{"id":"ec05b616-fdda-4624-8308-49bfb0d7d3c0","arxiv_id":"2506.19681","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PathLUPI uses transcriptomic pathway information as privileged training signal to improve WSI-only prediction of biomarkers, molecular subtypes, and survival across 49 oncology tasks.","lead":"This paper introduces PathLUPI, a method that uses gene expression data during training to teach a model to predict cancer mutations, subtypes, and survival from microscope slide images alone. The approach reports consistent gains over standard image-only models across dozens of tasks, which could make molecular insights available from routine pathology without sequencing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for the privileged signal: PathLUPI's reported gains could come from added capacity, not from transcriptomic supervision; a shuffled-transcriptome or equal-capacity WSI-only control is required.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the paper never ablates the privileged branch or permutes the transcriptomic input, so the causal attribution of the performance gain to transcriptomic supervision is not isolated. I agree with that assessment. PathLUPI is a carefully evaluated framework with broad TCGA coverage, external cohorts, and consistent point estimates, and the absolute-performance claims are not the problem. The problem is that the comparison baselines are simpler MIL architectures, while PathLUPI adds substantial machinery (RRT, cross-attention, pathway MLPs, alignment losses). Without a same-capacity WSI-only control or a shuffled-transcriptome control, the headline improvement cannot be assigned to the privileged information itself. This is a scientific attribution issue, not a suggestion of misconduct. The secondary issues noted by the reader, such as overconfident confidence intervals and the 'over half of tasks' phrasing, are real but less load-bearing than the missing control. The conditional verdict remains appropriate: the paper should be asked to supply the control experiment before the genome-anchored causal claim is taken at face value. No change to the reader's verdict is needed.","tokens_in":30205,"tokens_out":4666,"duration_ms":48225,"concrete_test":"Retrain PathLUPI on two representative internal tasks (e.g., CRC-BRAF and GBMLGG-IDH1) under three conditions: (a) original transcriptomic privileged input; (b) transcriptomic profiles randomly permuted across patients before each epoch, preserving marginal distributions but breaking case-level correspondence; (c) an equal-capacity WSI-only variant with the privileged branch removed and the alignment losses either disabled or redirected to a self-supervised reconstruction of the RRT features. Compare 5-fold AUC with bootstrap 95% CIs and a paired test across folds. If (b) or (c) matches or exceeds (a), the reported advantage is a capacity or architecture effect, and the genome-anchored claim fails; if (a) clearly beats both, the privileged-signal attribution is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is causal: the 4.48% mean AUC gain over the best WSI-only baseline is attributed to transcriptomic privileged information. That attribution is not isolated. Figure 6 ablates only the visual backbone (ResNet50, PLIP, UNI, CONCH); no experiment removes or permutes the privileged transcriptomic branch. PathLUPI differs from the ABMIL/CLAM/DTFD/TransMIL baselines not only by receiving RNA-seq at training time, but also by containing a region-aware re-embedding transformer (RRT), shared cross-attention, 50 pathway-specific MLPs, and three auxiliary alignment losses. Any or all of these added components could produce the gains through increased model capacity or better optimization, without the model ever using the biological content of the transcriptomes. The distilled branch is itself trained with access to the privileged branch and alignment losses, so it is not a WSI-only control; it is the hypothesis under test. A permutation control (shuffling gene-expression vectors across patients) would preserve input statistics and architecture while destroying phenotype-genotype correspondence. Without such a control, the paper's attribution of the improvements to genome anchoring is not supported, even though the point estimates across 49 tasks and external cohorts are credible and valuable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces PathLUPI, a learning-using-privileged-information (LUPI) framework that uses transcriptomic pathway signatures (Hallmark gene sets) as privileged supervision during training to produce genome-anchored histological embeddings. At inference, only whole-slide images are used. The authors evaluate PathLUPI on 49 molecular oncology tasks spanning 13 cancer types and 20 cohorts with 11,257 cases, including biomarker prediction, molecular subtyping, and survival prognosis. They report consistent improvements over ABMIL, CLAM, DTFD, and TransMIL, with a headline mean AUC increase of 4.48% over the best baseline in biomarker prediction and improved C-index in survival analysis. The paper also presents interpretability analyses linking attention patterns to known genotype-phenotype associations, and an ablation study comparing CONCH with PLIP, UNI, and ResNet50 as feature extractors.","tokens_in":30420,"tokens_out":3696,"duration_ms":41215,"significance":"If the central claim is causally valid, this is a substantial contribution: it provides a general recipe for using paired transcriptomic data at training time to improve WSI-only molecular prediction, it is evaluated on an unusually broad and clinically relevant benchmark, and it includes external cohort validation and interpretability analyses. The scale of the evaluation, the public data sources, and the promised code release are strengths. However, the causal attribution that the improvements come specifically from the privileged transcriptomic signal is not yet isolated, and the statistical presentation overstates precision. The contribution is therefore significant but needs additional experimental and statistical work before the main claim is fully supported.","major_comments":[{"comment":"The central causal claim that transcriptomic privileged information drives the reported gains is not isolated. Figure 6 ablates only the visual backbone (ResNet50, PLIP, UNI, CONCH). No experiment removes, shuffles, or permutes the transcriptomic branch. Relative to the baselines, PathLUPI adds a region-aware re-embedding transformer, shared cross-attention, 50 pathway-specific MLPs, and three alignment losses, so an equal-capacity WSI-only model or a permuted-transcriptome control (e.g., shuffling gene expression vectors across patients while preserving the architecture) is required to attribute the mean 4.48% AUC improvement to genome anchoring rather than to added capacity or optimization changes.","section":"Methods: Genome-anchored representation learning; Results: Validating the impact of foundation model embeddings"},{"comment":"The reported 95% confidence intervals are implausibly narrow. For example, Extended Data Table 2 reports AUC values such as 0.849 (0.848-0.850) for a five-fold cross-validated estimate. Bootstrapping out-of-fold predictions with 1,000 resamples captures uncertainty of the metric for a fixed set of fits, not variability across the five training runs and random splits. The one-sided Wilcoxon signed-rank test is appropriate for paired task-level comparisons, but the manuscript does not specify whether the test is over tasks, folds, or bootstrap replicates, and the reported P<0.001 for molecular subtyping is not attainable with a task-wise test on 8 tasks (minimum one-sided p = 1/256 ≈ 0.0039). The statistical methods and interval estimates need to be clarified and, if necessary, corrected.","section":"Methods: Implementation details; Extended Data Tables 1-4"},{"comment":"The alignment loss weight λ is introduced in the total loss L_total = L_sup + λ(L_rec + L_attn + L_rep), but no value or range for λ is reported in Extended Data Table 8 or anywhere else. The number of latent spatial regions R is stated as 50, and the top attention patch fraction is used in the interpretability analysis, but the loss weight and the threshold choice for 'top 1%' patches are not justified or varied. Since the contribution is specifically about the privileged alignment mechanism, the sensitivity of the main result to λ should be reported.","section":"Methods: Multi-level alignment and training objective; Extended Data Table 8"}],"minor_comments":[{"comment":"The number of tasks is inconsistent: the Abstract and Results state 49 tasks, while the Introduction states 48 tasks. The Extended Data Tables sum to 49 tasks, so the Introduction should be corrected.","section":"Introduction and Results"},{"comment":"The sentence 'By integrating transcriptomic priors via the LUPI paradigm, PathLUPI significantly improves biomarker prediction accuracy' is duplicated verbatim. One copy should be removed.","section":"Results: PathLUPI enhances biomarker prediction"},{"comment":"The sentence beginning 'Evaluation across these diverse cohorts revealed PathLUPI’s strength...' is followed by a partial repetition ('revealed PathLUPI’s strength in survival prognosis, marked by...'). This should be cleaned up.","section":"Results: PathLUPI advances survival prognosis"},{"comment":"The phrase 'depicted in pretraining strategies Figure 6b revealed that no single foundation model was universally optimal' is grammatically broken and appears to contain a stray heading fragment. The sentence should be rewritten.","section":"Results: Validating the impact of foundation model embeddings"},{"comment":"The ethics statement should clarify whether the local institutional review board approvals cover the use of de-identified WSIs from Center-1 and Center-2, and whether the same approvals cover both the private cohorts and the public TCGA/CPTAC/EBRAINS data.","section":"Methods: Patient cohorts and ethics"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the benchmark is a real asset. The main risk is that the headline claim is causal but the decisive control experiment is missing; a permutation or no-privilege ablation is a standard and feasible fix. The statistical reporting also needs correction, especially the CI construction and the Wilcoxon test specification. I would not reject, but I would not accept without the control experiment and statistical revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jin et al. report PathLUPI, a dual-branch LUPI framework that uses transcriptomic pathway signatures during training to produce WSI-only predictors. What's genuinely useful here is the scale: 49 tasks across 20 cohorts, with external validation and consistent point estimates. The architecture—privileged branch with pathway-level MLPs, distilled branch with pseudo-pathway reconstruction, shared cross-attention—is a sensible recipe, and the interpretability analyses (CellViT++ cellular composition, Shapley pathway attribution) add value. This is a solid empirical contribution to computational pathology.\n\nThe soft spot is the one the stress-test flags, and it lands. The paper attributes the 4.48% mean AUC gain to the privileged transcriptomic supervision, but no experiment isolates that signal. Figure 6 only swaps the visual backbone; nothing removes or shuffles the transcriptomic branch. PathLUPI adds a region-aware re-embedding transformer, cross-attention, 50 pathway MLPs, and three alignment losses. A shuffled-transcriptome control (or an equal-capacity WSI-only model with the same modules minus the omics) is necessary before 'genome-anchored' causality is supported. Without it, the point estimates are credible but the mechanism is not established.\n\nThere are two smaller issues. The 95% CIs are implausibly tight (e.g., 0.849, 0.848–0.850); they appear to be bootstrap intervals on aggregated out-of-fold predictions, which ignore resampling variance across fits. Report fold-level or repeated-run variance. And the Code Availability section says the code is on GitHub but gives no URL—that should be fixed.\n\nThe benchmark itself is valuable and the method is worth engaging. I would not cite the causal claim as it stands, but the empirical recipe and the task collection are useful. This deserves a serious referee: the missing control is addressable, and the corrected statistics are manageable.\n\nRecommendation: send to peer review with a required control ablation and variance-corrected CIs. The work is coherent, clearly presented, and the central question—whether privileged transcriptomics helps—is important.","headline":"A well-executed LUPI benchmark for WSI molecular prediction; the causal claim is not isolated without a no-privilege control, and the CIs are overconfident, but the scale and consistency warrant serious refereeing.","tokens_in":31016,"tokens_out":2658,"would_cite":false,"duration_ms":28117,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PathLUPI, a training-time privileged-information framework, claims that anchoring whole-slide-image embeddings to transcriptomic pathway signatures improves slide-only molecular prediction across 49 oncology tasks, with a mean AUC gain of…","keywords":["computational pathology","whole-slide images","learning using privileged information","transcriptomic pathway signatures","Hallmark gene sets","biomarker prediction","molecular subtyping","survival prognosis"],"falsifier":"Train PathLUPI twice on the same tasks: once with the privileged branch's gene-expression vectors randomly permuted across patients and once with the branch removed but the WSI branch given equal parameters and the same alignment losses. If either variant keeps the reported AUC and C-index, the transcriptomic supervision is not causally load-bearing; if both drop toward the WSI-only baselines, the paper's attribution is supported.","tokens_in":29988,"feed_emoji":"🧬","tokens_out":11854,"duration_ms":111577,"temperature":0.7,"pith_summary":"The paper sets out to prove that routine H&E-stained whole-slide images alone can carry enough molecular information to predict mutations, subtypes, and survival, provided the histology model is trained with genomic guidance it will not see at inference. PathLUPI does this by giving the model, during training, a second input: the patient's bulk RNA-seq profile split into 50 Hallmark pathways, then forcing a slide-only branch to imitate the pathway-aware representations. Across 49 tasks, 20 cohorts, and 11,257 cases, the slide-only branch consistently beats models trained without the privileged transcriptomic signal, including a mean 4.48% AUC improvement over the best baseline in biomarker prediction. The paper also argues the resulting attention maps reveal tissue-level morphological signatures tied to specific mutations and pathways, which would make the method useful for hypothesis generation as well as prediction. If this holds, molecular-style screening could be delivered from the existing pathology workflow rather than from sequencing.","feed_headline":"Transcriptomic training lifts slide-based biomarker AUC by 4.48%","feed_subtitle":"PathLUPI embeds pathway gene expression as training-only guidance, beating slide-only baselines on 49 oncology tasks.","key_machinery":"The load-bearing object is a dual-branch learning-using-privileged-information architecture with a shared cross-attention block. The privileged branch encodes each patient's gene expression as 50 pathway-level vectors, one per Hallmark gene set, each produced by a dedicated multilayer perceptron, and lets those vectors attend over re-embedded WSI patch features, producing pathway-aware slide representations. The distilled branch constructs pseudo-pathway vectors from the WSI features alone, passes them through the same shared cross-attention, and is aligned to the privileged branch by three losses: reconstruction, attention-map alignment, and representation consistency. A region-aware re-embedding transformer refines patch features before attention. At inference the privileged branch is discarded, and the distillation branch alone produces genome-anchored embeddings from WSI input.","core_discovery":"PathLUPI's central claim is that transcriptomic data should be treated not as another prediction target but as privileged training information. During training a privileged branch fuses pathology-foundation patch features with RNA-seq embeddings organized by 50 Hallmark pathways through cross-attention, while a distilled branch sees only the whole-slide image and is trained, through reconstruction, attention-alignment, and representation-consistency losses, to reproduce the privileged branch's outputs. At inference only the distilled branch runs, so predictions use the slide alone. The paper reports that this design beats four WSI-only baselines on nearly all of 25 biomarker tasks, 8 subtyping tasks, and 16 survival tasks, with a mean AUC gain of 4.48% over the strongest baseline in biomarker prediction, an internal mean subtyping AUC of 0.856, an internal mean survival C-index of 0.693, AUC at least 0.80 in 14 biomarker and subtyping tasks, and smaller but consistent gains on external cohorts. It further claims the learned attention is interpretable: high-attention patches are enriched in cell types and pathways known to accompany specific mutations such as BRAF in colorectal cancer and EGFR in lung adenocarcinoma.","pith_inferences":["Treating the framework as a recipe, the same privileged-pathway design could be applied to methylation, proteomics, or spatial transcriptomics as the privileged modality, since nothing in the architecture is RNA-specific beyond the per-pathway encoders.","The attention maps could be tested as a cheap pre-screen for spatial transcriptomics: tissue regions PathLUPI weights most heavily should, if the biological claim is right, be the same regions where the corresponding pathway programs are spatially enriched.","A permutation control, where gene-expression vectors are shuffled across patients while all other training conditions stay fixed, would isolate whether pathway content or merely the presence of a second input drives the reported AUC gains.","The same dual-branch distillation could transfer other structured priors into slide encoders, such as copy-number signatures or tumor microenvironment cell fractions, to produce slide-only predictors for endpoints not tested here."],"forward_implications":["If PathLUPI's claim is right, a deployed system can predict actionable biomarkers, molecular subtypes, and prognosis from the standard H&E slide after a one-time training phase that needs paired transcriptomics, removing sequencing cost and turnaround time from the routine workflow.","The external-cohort results imply the genome-anchored embeddings transfer across institutions and staining conditions better than WSI-only baselines, so the benefit should not be confined to the training distribution.","The interpretability analyses imply that slide-only attention can localize tissue regions whose morphology tracks specific pathway activities, giving pathologists a visual hypothesis-generation tool for genotype-phenotype links.","The backbone ablation implies that future pathology foundation models with better pretraining will directly improve PathLUPI's molecular prediction, making the framework a compounding rather than one-off gain."],"supporting_citations":[{"why":"Supplies the learning-using-privileged-information paradigm that lets omics guide training but not inference.","marker":"[20]"},{"why":"Supplies the Hallmarks-of-Cancer framework used to structure transcriptomic profiles into biological pathway groups.","marker":"[22]"},{"why":"Defines the Hallmark gene-set collection that splits gene expression into the 50 pathways used by PathLUPI.","marker":"[79]"},{"why":"Provides the pathology-specific patch encoder whose embeddings PathLUPI refines, and one of the backbones compared in the ablation.","marker":"[17]"},{"why":"Provides the region-aware re-embedding transformer used to refine patch features before cross-modal attention.","marker":"[78]"},{"why":"Provides ABMIL, one of the four WSI-only baselines that PathLUPI must outperform.","marker":"[39]"},{"why":"Provides CLAM, a WSI-only baseline and the toolbox used for patching and attention heatmap generation.","marker":"[40]"},{"why":"Provides TransMIL, the strongest baseline in internal survival prognosis and a core comparator.","marker":"[42]"},{"why":"Supplies the gradient-based Shapley value approximation used to identify influential pathways in interpretability analyses.","marker":"[58]"}],"fun_headline_variants":["PathLUPI uses RNA as privileged training to beat slide-only models","Genome-anchored embeddings raise slide biomarker AUC by 4.48%","Privileged transcriptomics sharpens molecular prediction from WSIs","Training with RNA improves image-only oncology predictions","Slide-based molecular prediction gains from genomic guidance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the privileged transcriptomic branch, and not the extra architecture, losses, or parameters, is what causes the reported gains, yet the paper never trains PathLUPI without transcriptomic input or with scrambled transcriptomic profiles.","fun_headline_variants_meta":{"raw":{"variants":["PathLUPI uses RNA as privileged training to beat slide-only models","Genome-anchored embeddings raise slide biomarker AUC by 4.48%","Privileged transcriptomics sharpens molecular prediction from WSIs","Training with RNA improves image-only oncology predictions","Slide-based molecular prediction gains from genomic guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000409,"raw_usage":{"total_tokens":2152,"prompt_tokens":1004,"completion_tokens":1148,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1063}},"tokens_in":620,"tokens_out":1148,"duration_ms":8901,"temperature":1.0,"reasoning_tokens":1063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:28:27.535194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PathLUPI twice on the same tasks: once with the privileged branch's gene-expression vectors randomly permuted across patients and once with the branch removed but the WSI branch given equal parameters and the same alignment losses. If either variant keeps the reported AUC and C-index, the transcriptomic supervision is not causally load-bearing; if both drop toward the WSI-only baselines, the paper's attribution is supported.","supporting_citations":[{"cited_title":"& Vashist, A","cited_arxiv_id":null,"evidence_quote":"Supplies the learning-using-privileged-information paradigm that lets omics guide training but not inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Hallmark gene-set collection that splits gene expression into the 50 pathways used by PathLUPI."},{"cited_title":"Y .et al.A visual-language foundation model for computational pathology.Nat","cited_arxiv_id":null,"evidence_quote":"Provides the pathology-specific patch encoder whose embeddings PathLUPI refines, and one of the backbones compared in the ablation."},{"cited_title":"InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11343–11352 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the region-aware re-embedding transformer used to refine patch features before cross-modal attention."},{"cited_title":"Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images","cited_arxiv_id":null,"evidence_quote":"Provides CLAM, a WSI-only baseline and the toolbox used for patching and attention heatmap generation."},{"cited_title":"neural information processing systems34, 2136–2147 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides TransMIL, the strongest baseline in internal survival prognosis and a core comparator."},{"cited_title":"& Yan, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-based Shapley value approximation used to identify influential pathways in interpretability analyses."}],"review_version":2}