{"id":"2b572d56-b89a-40b3-97cb-887cf543bd0b","arxiv_id":"2506.15853","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HistoStainAlign uses contrastive alignment between paired H&E and IHC whole-slide embeddings to predict P53, PD-L1, and Ki-67 status from H&E slides alone, achieving moderate F1 scores on small internal datasets.","lead":"HistoStainAlign is a deep learning framework that predicts three immunohistochemistry (IHC) stain patterns from standard H&E-stained whole-slide images, using paired but unregistered IHC slides during training. It reports moderate weighted F1 scores (0.72 to 0.83) on small gastrointestinal and lung datasets, with statistically significant improvement over a fine-tuned baseline for only one of three stains.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported weighted F1 scores are close to the trivial majority-class baseline, and the paper's own significance tests contradict the claim of consistent improvement across all three tasks.","rationale":"The reader's verdict of CONDITIONAL already identifies several problems, including the missing majority baseline and the fact that only one of three tasks shows significant improvement. I agree with that overall assessment. However, I see the near-majority F1 problem as more load-bearing than the reader's stated weakest assumption about unregistered paired IHC slides. The unregistered-pairing concern is real, but it is a hypothesis about why contrastive training might underperform; the near-majority issue is directly visible from Table 1 and Table 2 without any additional assumptions. Even if the paired IHC signal were perfectly clean, the current evidence does not demonstrate consistent improvement, because the reported p-values and AUC ordering already contradict the consistency claim. The absolute F1 values also need a majority baseline before they can be interpreted as evidence of clinically useful prediction. I credit the authors for using patient-level fold splitting and for reporting confidence intervals and a cosine-similarity analysis, but these do not resolve the missing trivial baseline. The concrete check is inexpensive and can be run with the class counts already in the paper, so the verdict should remain conditional until such a baseline and corrected significance analysis are provided.","tokens_in":12556,"tokens_out":6697,"duration_ms":75888,"concrete_test":"Add an 'Always-Majority' baseline row to Table 2 computed directly from Table 1 class counts (weighted F1 equals the majority-class proportion for each task), and also report balanced accuracy and Matthews correlation coefficient for all models. Independently recompute the Wilcoxon comparisons on full test-fold prediction scores with a multiple-comparison correction across the three tasks. If the margins over the majority baseline remain approximately +0.002, +0.007, and +0.051 and only PD-L1 survives correction, the abstract and discussion should be revised to state that HistoStainAlign significantly improves PD-L1 prediction while showing non-significant trends for P53 and Ki-67.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that the headline weighted F1 scores are never compared with the trivial majority-class predictor, and the paper's own significance tests contradict the 'consistently outperformed' claim. From Table 1, P53 has 132/180 wild-type (73%), PD-L1 has 149/181 low (82%), and Ki-67 has 41/61 low (67%). An all-majority classifier therefore achieves weighted F1 = 0.733, 0.823, and 0.672, respectively. HistoStainAlign reports 0.735, 0.830, and 0.723, i.e., margins of +0.002, +0.007, and +0.051. The 'Statistical Comparison of Performance of Models' section reports Wilcoxon p=0.32 for P53, p<0.05 for PD-L1, and p=0.34 for Ki-67, and Table 2 shows HistoStainAlign AUC is lower than GigaPath-Finetuned for both P53 (0.726 vs 0.750) and Ki-67 (0.663 vs 0.672). Thus the only statistically supported improvement is PD-L1, and the absolute F1 numbers for P53 and PD-L1 are essentially indistinguishable from guessing the majority class. The central claim that 'the model consistently outperformed baseline and fine-tuned GigaPath variants across all classification tasks' is not supported by the reported evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HistoStainAlign is a contrastive learning framework that predicts IHC biomarker patterns (P53, PD-L1, Ki-67) from H&E-stained whole-slide images (WSIs). The method fine-tunes Prov-GigaPath using paired but unregistered H&E and IHC slides during training, optimizing inter-modality, intra-modality, and classification losses; at inference only H&E images are used. The authors evaluate on three in-house datasets (180, 181, and 61 WSIs) and report weighted F1 scores of 0.735, 0.830, and 0.723 for P53, PD-L1, and Ki-67, respectively, alongside comparisons with CONCH, UNI, and GigaPath baselines. They claim the proposed framework consistently outperforms baseline and fine-tuned GigaPath variants across all tasks.","tokens_in":12774,"tokens_out":4989,"duration_ms":56315,"significance":"The paper addresses a clinically relevant problem: predicting IHC status from H&E slides to reduce cost and turnaround time. The method leverages a modern foundation model (Prov-GigaPath), avoids patch-level annotations and tissue registration, and uses patient-level data splits with bootstrapped confidence intervals. The inclusion of a self-supervised variant and use of the public ACROBAT dataset for Ki-67 are positive aspects. Code and model are promised to be released. However, the reported evidence is substantially weaker than the claims: for two of the three tasks the improvement over GigaPath-Finetuned is not statistically significant, the absolute F1 gains over a trivial majority-class predictor are negligible (P53 and PD-L1), and the statistical test used is questionable. The framework is potentially useful, but the current manuscript overstates the results and needs a rigorous evaluation of baselines and significance.","major_comments":[{"comment":"The paper does not compare against a majority-class baseline. From Table 1, the class distributions yield majority-class weighted F1 scores of 0.733 for P53 (132/180 wild-type), 0.823 for PD-L1 (149/181 low), and 0.672 for Ki-67 (41/61 low). The reported HistoStainAlign F1 scores are 0.735, 0.830, and 0.723, i.e., margins of +0.002, +0.007, and +0.051. For P53 and PD-L1, these margins are within the reported confidence intervals and essentially indistinguishable from guessing the majority class. The authors must report a majority-class baseline and perform a significance test against it before claiming that the model has predictive value.","section":"Datasets / Table 1 and Model Performance / Table 2"},{"comment":"The Wilcoxon signed-rank test is applied only to 'prediction scores of true positive samples.' This is not a valid test for overall classifier comparison because it ignores true negatives, false positives, and false negatives. For example, a classifier that always predicts the majority class would have no true-positive scores for the minority class and could artifactually appear to perform well under this test. The authors should compare the full classification outputs (e.g., McNemar's test, DeLong test for AUC, or bootstrap over the full test set). The current test does not support the claim of significant improvement, including for PD-L1.","section":"Statistical Comparison of Performance of Models (p.17)"},{"comment":"The statement 'the model consistently outperformed baseline and fine-tuned GigaPath variants across all classification tasks' is directly contradicted by the authors' own results. The Statistical Comparison section reports p=0.32 for P53 and p=0.34 for Ki-67, and Table 2 shows HistoStainAlign AUC is lower than GigaPath-Finetuned for both P53 (0.726 vs 0.750) and Ki-67 (0.663 vs 0.672). The claims in the Abstract and Discussion must be revised to reflect that only PD-L1 shows a statistically significant improvement, and even that result rests on the questionable test described above.","section":"Discussion (p.20)"},{"comment":"The evaluation protocol is ambiguous: after training the model with the class-based loss, the authors state that 'a logistic regression model was fitted to the training set of embeddings.' It is unclear whether the reported metrics come from the trained classification head or from a separately fitted logistic regression on the slide embeddings. If the latter, the class-based loss and the classification head appear redundant and the end-to-end training procedure is not what is evaluated. The authors should clarify the exact evaluation pipeline, including how the validation and test sets are used with the logistic regression, and whether the logistic regression is trained on the same training fold used for the deep model.","section":"Evaluation (p.15)"}],"minor_comments":[{"comment":"The inter-modality loss formula uses symbols K+ and Q without defining them; the text defines q_i and k_i^+ but the cross-entropy arguments use matrix notation. Please define all variables, including the temperature τ and the batch composition.","section":"Loss Functions (Eq. 1)"},{"comment":"The text says '5-fold nested cross-validation,' but the described procedure (one 60/20/20 split per fold) is a standard 5-fold cross-validation, not nested cross-validation. Nested cross-validation requires an inner loop for hyperparameter tuning; if no such tuning was performed, please use the correct term.","section":"Evaluation (p.15)"},{"comment":"Table 3 shows that the paired cosine similarity of HistoStainAlign is lower than that of GigaPath-Finetuned for P53 (0.755 vs 0.901) and Ki-67 (0.753 vs 0.774), although the paired-versus-shuffled difference is larger. The manuscript emphasizes only the difference, which may mislead readers. The absolute values should be discussed, as they indicate that HistoStainAlign embeddings are less aligned in absolute terms, even if more discriminative.","section":"Cosine Similarity Analysis / Table 3"},{"comment":"There are several typos and formatting issues, e.g., 'staining patten samples' on page 8, 'ACROBAT23' in the datasets section, and inconsistent use of hyphens in 'HistoStainAlign-SSL.' A careful proofreading pass is recommended.","section":"Throughout"},{"comment":"The paper does not report the values of key hyperparameters (temperature τ, masking ratio, loss weights, batch size, learning rate, number of epochs). Providing these would improve reproducibility and allow readers to gauge the sensitivity of the method.","section":"Hyperparameters"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the methodology (using unregistered paired IHC slides for contrastive training) is interesting, but the current evidence does not support the strong claims. The missing majority-class baseline and the problematic significance test are load-bearing issues. If the authors add a majority baseline, use a proper statistical comparison, and substantially temper the claims, the paper could become publishable. The cosine similarity analysis is a nice sanity check but should be interpreted more carefully. I recommend major revision rather than reject because the underlying approach may still be valuable, but the current version oversells the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on 2506.15853. The framework is sensible: take Prov-GigaPath, fine-tune it with paired H&E and IHC slides using a TANGLE-style contrastive loss plus a masked-embedding consistency loss, and at inference use only H&E. That's a reasonable incremental extension of existing multimodal pathology work, and skipping registration and patch annotations is a real practical plus. The cosine-similarity analysis shows the contrastive loss does what it should: paired slides end up closer in embedding space than shuffled pairs. The experimental setup is also mostly careful: patient-level splitting, nested cross-validation, bootstrapped CIs, and a Discussion that honestly concedes the significance caveats. Credit where due.\n\nThe problem is the headline claim. The paper says the model 'consistently outperformed' baselines, but the authors' own Wilcoxon tests show only PD-L1 reaches p<0.05; P53 and Ki-67 are p=0.32 and p=0.34, and AUC is actually lower than GigaPath-Finetuned for both. More striking, the reported weighted F1 scores hug the majority-class floor: 0.735 vs 0.733 for P53, 0.830 vs 0.823 for PD-L1, 0.723 vs 0.672 for Ki-67. Only Ki-67 shows a margin worth discussing, and it is not significant. There is no majority-class baseline in the paper. That omission matters because the abstract and discussion present the results as if they demonstrate meaningful predictive improvement.\n\nA separate soft spot is the training-signal assumption: the paired H&E and IHC slides are unregistered, separate sections, and the framework assumes the IHC slide's global staining pattern is a reliable target for features learned from the H&E slide. The paper does not test how often that assumption breaks down. With small datasets (P53: 180, PD-L1: 181, Ki-67: 61 slides), the evidence is thin.\n\nThe paper is still worth a serious referee. The idea is coherent, the alignment results are plausible, and the limitations are at least partially acknowledged. But the version I read needs major revision: report majority baselines, give fold-wise variance, soften the consistency claim, and ideally make code and trained models available. I would not cite it in its current form; I might cite it after revision if the code ships and the evaluation is fixed. Bring it to a reading group as a case study in how easy it is to overstate gains that sit on the baseline floor.","headline":"Sensible contrastive training framework, but the headline IHC-prediction claim is undermined by missing majority baselines and the authors' own significance tests.","tokens_in":13416,"tokens_out":3300,"would_cite":false,"duration_ms":34777,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HistoStainAlign predicts P53, PD-L1, and Ki-67 IHC staining patterns directly from H&E whole-slide images by aligning paired slide embeddings during training.","keywords":["whole-slide images","immunohistochemistry prediction","H&E-stained pathology","contrastive learning","multimodal slide alignment","P53","PD-L1","Ki-67"],"falsifier":"Train HistoStainAlign with the same H&E slides but with IHC partners randomly shuffled across cases; if F1 scores on P53 and Ki-67 do not drop materially, then the paired-slide alignment is not the source of the gain and the method's central premise is unsupported.","tokens_in":12296,"feed_emoji":"🔬","tokens_out":11931,"duration_ms":108394,"temperature":0.7,"pith_summary":"HistoStainAlign is a deep learning training framework that predicts immunohistochemistry (IHC) biomarker status from ordinary H&E-stained whole-slide images, using paired IHC slides only during training. The paper shows that aligning H&E and IHC slide embeddings in a shared space with contrastive losses, plus a masked-embedding consistency loss, improves prediction of P53, PD-L1, and Ki-67 patterns over the same foundation model fine-tuned without those losses. On gastrointestinal and lung cohorts, the method reaches weighted F1 scores of 0.735, 0.830, and 0.723 respectively, and the PD-L1 gain over the fine-tuned baseline is statistically significant. The practical point is that a model trained this way could screen or prioritize cases for expensive IHC staining, reducing cost and turnaround time without requiring patch-level annotations or pixel registration between paired slides.","feed_headline":"Contrastive training lets H&E slides stand in for three IHC stains","feed_subtitle":"Aligning paired H&E and IHC slide embeddings during training, it reaches F1 0.735, 0.830, and 0.723 for three markers.","key_machinery":"The mechanical core is HistoStainAlign's three-loss training objective over slide-level embeddings produced by the Prov-GigaPath foundation model, a ViT tile encoder plus a LongNet-based slide encoder. The inter-modality loss is a symmetric contrastive loss adapted from the TANGLE framework: within a batch of paired H&E and IHC slides, each H&E embedding is matched to its paired IHC embedding against all others in the batch, and symmetrically each IHC embedding is matched back to its H&E partner. The intra-modality loss feeds the same H&E slide twice, once with 50% of patches randomly masked, and penalizes the mean squared error between the two slide embeddings, promoting local consistency. The class-based loss supervises the H&E slide embedding with the slide-level biomarker label. At inference, only the H&E slide is used; the IHC slide has disappeared.","core_discovery":"The central claim is that paired but unregistered whole-slide images of the same case can serve as a training signal that teaches a slide encoder to read molecular staining patterns from H&E morphology alone. HistoStainAlign takes the GigaPath tile and slide encoders, embeds both the H&E and the IHC slide of a case, and optimizes three losses together: a symmetric contrastive loss pulling paired H&E and IHC embeddings together while pushing unpaired pairs apart, a mean-squared-error loss forcing the slide embedding of an H&E slide to match its own embedding computed from 50% masked patches, and a supervised classification loss on H&E embeddings. Measured by nested five-fold cross-validation, the result is weighted F1 of 0.735 for P53, 0.830 for PD-L1, and 0.723 for Ki-67, with the PD-L1 improvement over GigaPath-Finetuned statistically significant and the others positive but not significant. The authors conclude that the contrastive alignment is responsible for the gain, supported by cosine-similarity analyses showing a larger gap between paired and shuffled cross-modality similarities than the fine-tuned baseline.","pith_inferences":["Extending beyond the paper, the unregistered-pair assumption could be stress-tested by using H&E and IHC sections cut from different blocks of the same tumor; if performance drops sharply when the tissue sections diverge, the model is relying on global case-level correlation rather than a stable morphology-to-molecule mapping.","Extending beyond the paper, the method's value for clinical triage would be clearer if the model produced spatial prediction maps from H&E and these were compared with the actual IHC stain distribution, since the current slide-level label cannot show whether the model learned localized correspondences or only aggregate case statistics.","A further extension would be to apply the same contrastive alignment at finer label granularity, for example regressing the continuous PD-L1 TPS or Ki-67 index instead of binarizing at 50% or 3%; the 1% PD-L1 positivity threshold used in clinical trials is a natural next target."],"forward_implications":["A model trained with HistoStainAlign can be run on H&E slides alone at test time, so the clinical workflow needs no additional staining, registration, or patch-level annotations.","For PD-L1 in lung resections, the framework's improvement over GigaPath-Finetuned is statistically significant, indicating that paired IHC slides during training can add real predictive signal beyond foundation-model fine-tuning.","The self-supervised variant pretrained on the public ACROBAT breast cancer Ki-67 dataset improves Ki-67 classification on gastrointestinal samples, suggesting the alignment objective transfers across tissue types.","Cosine-similarity measurements show HistoStainAlign embeddings separate matched from unmatched cross-modality pairs more strongly than the baseline, supporting the claim that the contrastive loss changes the embedding geometry as intended.","These results position the framework as a potential pre-screening tool that could prioritize cases for actual IHC staining rather than replace it."],"supporting_citations":[{"why":"Prov-GigaPath, the foundation model whose ViT tile encoder and LongNet slide encoder HistoStainAlign fine-tunes.","marker":"[14]"},{"why":"TANGLE, the transcriptomics-guided framework whose symmetric contrastive loss is adapted as the inter-modality loss.","marker":"[25]"},{"why":"ACROBAT, the public multi-stain breast cancer dataset used to pre-train the Ki-67 model with the self-supervised variant.","marker":"[23]"},{"why":"The open-source digital pathology library used to generate H&E tissue masks during patch extraction.","marker":"[24]"},{"why":"CONCH, the visual-language foundation model used as a zero-shot baseline for IHC prediction.","marker":"[26]"},{"why":"UNI, the general-purpose pathology foundation model used as a multiple-instance-learning baseline.","marker":"[27]"},{"why":"The PD-L1 IHC interpretation manual that defines TPS and supplies the 50% cutoff used for the binary label.","marker":"[16]"},{"why":"HEMnet, an earlier registration-based H&E-to-IHC molecular label transfer approach that HistoStainAlign contrasts with by avoiding registration.","marker":"[7]"}],"fun_headline_variants":["H&E-only AI predicts three IHC stains without patch labels","Cross-modality model maps H&E to three IHC biomarkers","Unregistered H&E and IHC pairs train stain predictor","Slide embeddings align to translate H&E into IHC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an unregistered IHC slide cut from a different physical section of the same case carries the molecular ground truth for the H&E slide, and that the one slide-level biomarker label applies uniformly across the entire slide.","fun_headline_variants_meta":{"raw":{"variants":["H&E-only AI predicts three IHC stains without patch labels","Cross-modality model maps H&E to three IHC biomarkers","Unregistered H&E and IHC pairs train stain predictor","Slide embeddings align to translate H&E into IHC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3865,"prompt_tokens":1115,"completion_tokens":2750,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":731,"completion_tokens_details":{"reasoning_tokens":2682}},"tokens_in":731,"tokens_out":2750,"duration_ms":20216,"temperature":1.0,"reasoning_tokens":2682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:48:32.000260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train HistoStainAlign with the same H&E slides but with IHC partners randomly shuffled across cases; if F1 scores on P53 and Ki-67 do not drop materially, then the paired-slide alignment is not the source of the gain and the method's central premise is unsupported.","supporting_citations":[{"cited_title":"https://www.nature.com/articles/s41586-024-07441-w","cited_arxiv_id":null,"evidence_quote":"Prov-GigaPath, the foundation model whose ViT tile encoder and LongNet slide encoder HistoStainAlign fine-tunes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ACROBAT, the public multi-stain breast cancer dataset used to pre-train the Ki-67 model with the self-supervised variant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The open-source digital pathology library used to generate H&E tissue masks during patch extraction."},{"cited_title":"Y., Chen, B., Williamson, D","cited_arxiv_id":null,"evidence_quote":"CONCH, the visual-language foundation model used as a zero-shot baseline for IHC prediction."},{"cited_title":"J., Ding, T., Lu, M","cited_arxiv_id":null,"evidence_quote":"UNI, the general-purpose pathology foundation model used as a multiple-instance-learning baseline."},{"cited_title":"PD-L1 IHC 22C3 pharmDx, Interpretation Manual, NSCLC 1% 50%","cited_arxiv_id":null,"evidence_quote":"The PD-L1 IHC interpretation manual that defines TPS and supplies the 50% cutoff used for the binary label."},{"cited_title":"J., Andor, N., Nguyen, Q","cited_arxiv_id":null,"evidence_quote":"HEMnet, an earlier registration-based H&E-to-IHC molecular label transfer approach that HistoStainAlign contrasts with by avoiding registration."}],"review_version":1}