Predicting Severe Methotrexate-Induced Mucositis: Can AI Flag High-Risk Patients Early?

• A machine learning model (XGBoost) utilizing pathway-specific genetic variants successfully predicted high-dose methotrexate (HD-MTX) induced mucositis in pediatric ALL patients with an AUC of 0.76.

• The model demonstrated a sensitivity of 0.61 and specificity of 0.79, identifying high-risk patients better than clinical factors alone.

• Removing the genetic data from the model caused a significant performance drop (AUC fell to 0.61), proving that germline genetics—specifically in inflammatory and repair pathways—drive risk prediction more than routine clinical metrics like age or sex.

• Clinical Burden: Severe mucositis affects 20–40% of pediatric patients receiving HD-MTX, often necessitating tube feeding, opioid infusions, and prolonged hospitalization.

• Workflow Impact: For pharmacy leaders, this toxicity translates into unplanned resource utilization: urgent TPN compounding, complex pain management (PCA), and disruptions to chemotherapy scheduling that can impact survival.

• Current Gap: Standard pharmacokinetics (PK) monitoring focuses on drug clearance but often fails to explain why some patients with "normal" clearance still suffer severe tissue injury; this study suggests the answer lies in intrinsic tissue sensitivity rather than just drug levels.

• Cohort: The researchers analyzed 278 pediatric Acute Lymphoblastic Leukemia (ALL) patients (86 cases with grade ≥ 2 mucositis, 192 controls) treated with HD-MTX (≥ 1000 mg/m²) across six Canadian academic centers.

• Method: Instead of looking for single genetic "smoking guns," they used a pathway-informed approach. They curated 18 biological pathways relevant to mucosal injury (using PharmGKB, literature, and biological databases) and aggregated the effects of genetic variants within these pathways.

• Modeling: They fed these genetic features, alongside clinical variables (age, sex, radiation), into an XGBoost machine learning classifier to predict toxicity risk.

• Pathway Signals: The analysis identified two specific biological pathways significantly enriched for mucositis-associated variants: IL6 signaling (P=0.048) and WNT/β-catenin (P=0.041), both of which are critical for inflammation and epithelial repair.

• Key Genetic Drivers: The model prioritized specific genes within these pathways. Notable variants included PRKCD and CSNK1A1 (associated with protective effects) and PIK3R2 and AGT (associated with increased risk).

• Performance: The full model achieved fair-to-good discrimination (AUC 0.76). Crucially, when genetic data was removed (ablation analysis), the model's ability to predict mucositis collapsed (AUC 0.61), confirming that clinical factors alone are insufficient for risk stratification in this setting.

• Biology-Informed Triage: The study moves beyond black-box AI by linking risk to specific biological mechanisms (inflammation and tissue repair), which could eventually guide targeted prophylaxis (e.g., cryotherapy intensity or early nutritional support).

• Operational Readiness: While promising, this is a discovery-phase signal. There is currently no off-the-shelf CLIA-validated assay for this specific panel of variants, meaning it cannot yet be deployed for real-time clinical decision support.

Strengths:

• The use of pathway enrichment allowed the detection of cumulative genetic effects that traditional Genome-Wide Association Studies (GWAS) often miss.

• The inclusion of SHAP analysis provided transparency, allowing clinicians to see exactly which features drove the risk score for an individual patient.

Limitations:

• With only 278 patients, the study is relatively small.

• The results were validated internally (cross-validation) but lack external validation in an independent cohort, which is a mandatory step before clinical adoption.

This study provides a compelling "proof of concept" that pathway-based genetic profiling can predict HD-MTX mucositis better than clinical factors alone. However, pharmacy leaders should view this as a research signal rather than an immediate practice change. External validation and assay standardization are required before this can drive dosing or prophylaxis protocols.