3614 Background: Neoadjuvant chemoradiotherapy (nCRT) is standard for locally advanced rectal cancer (LARC), yet responses are heterogenous and early identification of responders remains challenging. We developed a Transformer-based large language model that analyzes fragment-level patterns in circulating cell-free DNA (cfDNA) to predict pathological complete response (pCR) and postoperative recurrence risk independently of somatic mutation calling. Methods: A total of 510 plasma samples were collected from 102 LARC patients at five perioperative time points. The model was trained on post-nCRT cfDNA data from a training cohort (N = 62) to predict pCR, evaluated in an internal validation cohort (N = 40) and an external validation cohort (N = 96). Associations between model prediction scores, nCRT response, and recurrence-free survival (RFS) were assessed across time points. Fragment-level features contributing to model predictions were analyzed to provide biological interpretability. Results: The model demonstrated robust predictive performance, with AUCs of 0.909 (95% confidence interval CI: 0.812-1.000) in the training cohort, 0.903 (95% CI: 0.785-1.000) in the validation cohort, and 0.835 (95% CI: 0.754-0.916) in the external cohort. Patients achieving pCR had significantly higher prediction scores than non-pCR patients (Wilcoxon p < 0.001), and scores inversely correlated with tumor regression grade (Jonckheere-Terpstra test p < 0.001). At the predefined cutoff corresponding to 95% specificity in the training cohort, sensitivities were 0.73, 0.64, and 0.44 across datasets. High post-nCRT and post-surgical prediction scores were associated with improved RFS (log-rank p = 0.012 and p = 0.024, respectively), and all relapsed patients had consistently low post-nCRT scores. The top 1% of cfDNA fragments, ranked by model importance, were predominantly 50-100 bp with ~ 40% GC content, and the proportion of subnucleosomal particles was significantly higher in pCR versus non-pCR samples, providing biological insights into features driving the model. Conclusions: We developed a robust, mutation-independent framework for predicting nCRT response and postoperative recurrence risk in LARC, supporting individualized treatment decisions and post-surgical surveillance strategies.
Chen et al. (Wed,) studied this question.