acceptodds
Under review as a conference paper at ICLR 2027

Model-driven Collaborative Information Enhances Drug Activity Prediction in Low-resource Scenarios

Abstract

Drug development is hindered by the high cost and time consumption of identifying bioactive molecules, particularly in zero-shot and few-shot scenarios where experimental data is scarce. Existing methods often fail to generalize to novel proteins or assays, limiting their applicability in real-world drug discovery. Furthermore, these approaches struggle with the simultaneous processing of heterogeneous data types, including assay descriptions, molecular structures, and protein sequences. To address these limitations, we propose Assay Similarity-based Activity Prediction (ASAP), a novel framework that leverages a model-driven assay similarity and ranking loss to predict bioactivity. By learning the collaborative information between assays and mitigating batch effects, ASAP enables robust predictions for low-resource scenarios, such as novel assays with only text or protein sequence information. Extensive experiments demonstrate the superiority of ASAP in both zero-shot and few-shot settings. Our proposed approach overcomes the limitations of traditional pair-input models and provides a scalable, generalizable solution for accelerating drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
drug activity prediction, collaborative information, drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yma8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on biological assay prediction. It is a challenging problem given the sparsity of measured activities per assay. Particularly for new assays with only few compounds labeled. Therefore authors present ASAP. They first estimate across-assay similarities by evaluating the prediction accuracy for one assay based on trained model of another assay. Then they leverage the similarity matrix to establish a collaborative assay space using contrastive learning, drawing similar assays together, and pushing dissimilar assay away from each other, where the assays are encoded through the protein sequence and assay description in text. When it comes to bioactivity training, ASAP combines an molecular encoder and the assay encoder pretrained in the collaborative assay space and applies an MLP to predict a score that is optimized by the ranking across compounds on the specific bioactivity. The authors benchmark their approach in zero-shot and few-shot scenarios, and demonstrate superior performance in most settings based on correlation metrics like Pearson’s coefficient and Concordance Index.

Critical strengths
  • The ASAP aims at true generalization towards zero-shot and few-shot scenarios for new assays in low data regime with few measured compound activities. The authors formulate their solution appropriately so that, any new assay could be encoded in the collaborative assay space, either by the text description or protein sequence, giving a boost start to the new assay model training.
  • The authors formulate their experiments rigorously. Particularly for zero-shot and few-shot settings, they purposefully prevent the data leakage issue by removing any overlapping compounds and similar target assays between training and unseen testing datasets.
Critical weaknesses
  • It is unclear how biologically sound the assay similarity matrix is by the cross-assay predictivity. It can be largely affected by the compound data availability and diversity. Some assays may have very few measured compounds, some may severe class-imbalance among data with measurements, some may only cover very limited scope of chemical space. It’s likely not revealing the real biology in assay similarity given it’s very dependent on the existing limited data. Using that assay similarity matrix as a reference for contrast learning to establish the collaborative assay space could be risky, and therefore further affect the generalizability for the full framework. The authors showed a couple examples for biological relevance for computationally similar assays, and demonstrated the protein-level clustering on the latent space, but I would expect a more systematic biological validation given this is the biggest assumption of this paper. If most assays are protein target related, I find it surprising that the authors did not try to compare with reference datasets of protein-protein interactions.
  • While ranking based objective makes sense for training, using only ranking or correlation based metrics for evaluation is not ideal. The relative binding likelihood or bioactivity strengths could be helpful, but in many cases scientists would still take the absolute binding affinity/probability or bioactivity magnitude to guide their decisions. Missing benchmarks on the absolute assay value predictions makes it less actionable.
Questions for authors
  • How to put ASAP into practical use? Especially with a relative score outcome per outcome, how would biologists and chemists use it for compound mechanism understanding or screening?
  • The authors promotes generalization of their method through zero-shot and few-shot testing. There’s a strong focus on assay-level knowledge transfer which essentially enables the generalization. It has its value, but to what degree does it help the assay prediction be usable? Would it make more sense to have a better predictor module than MLP? Would it be possible to leverage additional measured data from other similar assays to improve the prediction of one particular assay?
Reviewer_hX1COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ASAP, a two-stage framework for drug–target bioactivity prediction aimed at low-resource settings (zero-shot and few-shot) where a new assay has only textual description and/or target sequence. Stage 1 extracts "collaborative supervision" between assays using a model-driven similarity (train an SVR on assay i with ECFP fingerprints, predict assay j's molecules, take the Pearson correlation), then pre-trains assay embeddings with a contrastive margin loss that pulls similar assays together and pushes dissimilar ones apart. Stage 2 trains a molecule encoder (ChemBERTa) jointly with the pre-trained assay encoder using a within-assay ListMLE ranking loss to reduce sensitivity to absolute activity scales. Experiments cover zero-shot (target–drug and text–drug) and few-shot transfer on Davis and KIBA, plus a ChEMBL pre-training ablation.

Critical strengths
  1. Well-motivated problem setting. The distinction between transductive DTA and inductive assay-level transfer, and the two zero-shot sub-settings (target–drug, text–drug), is clearly laid out and genuinely relevant to real drug discovery.

  2. I like the ranking loss instead of regression, which is a well-justified design choice for heterogeneous assay measurements, and the ListMLE formulation is appropriately applied.

Critical weaknesses
  1. Insufficient differentiation from Assay-Match[1], which shares the same core idea. ASAP's central premise—deriving a model-driven assay similarity and using it to transfer knowledge to unlabeled assays—is essentially the same as that of Assay-Match (Fan & Barzilay, 2026), which learns assay compatibility via TRAK data attribution and uses it to select training data for unlabeled test assays. Yet the paper devotes only a single sentence to this work in Related Work (Sec. 5), without clarifying the substantive difference between the two similarity signals or providing any empirical or ablation evidence that ASAP's SVR/Pearson-based similarity is preferable. Because both methods construct assay similarity from model behavior (SVR prediction correlation vs. TRAK attribution) and apply it to the same zero-shot transfer problem, the novelty claim is under-supported; the authors should at minimum (i) articulate why SVR-based similarity is better suited to this task, (ii) report the correlation between the two similarity signals, and (iii) compare downstream performance against an Assay-Match-based variant.

[1]. Fan, V., & Barzilay, R. (2026). AssayMatch: Learning to Select Data for Molecular Activity Models. Journal of Chemical Information and Modeling, 66(7), 3540–3549.

  1. All results are reported only via molecule-count-weighted w-Pearson and w-CI. Because assays in the curated data range from 20 to 10,000 molecules (Appendix B.1), a few large assays can dominate the aggregate, and the weighted metric systematically down-weights exactly the small assays that matter most in the low-resource setting the paper targets; no macro-averaged (per-assay) metrics, per-assay success rates, or assay-size-stratified results are reported, so it is impossible to tell whether ASAP's advantage is broadly distributed or driven by a few large assays.

  2. The ablation in Table 4 is too thin to isolate the paper's two central claims: there is no comparison of the model-driven SVR/Pearson similarity against alternative similarity sources, and no ranking-vs-regression ablation (Stage 2 with MSE on absolute activities) to substantiate the batch-effect argument.

Questions for authors
  1. Computational cost and scalability of Stage 1. The paper does not report the cost of constructing the SVR/Pearson similarity matrix, which requires O(D²) pairwise evaluations over 17,517 pre-training assays.

  2. How is the assay embedding obtained when one modality is missing, and is this case properly trained for? Figure 2 depicts the assay embedding as a concatenation of a text branch and a protein-sequence branch followed by an MLP, which presupposes both modalities are present.

  3. The paper argues that ASAP generalizes better than structure-based DTA methods because it does not require protein structural or interaction data, which are often unavailable for novel targets (Sec. 1, C2). However, "requiring fewer inputs" is not equivalent to "generalizing better." I mean fundamentally, generalization to a new assay still relies on a learned model. Why, then, should this approach be considered to have better generalization than methods based directly on protein structure and drug structure? Is there any intuitive or empirical explanation for this?"

Reviewer_xBdKOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes ASAP, a method to rank molecules by activity in a new assay that has few or no measurements. The method has two stages in stage 1, a model is fit to each training assay and used to predict every other assay. The Pearson correlation between these defines a "collaborative" similarity of the two assays. An assay encoder is pre-trained with a contrastive margin loss so that similar assays get nearby embeddings. In stage 2, the assay embedding and a molecule embedding are fed to an MLP that is trained with a listwise ranking loss within each assay. The model is pre-trained on ChEMBL and evaluated zero-shot from the protein sequence, zero-shot from the assay text, and few-shot by fine-tuning on a few assay.

Critical strengths
  1. The problem is relevant and the introduction motivates it well. Historical assay data is large but heterogeneous, and new assays often have no measurements, so transferring knowledge across assays is a important goal.

  2. The evaluation protocol has good elements. Molecules shared with Davis/KIBA and targets with ≥30% sequence identity are removed from pre-training, and scoring the ranking within each assay sidesteps incomparable units.

Critical weaknesses
  1. The introduction motivates ASAP with assays that differ in experimental conditions, protocols and readouts, with batch effects, and with phenotypic assays that have no known target. The only part of the method that addresses heterogeneity is the within-assay ranking loss, which removes a per-assay shift and scale but nothing else. The data then excludes the motivating cases: only "assays targeting single proteins" are kept, and Davis and KIBA are pure binding sets in which an "assay" is simply one protein. The paper therefore evaluates zero-shot protein-to-molecule transfer, not transfer across heterogeneous experimental procedures.

  2. The model-driven similarity is weakly motivated and overstated as novel. An SVR fit to assay is a smoothed version its structure–activity data, so predicting assay j with it amounts to comparing the structure–activity relations of the two datasets directly, through a particular smoother. Any similarity of this kind inherits the same problems the paper lists for conventional similarities: for assays with as few as 20 molecules it is very noisy, and it changes substantially with the fingerprint. The "transductive nature" argument applies equally to ASAP. Adding a new training assay requires new SVR fits against all other assays and re-pre-training, and the roughly 250 million cross-assay predictions for about 15,800 training assays are not costed. The construction of C is ad hoc. and merely maps the mean correlation to [0, 1], which does not make it "probability-like". Combined with w, two unrelated assays (correlation 0) get weight 0.5, i.e. equal pull and push, and a correlation of 0.2 gives only 0.62, so most pairs carry little signal. Using predictions of one assay's model on other assays is also not new: profile-QSAR (Martin et al., JCIM 2011) builds on exactly this. The contrastive margin loss (Eq. 2) is the standard contrastive loss of Hadsell et al. (2006) with soft labels, which is not cited.

  3. The gains over baselines are not shown to come from the collaborative similarity. In Table 4, ASAP without collaborative pre-training already reaches 27 Pearson, against at most 22 for every baseline, so most of the gap exists without the paper's main idea. The collaborative pre-training is ablated only on the ChEMBL test split, not on Davis/KIBA where the headline comparisons are made. A cheaper similarity (target sequence identity, ESM2 or text-embedding cosine) is never tested in place of the SVR similarity, which is the natural control. A molecule-only ranking that ignores the assay is also missing.

  4. On Davis, ASAP beats the best baseline by about one standard deviation. On KIBA, it is tied or behind. After removing censored entries, a Davis target has about 24 measured molecules on average. The claimed "superiority" is not established.

  5. The few-shot evaluation is not convincing. "Few-shot" here means 10–40% (Davis) or 5–20% (KIBA) of each assay's molecules, not the handful of labelled compounds of standard few-shot benchmarks such as FS-Mol, which is not used. The methods designed for this setting (FS-CAP, ActFound, AdaMBind) sit at roughly 2–12% Pearson by my reading of Fig. 3, far below the zero-shot DTA baselines. This points to problems with the re-implementations rather than to weak methods. Fig. 3 also uses broken axes and has no error bars.

  6. The text-based evaluation uses only the authors' own ChEMBL split, which is split by assay ID without target filtering, so test assays likely share targets with pre-training assays. The benchmarks used by the closest prior work (e.g. CLAMP, MHNfs) are not used, so the numbers cannot be compared with published results. ASAP scores on the same test set different scores (table 3 & 4), and the paper does not say which inputs each table uses. The evidence that the similarity is "biologically meaningful" consists of two hand-picked assay pairs explained after the fact and a t-SNE plot coloured by protein class.

Questions for authors
  1. How does ASAP compare with the same encoders and ranking loss when the SVR similarity is replaced by target sequence identity, or by ESM2/text-embedding cosine similarity? How does it compare with a molecule-only ranking that ignores the assay?
  2. How do ASAP and the baselines perform on assays that are not single-protein assays (cell-based or phenotypic assays), which motivate the paper?
  3. What is the wall-clock cost of computing the similarity matrix, and how is it updated when new assays are added?
  4. Why do FS-CAP, ActFound and AdaMBind score far below the zero-shot baselines in Fig. 3? How do the methods compare on FS-Mol with its standard support-set sizes?
  5. Which inputs are used in Table 3 and Table 4, and why does ASAP differ between them on the same test set?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.