acceptodds
Withdrawn from ICLR 2027

K-PRS: Zero-Training Kernel Regression with Frozen Foundation-Model Embeddings for Cold-Start Drug Response

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine learning models for drug response prediction frequently fail to generalize to novel compounds under cold-start evaluation. We show that this failure stems from shortcut memorization: conventional architectures memorize historical compound identities during training, causing performance on unscreened molecules to collapse below naive cell-line baseline averages. To eliminate this vulnerability, we introduce a closed-form kernel estimator over frozen biological and chemical representations, gated by a positive-semidefinite prior that restricts evidence propagation strictly to compounds sharing identical mechanisms of action. By avoiding parameter optimization on response labels, the model is structurally prevented from memorizing entity identities. Across two independent cancer screening benchmarks, a matched five-fold protocol, and synthetic tasks with controlled identity availability, our non-parametric estimator consistently outperforms parameterized baselines. Furthermore, because predictions are formulated as closed-form weighted averages, the estimation weights inherently provide an exact evidence trail, calibrated uncertainty intervals, and verifiable robustness certificates from a single unified computation without auxiliary surrogates. All algebraic claims are formally machine-checked. Our findings establish both a practical predictor and a broader inductive principle: when entity identity is absent at deployment, explicitly refusing to learn it yields substantially stronger out-of-distribution generalization.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Zero Training, Cold Start, Drug Response, Foundation Model

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

The paper proposes K-PRS, a zero-training kernel-regression framework for cold-start drug-response prediction. The method combines frozen scGPT cell embeddings, frozen ChemBERTa-2/ECFP4 drug representations, a pathway-aware PSD kernel, and a masked residualized Nadaraya-Watson estimator. The authors argue that response-supervised predictors can memorize compound-specific response identities and fail under leave-one-drug-out evaluation, whereas a closed-form retrieval-based estimator avoids this shortcut.

Critical strengths
  1. The paper highlights an important cold-start evaluation issue in drug-response prediction. Strict LODO is more realistic for unseen compounds, and K-PRS clearly outperforms the simple cell-mean baseline on GDSCv2.

  2. The paper provides extensive diagnostic analysis. Ablations, stress tests, uncertainty analysis, evidence decomposition, and formal verification make the estimator relatively transparent.

Critical weaknesses
  1. The methodological novelty is limited. K-PRS mainly combines frozen representations with classical kernel smoothing, pathway gating, residualization, and Nadaraya-Watson regression. The theoretical non-identifiability result is generic and applies equally to K-PRS and trained estimators.

  2. The main gains rely heavily on curated pathway annotations. The pathway-only variant already achieves 0.6926 versus 0.7085 for the full model, and performance drops substantially when pathway labels are removed. It is unclear whether such annotations are realistically available for genuinely new compounds.

  3. The evidence for the claimed "memorization trap" is not sufficiently comprehensive. The matched trained baselines are limited, and several recent published architectures are not rerun under the same protocol. Therefore, the extremely poor ridge-head result does not establish a general failure of learned drug-response models.

  4. The external validation is limited. Experiments are restricted mainly to GDSCv2 and GDSC1, both historical cancer-screening datasets rather than prospective cold-start medicinal-chemistry settings. The authors also acknowledge that these datasets are not independent clinical cohorts.

  5. The presentation quality is far from publication quality. Many figures are overly dense, with very small fonts, crowded annotations, and too many panels. In particular, the central pipeline and several key experimental figures are difficult to read at normal page scale, which substantially hurts clarity and accessibility.

Questions for authors
  1. For a genuinely unseen compound, how is its pathway or mechanism-of-action label assumed to be known? Could the authors clearly define the deployment information budget and report the primary comparison under a setting in which pathway annotations for all query compounds are unavailable, noisy, or predicted rather than curated?

  2. Can the authors evaluate stronger recent drug-response models under exactly the same LODO splits, frozen inputs, and tuning protocol? Without such comparisons, it is difficult to determine whether the observed "memorization trap" is a general problem of learned response models or primarily a failure of the particular trained baselines considered here.

  3. What is the methodological gain over a much simpler pathway-restricted nearest-neighbor or Nadaraya-Watson baseline? Since the pathway prior supplies most of the improvement and the final gain from the dual frozen representations is relatively small, a simpler matched retrieval baseline would help isolate the actual contribution of K-PRS.

  4. How does the method behave under a stronger notion of novelty, such as scaffold-held-out or mechanism/pathway-held-out compounds? Standard drug-level LODO can still leave chemically or mechanistically close drugs in the reference set, which is particularly favorable to a kernel retrieval method.

  5. Could the authors redesign the main figures? In their current form, the small fonts, dense annotations, and large number of panels make the central method and empirical findings difficult to follow.

Reviewer_xHvjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets drug cold-start prediction, where the query compound's entire response column is deleted. The authors argue that trained fusion heads fail here because they memorise drug identity, and show that a ridge head over frozen features reaches 0.148 mean per-drug Pearson while reusing each cell line's mean response reaches 0.644.

Their method, K-PRS, fits no response-supervised parameter. Frozen scGPT and ChemBERTa-2 plus ECFP4 define the geometry, a PSD pathway gate restricts transfer to same-mechanism drugs, and a masked residual-centred Nadaraya-Watson average produces the prediction. On GDSCv2 it reaches 0.7085 against 0.644 for the cell mean, with GDSC1 replicating the gap. Because the estimator is closed form, the same weights also yield evidence attribution, conformal intervals, and a ranking certificate, with 38 supporting claims checked in Lean.

Critical strengths
  1. The matched-control negative result is a real contribution, independent of whether K-PRS is the right answer.

The paper's most valuable component is not the estimator but the controlled demonstration that trained response heads over strong frozen representations fail to clear a naive prior under drug cold-start. Ridge, kernel-ridge, and fusion heads get exactly the same frozen features and the same LODO split as K-PRS, so the gap can't be blamed on representation quality or preprocessing. Most convincing is the grouped five-fold result with nested tuning, where an interaction fusion head reaches 0.6379 against 0.6437 for the cell mean and loses to that naive prior on 226 of 295 drugs — running the honest protocol and still reporting the failure makes it hard to dismiss as under-tuning. This is the one claim that survives regardless of the method's fate: that apparent cold-start progress here can be an artifact of omitting a cell-mean baseline.

  1. The audit artifacts are non-vacuous, and the authors report where they come out badly.

With no learned head, the evidence decomposition, leave-one-out effect, conformal intervals, and ranking certificate all derive from the same weights that produce the prediction, so they describe the actual predictor rather than a post-hoc surrogate. More to the point, the authors let the artifacts embarrass them: certified embedding radii have median 6.5e-9 and none clears their own preregistered 0.01 threshold; coverage is always reported with interval width; the Lean development refutes their natural J − P form of the prior and corrects their gradient constant. A paper that preregisters a threshold and then reports missing it by seven orders of magnitude is unlikely to be quietly favourable elsewhere, which raises my confidence in the numbers generally.

Critical weaknesses
  1. The gain is carried by hand-curated pathway annotations, not by the frozen foundation-model embeddings in the title.

The ablation undercuts the paper's own framing. Over the cell-mean baseline (0.6442), the pathway-only kernel reaches 0.6926 — roughly three-quarters of the total improvement — while chemistry-only smoothing with frozen ChemBERTa-2 and ECFP4 reaches 0.6561, barely above the naive prior. The full dual representation adds 0.0159 on top of pathway-only. So the title's foundation-model embeddings are close to decorative, and the working ingredient is a mechanism-of-action lookup table.

This is consequential because it strains the central "zero-training, label-free" claim. Avoiding response supervision is not the same as avoiding supervision: curated target-pathway assignments are strong, response-correlated external knowledge, and the trained baselines are not given an equivalent prior. The honest comparison is a fusion head with the same pathway gate, which the paper does not run. Without it, the headline result may show only that mechanism annotations help — a known fact — rather than that refusing to learn identity is the better inductive principle.

  1. The memorisation trap — the paper's main intellectual claim — rests entirely on the authors' own baselines, one of which is implausibly bad.

A ridge head at 0.148 while the cell mean reaches 0.644 is not a subtle shortcut effect; it is a model performing far worse than predicting a constant per cell line. The more parsimonious explanation is a defect in the training or target-centering setup rather than a general property of response-supervised heads. The paper hedges ("a mechanism hypothesis, not a claim that every optimiser or architecture must fail") and then builds its narrative and title on that hypothesis anyway, while conceding it did not rerun any published architecture under the matched protocol.

Compounding this, the primary metric cannot see half the estimator: per-drug Pearson is computed across cells within a drug and is therefore invariant to the additive drug offset, so the entire separately-interpolated offset channel contributes nothing to the headline number. RMSE is 2.294 log-IC50 units and conformal widths run 7.46–14.29, which suggests the absolute predictions are not usable for the prioritisation task the impact statement describes. The diagnosis that motivates the method is thus under-evidenced, and the metric that validates the method tests only part of it.

Questions for authors

1.What does the fusion head do if you give it the same pathway information K-PRS gets? Right now the trained baselines have no mechanism prior, so the comparison can't separate "refusing to learn identity helps" from "pathway annotations help." This is the result most likely to move my assessment.

2.Did the trained baselines get the same per-drug residualisation? K-PRS centres each drug by its observed mean first. If ridge was fit on raw log-IC50, 0.148 could be a target-scaling artifact rather than memorisation.

3.How was ε chosen? Figure 5C shows a sweep with the used value marked, but the text says it never reads labels. What label-free criterion picked it?

4.Per-drug Pearson is invariant to the additive offset, which seems to make the offset channel invisible to your headline metric. Any evidence it helps? Per-drug RMSE for the baselines, alongside your 2.294, would settle it.

5.How many of the 295 drugs have usable pathway annotations, and what does the unknown-pathway fallback cost on the rest? A genuinely novel scaffold is exactly where annotation is likely missing.

Reviewer_VB7bOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The authors propose K-PRS, a zero-training framework to perform conditional predictions on new drug candidates using embedding models from foundation models. Cell-embeddings from scGPT and molecule embeddings from ChemBERTa-2 and ECFP4 from a set of known drugs and extend the prediction to the new candidate. The extension is performed using a pathway-aware positive semididefinite kernel and a masked Nadaraya-Watson estimator. This results in auditable predictions and allow for counterfactual assessment .

Critical strengths
  • The proposed method shows gain over the listed baselines with the proposed interpretability potentially being a very useful feature
Critical weaknesses
  • I think the paper lacks substantial background to frame the problem that the authors are trying to solve and especially how it relates to a broader class of problems. Furthermore, in general the presentation requires significant work with many parts of the text being difficult to follow, redundant, and potentially misleading. Much of the main text mathematical notation is under defined and extremely difficult to follow.
  • I think a key claim for the paper is that the cold-start problem is due to short-cut learning. But such a thing should be shown empirically. There is no reason to believe, a priori, that a high-dimensional vector representing a drug lies outside the convex hull of a set of previous inputs. I'm not certain that the ridge-head experiments are sufficient to support this conclusion.
Questions for authors
  • Line 137: "K-PRS places all of these views on one closed-form estimator" this is a strong statement if the proposed method is a unification of the many aforementioned post-hoc uncertainty and certification mechanisms listed before. Could you please clarify and add information about why that is the case?
  • Line 189: What is K_d and K_chem?
  • What does the per-drug Pearson distribution show?

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

Related papers

Discussion (0)

Sign in to comment.