Microsatellite Instability (MSI)

MSI is fundamentally caused by a loss-of-function in the DNA mismatch repair (MMR) machinery. Key genes involved include MLH1, MSH2, MSH6, and PMS2. When these proteins are dysfunctional—often due to germline mutations (Lynch syndrome) or somatic MLH1 promoter hypermethylation—errors introduced during DNA replication, particularly in repetitive microsatellite tracts, are not corrected. This leads to length polymorphisms and a hypermutator phenotype, driving tumorigenesis through frameshift mutations in coding regions of tumor suppressor genes.

Tumors are stratified into three categories based on the instability of a standard panel of five microsatellite markers (Bethesda panel):

• MSI-High (MSI-H): Instability in ≥2 markers (≥30% of loci). Characteristic of Lynch syndrome and ~15% of sporadic colorectal cancers.
• MSI-Low (MSI-L): Instability in only 1 marker.
• Microsatellite Stable (MSS): No markers show instability. Clinically, the binary distinction of MSI-H/dMMR vs. MSS/pMMR is the most actionable, directly guiding immunotherapy decisions.

Deep learning models, particularly Multiple Instance Learning (MIL) frameworks like CLAM, can now predict MSI status directly from routine Hematoxylin and Eosin (H&E) stained whole slide images. These models bypass the need for expensive and time-consuming molecular testing (PCR or NGS). The AI learns to identify subtle morphological correlates of the hypermutator phenotype, such as tumor-infiltrating lymphocytes (TILs), Crohn's-like lymphoid reactions, and poor differentiation, which are strongly associated with MSI-H tumors.

MSI-H/dMMR is the first pan-cancer, tissue-agnostic biomarker approved by the FDA for immune checkpoint inhibitor therapy. The high mutational burden generates abundant neoantigens, provoking a robust immune response that is held in check by upregulated PD-1/PD-L1 pathways. Pembrolizumab and nivolumab are approved for unresectable or metastatic MSI-H/dMMR solid tumors regardless of tissue of origin. This makes accurate MSI screening via AI a critical tool for expanding access to life-prolonging immunotherapy.

Universal MSI testing (or IHC for MMR proteins) is recommended for all newly diagnosed colorectal and endometrial cancers to screen for Lynch syndrome, the most common hereditary cancer predisposition syndrome. Lynch syndrome carriers have a germline mutation in an MMR gene and face an 80% lifetime risk of colorectal cancer. AI-driven MSI prediction from H&E slides offers a cost-effective, scalable pre-screening tool that can flag high-risk patients for confirmatory germline testing, streamlining genetic counseling workflows.

AI models predicting MSI status from histology are not black boxes; they consistently focus on specific morphological features. Key attention regions include:

• Intratumoral and peritumoral lymphocytes: Dense immune infiltration.
• Tumor budding: Invasive front architecture.
• Mucinous or medullary differentiation: Histological subtypes.
• Poor tumor differentiation: Lack of gland formation. These features, quantified by the model's attention mechanism, provide an interpretable link between the MSI genotype and its visible histological phenotype.

The cellular proofreading system that corrects base-pairing errors during DNA replication. Loss of function in key MMR proteins—most commonly MLH1, MSH2, MSH6, and PMS2—results in the accumulation of replication errors, particularly in repetitive microsatellite regions. This deficiency is the root cause of the MSI phenotype. Immunohistochemistry (IHC) staining for these four proteins is the standard clinical assay for dMMR status, and AI models trained on H&E slides learn to infer their functional status from morphological patterns.

An inherited autosomal dominant cancer predisposition syndrome caused by germline mutations in MMR genes. It is the most common hereditary cause of colorectal and endometrial cancers. Universal tumor screening for MSI or dMMR is recommended for all newly diagnosed colorectal cancers to identify patients who should undergo germline genetic testing. AI-based MSI screening from routine histology offers a cost-effective, scalable pre-screening tool that can flag high-risk patients without requiring additional molecular assays.

MSI-high tumors produce abundant neoantigens due to their hypermutated state, making them highly immunogenic and responsive to immune checkpoint inhibitors like pembrolizumab (anti-PD-1). The FDA granted the first tissue-agnostic approval for pembrolizumab based on MSI status alone, regardless of tumor origin. This therapeutic implication makes accurate MSI classification a critical clinical decision point, and AI-driven detection from H&E slides can accelerate treatment eligibility determination.

The historical gold standard for MSI detection, analyzing a panel of five Bethesda markers (BAT-25, BAT-26, D2S123, D5S346, D17S250). Instability in two or more markers classifies a tumor as MSI-High. While precise, PCR requires matched normal tissue, dedicated molecular laboratory infrastructure, and specialized expertise. Deep learning models trained on whole slide images offer a tissue-sparing, infrastructure-light alternative that can be deployed at scale in routine pathology workflows.

A quantitative measure of the total number of somatic mutations per megabase of tumor genome. MSI-High tumors typically exhibit elevated TMB, and both biomarkers independently predict immunotherapy response. AI models can simultaneously predict MSI status and TMB from H&E-stained slides, demonstrating that morphological features encode multiple genomic phenotypes. This multi-task learning capability makes computational pathology a powerful tool for comprehensive biomarker screening from a single digital image.

Training MSI prediction models requires only slide-level labels (MSI-High vs. Microsatellite Stable) rather than expensive pixel-level annotations. Multiple Instance Learning (MIL) architectures aggregate features from thousands of unlabeled tissue patches to produce a single slide-level classification. Key approaches include:

• Attention-based MIL: Learns to weight diagnostically relevant regions
• Graph neural networks: Models spatial relationships between patches
• Transformer-based aggregation: Captures long-range tissue architecture dependencies This weak supervision paradigm enables training on large, routinely collected clinical cohorts.