Limits of Predicting Colorectal Cancer PDX Engraftment: A Specimen-Grouped Machine Learning Analysis of 1007 Establishment Attempts

Background: Patient-derived xenograft (PDX) models retain clinically relevant features of the donor tumor, but the establishment of colorectal cancer (CRC) PDX models succeeds in only a fraction of attempts. Whether routinely available clinicopathological and experimental variables can predict engraftment for a previously unseen tumor has not been established under a validation design that respects the clustered structure of PDX datasets. Methods: We retrospectively analyzed 1007 first-generation CRC-PDX establishment attempts from 381 specimens (328 patients). Primary analysis used specimen-grouped validation with nested grouped cross-validation; all preprocessing was performed within training folds. Secondary analyses included patient-grouped, attempt-level, specimen-level and temporal validation. Results: Attempt success was 46.3%; 68.8% of specimens yielded ≥1 PDX. Within-specimen correlation was substantial (0.31). Under specimen-grouped validation, pooled AUC was 0.550 (95% CI 0.508–0.591), PR-AUC was 0.500, Brier was 0.253, and sensitivity/specificity was 0.42/0.64. Patient-grouped validation was similar (0.548), while attempt-level splitting inflated AUC to 0.68–0.69, confirming bias from specimen clustering rather than preprocessing leakage. Specimen-level prediction reached 0.636; calendar-time temporal validation yielded 0.584. Sensitivity analyses were stable. Conclusions: Routine variables offer only marginal discrimination for CRC-PDX engraftment in new tumors. Moderate attempt-level performance reflects repeated sampling of the same tumors, not a generalizable signal. These results do not support routine-variable-based triage and identify molecular and tissue-quality factors as the necessary next step.