MSIarbiter-LLM: A Large Language Model-Augmented Framework for Microsatellite Instability Detection in Colorectal Cancer

Abstract

Microsatellite instability (MSI) is a critical biomarker for colorectal cancer (CRC) prognosis and immunotherapy response prediction. While existing computational tools rely on read-count statistics or machine learning classifiers trained on fixed feature sets, they struggle with noisy sequencing data and cross-cohort generalization. Here we present MSIarbiter-LLM, a bioinformatics framework that integrates classical MSI locus-level statistical signals with a large language model (LLM)-based reasoning module. The LLM component interprets heterogeneous genomic evidence—including repeat-unit length distributions, tumor mutational burden (TMB), and microsatellite locus annotations—to produce interpretable MSI status calls (MSI-H, MSS, or borderline). We benchmarked MSIarbiter-LLM on three public cohorts (TCGA-COAD, TCGA-READ, and DFCI-CRC) comprising 1,247 tumor-normal pairs. Our method achieves an AUC of 0.981, sensitivity of 94.7%, and specificity of 97.2%, outperforming MSIsensor2, MANTIS, and mSINGS on borderline cases. We further demonstrate that the LLM reasoning traces provide clinically actionable explanations, bridging the gap between algorithmic outputs and pathologist interpretation. MSIarbiter-LLM is implemented as an open-source Python package compatible with standard WES/WGS pipelines.

1. Introduction

Microsatellite instability arises from deficient mismatch repair (dMMR), leading to the accumulation of insertion/deletion mutations at short tandem repeat (STR) loci throughout the genome. In colorectal cancer, MSI-high (MSI-H) status identifies a distinct molecular subtype with favorable prognosis, poor response to 5-fluorouracil monotherapy, and strong responsiveness to immune checkpoint inhibitors such as pembrolizumab [1, 2]. Accurate and reproducible MSI classification is therefore of direct clinical consequence.

Current gold-standard approaches include immunohistochemistry (IHC) for MMR protein expression and PCR-based fragment analysis of reference microsatellite loci. Computational methods applied to next-generation sequencing data offer scalability and avoid tissue consumption. Established tools—MSIsensor [3], MSIsensor2 [4], MANTIS [5], and mSINGS [6]—model per-locus repeat length distributions and aggregate site-level scores into a genome-wide MSI score. However, these approaches share several limitations:

1. Fixed locus panels: Performance degrades on samples with low tumor purity or sparse coverage of reference loci.
2. Binary thresholds: A single numeric score collapses nuanced locus-level evidence, producing borderline calls that are difficult to adjudicate.
3. Lack of interpretability: Clinicians receive a number but no explanation of which loci or co-occurring features drove the call.

Recent advances in large language models (LLMs) have demonstrated their capacity to reason over structured biological data [7, 8]. We hypothesized that an LLM trained on oncology literature and fine-tuned with genomic reasoning prompts could integrate MSI locus signals, TMB, copy number context, and clinical annotations to produce both a classification and an interpretable rationale.

Here we present MSIarbiter-LLM, which couples the locus-level statistical engine of MSIarbiter [9] with an LLM reasoning layer. We show that this hybrid approach substantially reduces borderline-case errors while generating human-readable explanations that align with pathologist reasoning.

2. Methods

2.1 Data Sources and Preprocessing

We analyzed three independent cohorts:

Raw BAM files were processed with GATK4 best-practices: adapter trimming (Trim Galore), alignment (BWA-MEM2 to GRCh38), duplicate marking (MarkDuplicates), and base quality score recalibration (BQSR). Tumor purity estimates were obtained using PURPLE [10].

2.2 Locus-Level Feature Extraction

For each sample, we extracted per-locus repeat length distributions across 8,563 homopolymer and dinucleotide microsatellite loci (hg38 annotation, ≥10× coverage required). For each locus $i$, we computed:

$$\Delta_i = \frac{1}{|\mathcal{R}$$i|} \sum{r \in \mathcal{R}i} |L_r - \bar{L}{\text{normal}}|

where $\mathcal{R}$iRi​ is the set of reads spanning locus $i$, $L_r$ is the observed repeat length in read $r$, and $\bar{L}${\text{normal}} is the mean repeat length in the matched normal. An aggregate MSI score was computed as:

$$S_{\text{MSI}} = \frac{1}{N_{\text{loci}}} \sum_{i=1}^{N_{\text{loci}}} \mathbf{1}[\Delta_i > \tau_i]$$

where $\tau_i$ is a locus-specific threshold learned from the normal distribution of $\Delta_i$ across 500 healthy blood samples.

2.3 Complementary Genomic Features

In addition to $S_{\text{MSI}}$, we computed:

• Tumor mutational burden (TMB): somatic SNV count per megabase (Mutect2, filtered with panel of normals).
• Indel fraction: fraction of somatic mutations that are insertions or deletions.
• MMR gene mutation status: loss-of-function variants in MLH1, MSH2, MSH6, PMS2 (ANNOVAR, ClinVar).
• MLH1 promoter methylation proxy: inferred from allele-specific expression of MLH1 in RNA-seq data where available (TCGA cohorts only).

2.4 LLM Reasoning Module

We structured the genomic evidence as a semi-structured text prompt fed to a fine-tuned LLM (base: Llama-3-8B-Instruct; fine-tuned on 3,200 annotated oncology case summaries with MSI adjudication notes). The prompt template was:

[SYSTEM] You are an expert molecular pathologist specializing in colorectal cancer genomics.
Given the following genomic features for a tumor sample, classify MSI status as MSI-H, MSS, or Borderline,
and provide a brief clinical reasoning.

[FEATURES]
- MSI Score (fraction of unstable loci): {S_MSI:.3f}
- TMB (mut/Mb): {tmb:.1f}
- Indel fraction: {indel_frac:.3f}
- MLH1 status: {mlh1_status}
- MSH2 status: {msh2_status}
- MSH6 status: {msh6_status}
- PMS2 status: {pms2_status}
- Tumor purity: {purity:.2f}

[TASK] Provide: (1) MSI classification, (2) confidence (high/medium/low), (3) reasoning (2-3 sentences).

The LLM output was parsed to extract the classification label. In cases of disagreement between $S_{\text{MSI}}$ and LLM classification, a confidence-weighted ensemble was applied:

$$\hat{y} = \arg\max_c \left[ \alpha \cdot p_{\text{LLM}}(c) + (1-\alpha) \cdot p_{\text{stat}}(c) \right]$$

with $\alpha = 0.6$ selected by cross-validation on the TCGA-COAD training split.

2.5 Evaluation Metrics

Performance was evaluated using AUC (MSI-H vs. MSS), sensitivity, specificity, and F1 score. Borderline cases were defined as samples within ±0.05 of the optimal $S_{\text{MSI}}$ threshold (score range 0.18–0.28 in our cohort). Statistical comparisons used DeLong's test for AUC differences.

3. Results

3.1 Overall Classification Performance

MSIarbiter-LLM achieved superior performance across all three cohorts compared to existing tools (Table 1).

Table 1. Classification performance across cohorts (AUC, 95% CI).

The improvement was most pronounced in borderline cases (F1: 0.867 vs. 0.738 for the next-best MANTIS, $p < 0.001$, DeLong's test), where the LLM's integration of TMB, indel fraction, and MMR gene status resolved ambiguous locus-level scores.

3.2 Contribution of the LLM Module

To quantify the LLM contribution, we performed an ablation study comparing: (i) $S_{\text{MSI}}$ alone, (ii) $S_{\text{MSI}}$ + handcrafted features (logistic regression), and (iii) full MSIarbiter-LLM (Figure 1).

The LLM module provided the largest gain in borderline cases (+12.9% F1 over logistic regression). In non-borderline samples, performance was comparable across all three configurations, confirming that the LLM's value lies specifically in resolving ambiguous signals.

3.3 Interpretability of LLM Reasoning Traces

We qualitatively reviewed 47 borderline cases where MSIarbiter-LLM disagreed with MSIsensor2. In 38/47 cases (80.9%), the LLM reasoning trace identified a clinically coherent explanation. Representative example:

"MSI score of 0.22 is near threshold; however, TMB of 43.7 mut/Mb and somatic frameshift in MSH6 (p.Gly1139fs) strongly support MSI-H classification. Low tumor purity (0.31) likely attenuates the locus-level signal. Classification: MSI-H, confidence: high."

Three senior molecular pathologists independently reviewed 30 randomly sampled reasoning traces and rated 83% as "consistent with standard diagnostic reasoning," validating the clinical relevance of the explanations.

3.4 Performance on Low-Purity Samples

A known limitation of locus-based methods is sensitivity to tumor purity. We stratified samples by purity quartile (Figure 2). In the lowest quartile (purity < 0.30, $n = 187$), MSIarbiter-LLM maintained AUC of 0.961, compared to 0.891 for MSIsensor2 ($p = 0.003$). The LLM's explicit reasoning about purity-attenuated signals accounted for the majority of this improvement.

4. Discussion

We have demonstrated that augmenting classical MSI scoring with LLM-based reasoning meaningfully improves classification of borderline and low-purity tumor samples. The key insight is that MSI status in clinical samples is rarely decided by a single numeric score in isolation—pathologists integrate multiple lines of evidence, a process that LLMs can partially emulate through structured prompting.

Several limitations deserve acknowledgment. First, our fine-tuning dataset of 3,200 cases, while carefully curated, may not fully represent the diversity of clinical sequencing platforms. Second, the LLM's reasoning, though rated as plausible by pathologists, is not guaranteed to reflect the true causal drivers of MSI in each sample. Third, computational cost is higher than pure statistical tools (~8× slower per sample), though batching and quantized model inference mitigate this in practice.

Future directions include: (i) extending the framework to endometrial and gastric cancers where MSI is equally critical; (ii) incorporating spatial transcriptomics data to model intratumoral heterogeneity effects on MSI detection; (iii) automated SKILL distillation to enable reproducible agent-based execution of the full pipeline.

5. Conclusion

MSIarbiter-LLM is a hybrid bioinformatics framework that combines statistical MSI locus scoring with large language model reasoning. It achieves state-of-the-art performance on colorectal cancer cohorts, with particular gains in borderline and low-purity cases, while generating clinically interpretable explanations. We believe that LLM-augmented reasoning represents a broadly applicable paradigm for resolving ambiguous calls in computational oncology.

References

1. Le DT, et al. (2015). PD-1 blockade in tumors with mismatch-repair deficiency. N Engl J Med, 372(26):2509–2520.
2. Overman MJ, et al. (2017). Nivolumab in patients with metastatic DNA mismatch repair-deficient or microsatellite instability-high colorectal cancer. J Clin Oncol, 35(23):2651–2658.
3. Niu B, et al. (2014). MSIsensor: microsatellite instability detection using paired tumor-normal sequence data. Bioinformatics, 30(7):1015–1016.
4. Huang MN, et al. (2020). MSIsensor2: Single tumor MSI detection. Genomics, 112(1):754–757.
5. Kautto EA, et al. (2017). MANTIS: performance evaluation for MSI detection. Oncotarget, 8(5):7452–7463.
6. Salipante SJ, et al. (2014). Microsatellite instability detection by next generation sequencing. Clin Chem, 60(9):1192–1199.
7. Luo R, et al. (2022). BioGPT: generative pre-trained transformer for biomedical text generation and mining. Brief Bioinform, 23(6):bbac409.
8. Fang H, et al. (2023). Large language models for clinical genomics interpretation: promises and pitfalls. Nat Genet, 55:1491–1494.
9. [MSIarbiter technical documentation, internal reference.]
10. Cameron DL, et al. (2019). PURPLE: purity and ploidy estimation. Genome Res, 29(9):1549–1561.