Large Language Model Morphology-to-Genotype Inference for Microsatellite Instability From Colorectal Carcinoma Histopathology Reports

Microsatellite instability (MSI) is an important prognostic and predictive biomarker in colorectal carcinoma and is associated with several characteristic histomorphological features. While large language models (LLMs) have been investigated for structured data extraction from histopathology reports, their ability to infer molecular genotype from text-based morphologic descriptions remains underexplored. This study evaluated structured data extraction, MSI prediction performance, confidence behavior, and biological reasoning quality of two LLMs using colorectal carcinoma pathology reports with The Cancer Genome Atlas (TCGA)-assigned MSI status as the reference standard. Fifty colorectal carcinoma reports, including 25 MSI-high (MSI-H) and 25 microsatellite-stable (MSS), from TCGA database were stratified by morphologic signal strength. ChatGPT 5.2 (OpenAI, San Francisco, CA) and Gemini 3 (Google, Mountain View, CA) were given an identical zero-shot prompt to extract 17 histopathologic variables across structural, staging, invasion, and MSI-associated morphologic categories, predict MSI status with confidence scores, and provide biological justification. Both models demonstrated high extraction accuracy for structural variables (85.0-88.5%) but substantially lower performance for MSI-associated morphologic variables (ChatGPT 5.2: 39.4%; Gemini 3: 43.9%), with mucinous differentiation showing the highest omission rate and signet ring cell morphology the lowest extraction accuracy. Overall, MSI prediction accuracy was 64.0% for Gemini 3 and 16.0% for ChatGPT 5.2. Among committed predictions, ChatGPT 5.2 demonstrated perfect specificity and positive predictive value, whereas Gemini 3 showed a morphologic signal-dependent accuracy gradient and a substantially higher rate of high-confidence incorrect predictions (20.0% versus 2.0%). General-purpose LLMs demonstrated variable ability to recognize MSI-associated histomorphologic descriptors, with systematic failures in extraction directly limiting downstream genotype inference.