acceptodds
Under review as a conference paper at ICLR 2027

HRD-BioAgent: Evidence-Grounded Agentic Reasoning for Multimodal Homologous Recombination Deficiency Prediction

Abstract

Homologous recombination deficiency (HRD) is commonly characterized by genomic scars, but its phenotype is also reflected across mutation, transcriptomic, epigenetic, and histomorphologic modalities. Conventional multimodal predictors can estimate HRD status, yet they provide limited support for reasoning about whether biological evidence is consistent with a patient-level prediction, particularly when definitive genomic measurements are unavailable. We introduce HRD-BioAgent, an evidence-grounded agentic framework that combines multimodal HRD prediction with adaptive biological evidence acquisition. HRD-BioAgent organizes modality-specific signatures in a provenance-aware HRD Evidence Map and uses a ReAct workflow to determine when additional genomic evidence is needed. For a patient requiring additional evidence, the agent can invoke a genomic analysis tool to derive SBS3 mutation-signature evidence from raw trinucleotide mutation profiles, while numerical predictions are produced by a calibrated fold-local meta-model rather than an LLM. We separate the value of the evidence that is integrated from the policy that decides when to acquire it. Under an "acquire-all" reference policy, in which every patient receives the complete evidence set, the evidence-integration model reaches AUROC 0.923 on TCGA-BRCA (); on a matched complete-case subset (), adding SBS3 to a non-agent OOF stacking baseline raises AUROC from 0.878 to 0.906 (two-sided paired-bootstrap ), and integrating the complete evidence set reaches 0.916, a further numerical increment that is not statistically distinguishable from chance under two-sided testing (AUROC , ). The paper's contribution therefore rests on SBS3 retrieval, provenance, and auditable concordance rather than on this increment. A "confidence-gated adaptive" policy that requests additional evidence only for patients whose prediction remains uncertain retains AUROC 0.907 against an acquire-all reference of 0.919 on the same cohort () while scheduling 58% fewer evidence-analysis steps. HRD-BioAgent also provides patient-level reasoning traces, provenance, and prediction-evidence concordance, enabling auditable HRD prediction under incomplete multimodal measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM agents, multimodal learning, missing modality, biomarker prediction, computational pathology, evidence-grounded reasoning, cancer genomics, interpretability

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_F8d7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HRD-BioAgent, a framework designed to predict homologous recombination deficiency (HRD) status—defined by genomic scar scores—by integrating data from whole-slide pathology images, transcriptomics, mutations, and methylation profiles. The framework employs an "HRD Evidence Map" to track the provenance, target population, and type of evidence; utilizes conditional tool invocation to incorporate supplementary information, such as SBS3 mutational signatures; and generates probability estimates via a calibrated statistical meta-model trained using an in-fold approach. The large language model (LLM) serves to orchestrate the workflow and synthesize textual output, rather than directly generating numerical predictions.

Critical strengths

1.The paper explicitly distinguishes between empirical data, model predictions, and secondary inferences; it delegates the generation of numerical probabilities to calibrated statistical models while maintaining a record of evidence sources and missing data. This design facilitates auditing and prevents the misinterpretation of unmeasured modalities as negative results. Deterministic checkers provide concrete, testable safeguards against process errors.

2.Tables 2 and 3 compare "stacking," static SBS3 enhancement, and the full system using the same set of patients, explicitly acknowledging that the final incremental step did not reach statistical significance. Table 7 also reports competitive results using L1 feature selection, without interpreting these as an advantage in agent reasoning. These analyses provide a useful basis for evaluating the sources of performance.

Critical weaknesses
  1. The most distinct performance gain in Table 3 arises from the static inclusion of SBS3—a mutational signature with well-established utility in predicting homologous recombination deficiency (HRD), as seen in tools like HRDetect. Since the evidence features utilized by the full system differ from those used in the static SBS3-only baseline, the observed improvement cannot be attributed solely to the reasoning process or the orchestration logic. Furthermore, the L1 baseline in Table 7 achieves performance comparable to that of the human-curated evidence set. The adaptive strategy described in Appendix J—which determines whether to escalate an analysis based on fixed probability thresholds—has not been benchmarked against a standard deterministic cascade (i.e., one employing the same inputs, tools, calibration methods, and thresholds). Moreover, the paper lacks sufficient documentation regarding the specific LLMs employed, the prompts used, the selection of tools, and the operational stability of the system. Consequently, it remains unclear whether the language model actively influences decision-making or if the same outcomes could be achieved through standard stacking combined with rule-based routing. The "85/85" error detection metric primarily validates the execution of predefined checks, rather than the biological accuracy of the synthesized textual evidence. For a paper that positions itself primarily around the concept of "agents," this represents a significant gap in the clear attribution of its core contributions.

  2. Sections 3.2–3.3 require that evidence estimation and feature selection be restricted to the training folds; however, Section 5 explicitly acknowledges that the priors for BRCA-related evidence were not entirely determined within the outer training folds. It must be clarified which features or thresholds utilized labels from the full cohort, and whether the primary findings remain robust under strict data isolation. Furthermore, a mere statement that "out-of-fold" (OOF) predictions were used is insufficient to verify the complete nested procedure involving base models, meta-models, and isotonic calibration. There is also a direct contradiction in the rules for "concordance" analysis: Section 4.5 of the main text requires at least two out of the four pieces of evidence to be positive, whereas Appendix A modifies this rule to require at least one positive piece of evidence excluding TP53—while simultaneously noting that only 25 of the original 105 "Concordant+" cases remain classified as such. Under the aforementioned definition, any patient who met the "two positive pieces of evidence" criterion must have had at least one piece of evidence other than TP53; therefore, assuming the predictions remained unchanged, such cases should not have lost their "concordant" status. Additionally, Figure 3 labels a case as "Concordant+" even though it lacks mutations in TP53, BRCA1, or BRCA2, has no methylation data, and only meets the E2F criterion—a classification that conflicts with the rules stated in the main text. Given that "auditable concordance assessments" constitute the core output of this paper, these discrepancies must be resolved through the establishment of consistent rules, the presentation of patient-level results, and the provision of an executable implementation.

Questions for authors

1.Please report the LLM used, its version, the prompts employed, and the tool-calling mechanism; furthermore, specify which actions are determined by the LLM and which are governed by fixed rules. Given identical evidence (across 11 items) and training protocols, how do static meta-models and deterministic cascades compare to a full agent in terms of predictions, invocation costs, and reporting accuracy?

2.Which BRCA evidence priors or thresholds utilize labels from patients in the hold-out test set? Please provide the complete nested pseudocode—covering the base model, meta-model, calibrator, and the selection of routing thresholds—and report the primary results obtained under strict isolation.

3.Please explain how the reduction from 105 cases to 25 cases in Appendix A is derived from the stated rules, and clarify why Figure 3 is labeled "Concordant+." Were different rules applied to Figure 3, Table 5, and Appendix A? How would the group sizes and conclusions change if a consistent set of rules were applied throughout?

4.In Table 13, when a specific modality is removed, are all evidence features derived from that modality also removed—rather than just the predictions from the corresponding base models? How do the multi-stage models described in Appendix J handle evidence that has not yet been acquired, and are the routing thresholds determined solely based on the training data? Please also provide information regarding the uncertainty stemming from differences in strategies, as well as the actual invocation rate of SBS3.

2 comments
AuthorsOpenReview ↗

Response to Reviewer F8d7

Thank you. We accept your points. G1 to G4 refer to the general response; numbers refer to the revised PDF.

W1.

  • (i) Gain comes from static SBS3. Accepted. The feature also needed correction: the submitted "SBS3 reference" vector was not COSMIC SBS3 (Pearson 0.05) but nearly uniform; the final agent uses the cosine to COSMIC v3.3 SBS3 (univariate AUROC 0.873 vs 0.822). Fitted SBS3 activity is non-zero in 2.5% of patients, and we do not claim that the cosine quantifies the SBS3 process (Appendix E.1).
  • (ii) Feature mismatch. The fusion baselines already had a 4,657-gene mutation vector including BRCA1, BRCA2 and TP53. Adding five evidence features changes their AUROC by at most +0.006, none significant (Table 30; partial parity). In the revision, stacking and the agent's K3 configuration share identical base learners (Appendix A.3).
  • (iii) L1 baseline. Data-driven selection is reported next to the agent (Table 7, point estimates). It reaches 0.923 to 0.925 in BRCA against 0.930, is better than the agent in LUAD (0.848 vs 0.822), and is worse in UCEC (0.907 and 0.872 vs 0.934) and OV (0.878 and 0.857 vs 0.897). A selector without knowledge candidates is significantly worse in BRCA, UCEC and OV (Table 13).
  • (iv) Deterministic cascade. Benchmarked in Appendix F.1. A scripted rule reproduces the ReAct routing exactly (Δ = 0.000 [0, 0]). Confidence-gated cascades reach 0.904 to 0.913, a threshold gate 0.911, the scripted rule 0.918 and acquire-all 0.922 to 0.923. We claim no gain from routing (G1).
  • (v) LLM. No LLM takes part in prediction or routing. The narration study uses gpt-4o-2024-11-20 at temperature 0; prompts and judge protocol are in Appendix D.4.
  • (vi) 85/85 validates execution only. Agreed. The audit is now described as an internal consistency check, because renderer and parser share a template (§4.4). The independent test is Table 2: a GPT-4o writer makes a strict critical error in 1, 2 and 13 of 100 reports from tagged, untagged structured and raw-table inputs.

W2. Full-cohort information; nested procedure. In the submitted version, pathway selection and expression normalisation used the full cohort, and routing thresholds were fixed constants. The revised procedure is written out in Appendix A.4: (1) outer split; (2) unimodal learners fitted on the outer training partition, with out-of-fold scores for training rows; (3) knowledge features mapped without labels, direction check and z-scoring on training rows; (4) configuration chosen by inner 3-fold cross-validation; (5) penalty, calibration (isotonic if N ≥ 300, otherwise Platt) and sensitivity-90 threshold from inner out-of-fold predictions; (6) one refit, and the test fold scored once. A label guard recorded 0 violations. The residual dependence and its nested bound are in G2.

W3. Concordance contradiction; Figure 3. See G3. Patient-level results are in Table 21 and release states across cohorts in Table 23. The same code computes the release state for both tables and the case figure.

Q1. LLM details: W1(v). Predictions: W1(iv). Invocation cost: 905 spectrum calls

Thanks for your rebuttal. However, I believe this article has room for further revision, and you failed to write the AI ​​usage instructions as required by ICLR, which is a major oversight. I am maintaining my original score.

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

This paper proposes a multimodal framework for predicting homologous recombination deficiency from whole-slide histopathology images, RNA-seq, somatic mutations, and DNA methylation across breast, ovarian, and endometrial cancer cohorts in TCGA. Rather than relying on a single end-to-end black-box classifier, the authors pair unimodal predictors with a curated HRD Evidence Map and a ReACT-style workflow that retrieves mutational signatures such as COSMIC SBS3 when initial predictions are uncertain. Numerical risk estimation is handled by a fold-local calibrated logistic regression meta-model, while a rule-based critic checks provenance constraints and generates patient-level concordance reports.

Critical strengths

The key contribution of this work is its clinical system design, which explicitly separates measured biological evidence, model predictions, and secondary inferences. By preserving feature provenance and avoiding the relabeling of predicted molecular features as direct genomic evidence, the framework improves interpretability and clinical trust. It also restricts the language model to workflow orchestration and report synthesis, while assigning probability estimation to a calibrated statistical model to ensure reproducibility and avoid hallucinated risk estimates.

Critical weaknesses

My primary reservation is that the core methodology looks incremental, and the high-level agentic framing is largely disconnected from how the system actually learns and predicts. Strip away the agent terminology, and the underlying predictive architecture is a classical two-stage stacking pipeline: off-the-shelf unimodal models produce out-of-fold probabilities, which are concatenated with eleven domain features and fed into a standard logistic regression classifier with isotonic calibration. Because the language model is completely barred from participating in representation learning or probability estimation, and because the main model reported throughout the paper simply acquires all eleven features for every patient by default, there is no autonomous reasoning or decision-making in the primary workflow. Even when selective feature acquisition is explored in the appendix, the policy is not learned; it is just a hard-coded two-threshold rule applied to the initial scalar probability. Calling a thresholded feature lookup and a logistic regression stacker an evidence-grounded ReACT agent overstates the algorithmic novelty of the work.

The other two main architectural components, the provenance-aware Evidence Map and the deterministic critic, are essentially software engineering data structures rather than machine learning contributions. At a design level, the Evidence Map is a static configuration dictionary that stores thresholded summary statistics derived from classical univariate tests like DESeq2 and fGSEA. Rather than learning complex cross-modal interactions or discovering latent biological representations end-to-end, the model relies on manually curated tabular features and pre-specified cutoffs. Likewise, the deterministic critic is a collection of basic input validation assertions that check whether probabilities fall between zero and one, whether fold identifiers are valid integers, and whether cohort labels match. While input validation and metadata tracking are good software engineering practices for clinical pipelines, presenting basic schema checks and static lookup tables as core methodological pillars highlights the lack of technical innovation suitable for ICLR.

Beyond the lack of algorithmic novelty, there is a fundamental conceptual contradiction in how the problem formulation and feature allocation are designed. The paper motivates its multimodal approach by arguing that definitive genomic HRD measurements are often unavailable in clinical practice, yet the agent's ability to resolve uncertain patients hinges on extracting the COSMIC SBS3 mutational signature and BRCA mutation status from whole-exome sequencing variant files. In a realistic clinical setting, if whole-exome sequencing and methylation arrays are already available on disk to query as a tool, the laboratory already possesses the sequencing data required to compute genomic HRD scar scores directly. Moreover, computing SBS3 cosine similarity from an existing variant file is a trivial vector dot product that takes microseconds, so motivating adaptive feature retrieval as a way to save analysis steps optimizes an artificial proxy rather than actual clinical assay cost or computational latency. This design flaw is compounded by an artificial feature split in the baseline comparisons, where key causal HRD drivers like BRCA1/2 mutations and SBS3 are withheld from the baseline multimodal fusion models and reserved exclusively for the second-stage meta-model.

Questions for authors
  1. What is the actual role of the language model and the ReACT loop if numerical integration is performed entirely by a static logistic regression meta-model and selective retrieval is governed by fixed probability thresholds? Could the entire prediction, routing, and critic pipeline be executed by a brief deterministic script without invoking a language model at all?

  2. From a system design perspective, why were known HRD drivers such as BRCA1/2 mutations and the SBS3 mutational signature excluded from the baseline multimodal predictors and only introduced inside the BioAgent meta-model? Since the main acquire-all agent uses the exact same calibrated logistic regression estimator as the stacking baseline, wouldn't a static stacking model given all eleven features produce the exact same predictions as HRD-BioAgent?

  3. How do you reconcile the clinical motivation that genomic HRD assays are unavailable with a workflow that requires whole-exome sequencing variant calls and methylation arrays on demand to compute SBS3 signatures and refine uncertain predictions?

  4. Given your acknowledgment in the discussion and appendix that breast cancer evidence priors were not fully selected within outer training folds and that pan-cancer map construction still requires verification for test-fold isolation, how can we be sure that the reported performance gains are free from feature-selection leakage?

Reviewer_9AVeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HRD-BioAgent combines multimodal HRD prediction with a biological evidence map, conditional evidence retrieval, and patient-level explanations. The problem is relevant, and separating measured assays from model-derived evidence is sensible. However, the evaluation does not establish the contribution of the agent, the evidence-selection procedure is not fully independent of the test data, and the concordance analysis contains a substantive inconsistency.

Critical strengths

The matched cohort comparisons and SBS3 ablation help identify where the predictive improvement comes from. I also appreciate the authors explicit acknowledgment that the additional improvement beyond stacking with SBS3 is not statistically significant. Keeping numerical prediction within a calibrated estimator makes the prediction component easier to inspect.

Critical weaknesses

A. The evidence-selection procedure compromises the held-out evaluation, the limitations explicitly state that brca evidence priors were not fully selected within the outer training folds. These priors determine which biological features enter the model so fitting the final estimator within each fold does not resolve the earlier exposure to test-set information. This leaves uncertainty about how much of the reported performance would survive a fully independent evaluation. The entire evidence-map construction and feature-selection procedure needs to be repeated within the outer training folds.

B, The experiments do not establish what the agent contributes. On the matched cohort, adding SBS3 to ordinary stacking raises AUC from 0.878 to 0.906; the complete method reaches 0.916. Tab. 7 also reports essentially the same AUROC for l1 feature selection and the curated evidence set 0.920 v 0.919. The authors acknowledge these results, but a static model receiving exactly the same evidence remains missing. Moreover, App J describes acquisition through fixed probability thresholds. A comparison against a scripted workflow with identical features, thresholds, and estimator is necessary to establish whether the agent adds predictive value, better evidence assessment, or useful flexibility.

C, The concordance results are inconsistent with the stated rules see sec. 4.5 requires at least two of four signals for evidence support: TP53 mutation, elevated E2F, BRCA1 methylation, and BRCA1/2 mutation. app A then removes TP53 and requires at least one remaining signal, but reports that only 25 of the original 105 Concordant+ patients retain support. Under the stated definitions, all 105 must retain support: satisfying the original rule necessarily requires at least one non-TP53 signal. Either the implemented rule differs from the description or the reported counts are incorrect. This needs to be resolved, particularly because auditable concordance is presented as a central contribution.

D, The practical benefit remains insufficiently demonstrated. All evaluations use tcga; the additional cancer analyses involve within-cancer training and evaluation rather than external validation. Table 12 reports 72 false negatives among 181 reference positive patients, implying approximately 60% sensitivity at the reported operating point. A clinically motivated operating point and its sensitivity-specificity tradeoff would be more informative. Likewise the reported 58% saving concerns defined analysis steps, with four pathway scores counted separately despite using the same RNA assay. The authors acknowledge that this does not measure runtime or financial savings leaving the practical advantage of selective acquisition unresolved.

Questions for authors

a, Which decisions in the reported experiments actually depend on the LLM? Please specify the model, prompts, and execution settings, and show what changes when orchestration is replaced with the published threshold rules.

b, After restricting all evidence discovery and selection to the outer training partitions, how do the main results change?

c, The SBS3 feature is cosine similarity to a reference signature. Does the result hold with fitted SBS3 activities, using a method such as SigProfilerAssignment, and how stable is the evidence at low mutation counts?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.