Immunohistochemistry (IHC) is one of the most important diagnostic tools in modern oncology. By applying antibodies that bind to specific proteins in tumor tissue, IHC reveals which proteins are expressed -- and in what amounts -- on or inside cancer cells. This information directly guides treatment: for example, the PD-L1 protein is expressed on the surface of certain cancer cells, and a patient's PD-L1 expression level determines eligibility for immunotherapy drugs like pembrolizumab. HER2 expression guides the use of targeted therapies in breast and gastric cancer.
As targeted therapies have expanded, so has the number of proteins that oncologists want to measure. New classes of drugs called antibody-drug conjugates (ADCs) and bispecific antibodies directly attack proteins on the tumor cell surface, and their effectiveness is linked to how much of that protein the tumor expresses. Quantifying this protein expression accurately -- the tumor proportion score (TPS), the percentage of positive tumor cells out of all tumor cells -- is increasingly critical to treatment decisions.
Currently, IHC scoring is done manually by pathologists, which is time-consuming, subject to inter-observer variability, and difficult to standardize across laboratories. Deep learning models have been developed to automate this scoring, but they have a fundamental limitation: each model must be trained separately on hundreds of annotated slides from a specific cancer type and specific antibody stain. A model trained to score PD-L1 in lung cancer cannot reliably score PD-L1 in bladder cancer, or score HER2 in breast cancer -- a limitation known as domain shift.
This study set out to develop a single Universal IHC (UIHC) analyzer -- a deep learning model trained on diverse cancer types and multiple antibody stains simultaneously -- that could accurately score IHC expression even for cancer-stain combinations it had never encountered during training. If successful, such a model would dramatically reduce the time and annotation cost needed to develop IHC scoring tools for new therapeutic targets.
The study trained eight distinct deep learning models that differed only in their training data composition. Single-cohort (SC) models were trained on one cancer type and one IHC stain -- for example, the P-L model trained on PD-L1 22C3 stained lung cancer slides, or the H-B model trained on HER2-stained breast cancer slides. These served as the benchmark, representing the current standard approach.
Multi-cohort (MC) models were trained on combinations of cancer types and IHC stains. The most comprehensive, designated PH-LUB, was trained on PD-L1 22C3 stained lung, urothelial, and breast cancers combined with HER2-stained breast cancers -- four cohorts spanning two antibody types and three cancer origins. All models were trained to accomplish the same task: detect tumor cells in each image patch and classify each as positive or negative for the IHC stain, then calculate the TPS across the entire slide.
Training was done at the patch level -- small 0.04 mm2 image tiles cut from whole slide images (WSI) -- because WSIs are too large to process as a single unit. Pathologists annotated each patch, marking individual tumor cells as positive (stained) or negative (unstained). After patch-level training, models were assembled to generate WSI-level TPS predictions that could be compared directly to pathologist-assigned TPS categories: below 1%, 1-49%, or 50% and above.
On test data matched to their training domains, single-cohort models performed well -- validating that the standard approach works within its intended scope. The P-L model (PD-L1 in lung) scored well on PD-L1 lung test slides; the H-B model (HER2 in breast) scored well on HER2 breast test slides. Multi-cohort models performed at least as well on these familiar test sets, confirming that expanding training data did not degrade performance within known domains.
The difference became pronounced on novel test sets -- cancer-stain combinations the models had never seen during training. For the challenging pan-cancer PD-L1 test set (22C3 stain across cancer types not in training), multi-cohort models trained with more cancer type diversity significantly outperformed the best single-cohort model (F1-score 0.722-0.745 for multi-cohort vs. 0.708 for the best single-cohort, p less than 0.001). On entirely novel antibody stains -- including MET, TROP2, HER3, DLL3, FGFR2, Claudin 18.2, and MUC16 -- multi-cohort models generally outperformed single-cohort models across all tested cancers.
The most revealing result came from MET staining in pan-cancer tissues: all four multi-cohort models significantly outperformed the best single-cohort model (F1-scores 0.762-0.795 vs. 0.744, all p less than 0.001). This is particularly important because MET is a drug target for therapies like amivantamab, and the ability to reliably score MET expression without training on MET-stained slides represents a genuine practical advance for oncology drug development.
At the whole-slide image (WSI) level -- where models must correctly classify each slide into TPS categories that directly correspond to clinical cutoffs -- the multi-cohort PH-LUB model was designated the UIHC model based on its consistently top performance across all test cohorts. It achieved a Cohen's kappa score of 0.578 and accuracy of 0.751 on the multi-stain pan-cancer test set. The best single-cohort model (H-B) reached only kappa 0.509 and accuracy 0.703 -- a meaningful gap at the level of clinical decision-making.
Cohen's kappa is a statistical measure of agreement that corrects for chance. A kappa of 0.578 reflects moderate-to-substantial agreement with pathologist scoring -- comparable to inter-pathologist agreement levels seen in clinical practice. On the specific 22C3 PD-L1 lung test set, the UIHC model was the only one to outperform the best single-cohort model (kappa 0.652 vs. 0.638), demonstrating that breadth of training did not compromise performance on the most clinically established stain type.
For novel antibody stains that lack established clinical scoring cutoffs, the UIHC model achieved an AUROC of 92.1% across a range of positivity thresholds -- outperforming all single-cohort models. This cutoff-independence is practically important because many emerging drug targets do not yet have consensus scoring systems, and a model whose performance is stable across different thresholds can support biomarker discovery before clinical standards are established.
To demonstrate practical utility, the researchers applied the UIHC model to a biological question: does the c-MET protein expression differ between lung cancers with MET gene amplification versus those with MET exon 14 splicing mutations -- two different genetic changes that theoretically both increase c-MET levels, but through different mechanisms? The UIHC model had never been trained on MET-stained slides, yet it was applied directly.
The UIHC model assigned TPS of 94.5% to MET amplification cases, compared to 77.1% for MET exon 14 skipping mutation cases and 75.7% for EGFR exon 20 insertion cases. This finding -- that amplification produces higher protein expression than splicing mutations -- aligns with known MET biology and is clinically relevant for deciding which patients might most benefit from large molecule drugs that target c-MET, such as amivantamab.
What makes this result significant is that the UIHC model quantified meaningful biological differences in a novel stain type without any MET-specific training data. This demonstrates the model's potential as a discovery tool for early drug development: pharmaceutical teams developing new antibody targets could use UIHC to quantify candidate protein expression in tumor archives without building a custom training set for each candidate -- a process that currently takes months of pathologist annotation time.
Pathological validation confirmed the UIHC model's accuracy in individual cases. In one lung cancer case stained for MET, a single-cohort model (P-L) incorrectly classified most tumor cells as negative (TPS 36%), while the UIHC model accurately identified positive cells (TPS 61%), closely matching pathologists' assessment of 75%. In a gastric cancer case stained for FGFR2, all single-cohort models failed, while the UIHC model correctly identified membranous staining even in complex tissue backgrounds.
The study's core contribution is demonstrating that training diversity drives generalization in IHC AI: exposing a model to more cancer types and more antibody stains during training produces a model that performs better on unseen combinations -- not just incrementally better, but meaningfully better where it matters most, in the novel domains representing new clinical and research needs.
The UIHC model's ability to assess novel stain types without additional annotation is practically transformative for oncology drug development. When a pharmaceutical company is developing a new antibody-drug conjugate targeting a cell-surface protein, the first question is: how much of that protein do tumors actually express? Currently, answering that question requires assembling hundreds of slides, having pathologists annotate individual cells for weeks, and training a new model. The UIHC approach would allow immediate quantification from existing slides without that pipeline.
The authors acknowledge current limitations: the model currently scores only tumor cells, not immune cells like lymphocytes and macrophages that are also clinically relevant for some IHC markers. It performs binary positive/negative classification rather than multi-level scoring required for some markers. And performance variability was observed across some staining techniques, suggesting that further expansion of the training cohort -- adding more stain types and more cancer origins -- is likely to continue improving performance. Future development also aims at multi-level protein quantification aligned with clinical scoring guidelines like ASCO/CAP HER2 criteria.