AI Single-Cell Mismatch-Repair IHC Quantification in Colorectal Cancer

The diagnostic and clinical question

Mismatch-repair (MMR) status is one of the most consequential biomarkers in colorectal cancer. Loss of expression of one or more of the four core MMR proteins — MLH1, PMS2, MSH2, and MSH6 — identifies a tumor as mismatch-repair deficient (dMMR). This single finding carries several distinct clinical implications. It flags patients who may have Lynch syndrome, an inherited cancer-predisposition condition, prompting germline evaluation and cascade testing of relatives. It informs prognosis and adjuvant management in early-stage disease. And, importantly, dMMR status is one of the tissue markers used to establish eligibility for the immune checkpoint inhibitor class of therapies, which show particular activity in mismatch-repair-deficient tumors [1].

In routine practice, MMR status is most commonly assessed by immunohistochemistry (IHC), a stain that visualizes whether each of the four proteins is present in tumor-cell nuclei. Interpretation, however, is not trivial. The pathologist must distinguish genuine tumor-cell nuclear staining from the staining of surrounding non-neoplastic cells (lymphocytes, stroma, normal epithelium), judge the adequacy of internal positive controls, and recognize patterns such as heterogeneous, weak, or focal staining that can produce equivocal or discordant calls. Interobserver variability and technical artifact are recognized challenges. This is precisely the kind of pattern-recognition task — quantitative, repetitive, and anchored in single-cell discrimination — where computational image analysis has been proposed as an aid [1].

What the method is and how it works

The approach described by Nowak and colleagues, termed AIMMeR, applies deep learning to whole-slide images of MMR IHC in colorectal cancer, with the goal of quantifying staining at single-cell resolution and rendering an automated deficient-versus-proficient classification [1].

Conceptually, the pipeline must solve a layered problem. First it must localize and segment individual cells across the slide. Second, it must determine, cell by cell, whether the nucleus shows positive MMR protein expression. Crucially, a correct call depends on evaluating staining specifically within the tumor-cell population rather than in reactive or stromal cells — because a tumor can be MMR-deficient while the admixed lymphocytes and stroma retain normal (positive) staining and serve as the built-in internal control. By operating at the single-cell level, the method is designed to make this distinction explicit and quantitative, converting what is often a gestalt human judgment into a measurable proportion of stained tumor cells [1].

The output is twofold: a continuous, quantitative readout of staining and a thresholded binary classification (dMMR vs proficient MMR, pMMR) that can be compared directly against the reference-standard pathologist interpretation. The developers tuned the decision threshold to prioritize sensitivity — the clinically appropriate choice for a screening application, where missing a deficient tumor (a false-negative) has greater consequences than flagging a proficient one for review [1].

The evidence

The central validation reported by Nowak (2024) tested AIMMeR against expert pathologist interpretation using colorectal cancer material drawn from large clinical trial cohorts — the SCOT and QUASAR2 studies [1].

Across this validation, the model achieved an area under the receiver-operating-characteristic curve (AUROC) of 0.98 for discriminating dMMR from pMMR tumors relative to the pathologist reference standard [1]. An AUROC of 0.98 indicates very strong overall discrimination, meaning that across the range of possible decision thresholds the model separated deficient from proficient cases with high fidelity.

Because the intended use is screening, the more operationally meaningful figure is the performance at a high-sensitivity operating point. At a sensitivity of 98%, the model maintained a specificity of at least 75% [1]. In practical terms, if the tool were deployed to triage MMR IHC slides, this configuration would be expected to correctly flag approximately 98% of truly deficient tumors while correctly clearing at least three-quarters of proficient tumors — the remainder of proficient cases being referred for human review as presumptive positives. This is the characteristic trade-off of a screening instrument: sensitivity is deliberately maximized at some cost to specificity, so that the human expert's attention is concentrated on the cases most likely to be deficient or genuinely ambiguous.

It is worth being precise about what these metrics do and do not establish. The reference standard is the pathologist interpretation of the same IHC — so the study demonstrates strong agreement with expert human reading of the stain, validated on substantial trial-derived cohorts [1]. It is an evaluation of IHC interpretation, not a head-to-head comparison against orthogonal molecular assays such as PCR-based microsatellite instability testing or next-generation sequencing, and the provided evidence does not report those comparisons.

Where this stands

The evidence supports AIMMeR as a validated-against-pathologist, single-cell dMMR-IHC quantification method with excellent discrimination (AUROC 0.98) and a high-sensitivity operating point (≥75% specificity at 98% sensitivity) in colorectal cancer cohorts derived from the SCOT and QUASAR2 trials [1]. That is a meaningful result for a screening-oriented tool, and it illustrates the strongest current rationale for computational MMR assessment: reproducible, quantitative, scalable triage of a biomarker that every colorectal cancer specimen is expected to be evaluated for.

Several important caveats frame how this should be understood. First, status: the work described here is a research and validation study. The provided evidence does not establish regulatory clearance or approval, and the tool should be regarded as investigational for routine clinical deployment absent such qualification. Readers should not infer FDA-cleared status from these data.

Second, the comparison standard matters. Demonstrating high concordance with expert IHC interpretation confirms that the algorithm reproduces expert reading; it does not, on its own, resolve cases where IHC itself is misleading — for example, the recognized phenomenon in which MLH1 or other proteins may retain antigenic staining despite underlying functional deficiency, or technically false-positive staining. The provided reference does not characterize performance in those specific scenarios, and they remain open questions.

Third, generalizability requires attention. Performance validated on trial-derived material must be confirmed across the heterogeneity of real-world laboratories — different antibody clones, staining platforms, fixation practices, and scanners — before equivalent performance can be assumed in any given practice. This is a general principle of computational pathology validation and applies here.

Finally, the clinical framing should remain careful. dMMR status contributes to downstream decisions, including Lynch syndrome screening pathways and determination of eligibility for the immune checkpoint inhibitor therapy class [1]. An AI tool that identifies dMMR status informs those pathways; it does not itself make a Lynch syndrome diagnosis (which requires germline confirmation) nor direct any individual patient's treatment. The appropriate conception is that of a quantitative, high-sensitivity assistant to the pathologist — concentrating expert review and improving reproducibility — rather than an autonomous substitute for diagnostic judgment.

The evidence suggests that single-cell deep-learning quantification of MMR IHC is a promising and well-validated approach to a high-value biomarker task. Confirming its performance prospectively, across diverse laboratories, and against orthogonal molecular references are the natural next steps before such tools enter routine diagnostic workflows.

References

1. Nowak KM, et al. AIMMeR: single-cell deep-learning quantification of mismatch-repair immunohistochemistry in colorectal cancer; AUROC 0.98 versus pathologist, with specificity ≥75% at 98% sensitivity (SCOT/QUASAR2 cohorts). Cell Reports Medicine. 2024. PMID 39293403; DOI 10.1016/j.xcrm.2024.101727.