Development and External Validation of a Transcriptome-Based Multivariable Prediction Model for Treatment-Free Remission in Chronic Myeloid Leukemia
Menée à l'aide d'une approche utilisant un algorithme d'apprentissage automatique et d'une analyse transcriptomique de cellules sanguines issues de 96 patients puis validée sur 70 patients supplémentaires, cette étude identifie une signature basée sur l'expression de 50 gènes pour prédire le maintien de la rémission après l'arrêt d'un traitement par inhibiteur de tyrosine kinase
Purpose : Treatment-free remission (TFR) is a major therapeutic objective in chronic myeloid leukemia (CML). However, nearly 50% of patients relapse after tyrosine kinase inhibitor (TKI) discontinuation, and no robust predictive biomarker is currently available.
Methods : We profiled peripheral blood cell transcriptomes at imatinib (IMA) discontinuation in patients from the multicenter STIM2 trial (n = 96) to develop a transcriptome-based model predicting TFR by 2 years. A DESEQ2-based machine learning approach was compared with classical machine learning algorithms. The signature was then externally validated in an independent real-world cohort of patients attempting IMA or nilotinib cessation (n = 70). The biologic processes associated with the signature were further explored.
Results : We identified a 50-gene signature discriminating patients with sustained 2-year TFR from those experiencing molecular relapse (area under the receiver operating characteristic curve [AUROC], 0.83 [95% CI, 0.73 to 0.93] and 0.75 [95% CI, 0.55 to 1.00] in the training and internal validation cohorts, respectively). The discriminative performance was confirmed in the external test cohort, both as a binary predictor of 2-year TFR (AUROC, 0.71 [95% CI, 0.58 to 0.83] overall; 0.77 [95% CI, 0.61 to 0.92] in IMA-treated patients) and as a time-to-event predictor (log-rank P = .0042). The high TFR-signature group showed a higher proportion of myeloid immune cells and natural killer T cells, with an enrichment in Hedgehog signaling, whereas the low TFR-signature group demonstrated a higher proportion of lymphoid cells with an enrichment in mTOR signaling and a trend for oxidative phosphorylation activation. T-cell receptor and immunoglobulin heavy-chain repertoire analyses showed significantly greater polyclonality in the high TFR-signature group.
Conclusion : These findings demonstrate that transcriptomic profiling at TKI discontinuation can predict TFR outcomes in patients with CML and provide biologic insights into the mechanisms underlying sustained TFR.
Journal of Clinical Oncology , résumé, 2026