the two-stage training strategy.

Critical strengths
  1. This paper targets a practically meaningful challenge in allosteric virtual screening: unknown binding sites and limited activity data. Leveraging Multi-Instance Learning and Contextual Sparse Pocket Aggregation, the authors conduct virtual screening using multiple candidate pockets without pre-specifying the binding pocket. The method enjoys good practicality and plug-and-play properties, yielding performance improvements across multiple pre-trained models. For instance, AlloMIT raises HypSeek’s Allosteric BEDROC from 17.09% to 23.25% and [email protected]% from 20.79 to 41.70, with comparable gains observed on DrugCLIP and LigUnity.

  2. The manuscript carries out substantial work on dataset construction and experimental validation. The authors integrate ASD2023 with data resources including DUD-E, DEKOIS and LIT-PCBA to build AlloVSBench. Target splitting, data deduplication, and adoption of apo/predicted structures are adopted to make the evaluation more closely mimic real-world allosteric virtual screening scenarios. Furthermore, the ablation studies are comprehensive, systematically comparing different Pocket Aggregation approaches, the CSPA component, and the two-stage tuning strategy. As an example, joint tuning alone achieves a BEDROC of 12.15%, while the full method reaches 23.25%. These experiments provide solid support for the effectiveness of the core designs.

Critical weaknesses
  1. Although AlloVSBench contains nearly 800k target–ligand pairs, the test set only includes 25 allosteric positives distributed across 19 targets, with merely 1–3 positive molecules per target. Consequently, early enrichment metrics such as BEDROC and EF can be sensitive to ranking shifts of a small number of molecules.

  2. While AlloMIT serves as a practical plug-and-play framework, its overall technical innovation is relatively limited. Built upon existing pre-trained virtual screening models, the method mainly adapts and optimizes established techniques including Multi-Instance Learning, Sparse Aggregation, and Parameter-Efficient Tuning for the allosteric screening task. Although these designs are reasonable and validated via ablation experiments, the work is more of a task-specific improvement over existing approaches rather than a highly original methodological innovation.

Questions for authors

Given that the test set contains only 25 allosteric positives distributed across 19 targets, how do the authors view the impact of this small positive sample size on experimental evaluation? Specifically, has the team considered that early enrichment metrics such as BEDROC and EF may be sensitive to ranking shifts of a small number of positive molecules, which could undermine the stability of the reported results?

1 comment
AuthorsOpenReview ↗

Response to Reviewer UkKQ

Thank you for this insightful comment. We have carefully considered your concern and revised our evaluation and its presentation accordingly. To make our revisions and supporting data easier to inspect, we provide an anonymous repository here.


Q1. Given that the test set contains only 25 allosteric positives distributed across 19 targets, how do the authors view the impact of this small positive sample size on experimental evaluation? Specifically, has the team considered that early enrichment metrics such as BEDROC and EF may be sensitive to ranking shifts of a small number of positive molecules, which could undermine the stability of the reported results?

A1. We agree that experimentally confirmed allosteric hits are scarce and that early-enrichment metrics may be sensitive to the ranks of individual hits. To examine this risk, we present complementary hit-level and target-level sensitivity analyses in Figure 1 and Figure 2.

First, we grouped hit ranks into bins and compared their changes under AlloMIT relative to HypSeek. As shown in Figure 1, most hits move to an earlier bin or remain in the same bin. This pattern shows that the gains are accompanied by broadly maintained or improved hit rankings, rather than isolated improvements masking widespread deterioration.

We further assessed sensitivity to target composition by exhaustively removing combinations of one, two, and three targets. As shown in Figure 2, AlloMIT retains positive gains in both BEDROC and EF across nearly all evaluated combinations.

Together, these analyses suggest that the gains extend beyond isolated hits and are robust to the tested changes in target composition.


Additional response to W2 — Methodological contribution

Reviewer comment (Critical Weaknesses, W2). While AlloMIT serves as a practical plug-and-play framework, its overall technical innovation is relatively limited... more of a task-specific improvement over existing approaches rather than a highly original methodological innovation.

Response. We agree that AlloMIT does not introduce an entirely new architecture. Rather, our contribution lies in a systematic investigation of the learning paradigm under allosteric pocket uncertainty. As shown in Section 4.4 and Figure 8 of the manuscript, our analyses explain why AlloMIT first scores each ligand against individual candidate pockets and then aggregates the resulting scores. This provides a methodological insight for weakly supervised screening with uncertain candidate pockets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.