Abstract 6906: Test-time compute for subtype classification in pediatric T-cell acute lymphoblastic leukemia using transcriptomics.

Abstract Background: T-cell Acute Lymphoblastic Leukemia (T-ALL) represents a highly aggressive hematologic malignancy characterized by profound molecular and clinical heterogeneity. While current T-ALL subtype classifications may not yet be universally integrated into clinical decision-making, the potential for accurate and rapid classification holds significant promise for future prognostic stratification and guiding targeted therapies. Traditional classification methods can be time-consuming and labor-intensive, highlighting the need for efficient, data-driven approaches. Leveraging transcriptomic data alone, we aimed to overcome these limitations by developing a novel approach for precise T-ALL subtype classification. Methods: We developed a machine learning pipeline for T-ALL subtype classification from RNA-seq data, introducing test-time compute paradigm. Initially, a Random Forest model was trained on a large discovery cohort (Polonen et al., 2024; 1,112 samples, 24,619 genes) to identify top predictive features by selecting the top 100 genes per subtype. At prediction time, the pipeline dynamically executes: (1) filtering preselected features from both training and test datasets (e.g., TARGET cohort, 264 samples, 22,688 genes); (2) batch correction via pycombat; (3) re-training of the Random Forest classifier on the processed training set; and (4) prediction on processed test data. This test-time pipeline is accessible via a Streamlit WebApp and Command Line. Results: We used F1 score to evaluate the performance of our classifier. Random Forest on initial training with discovery cohort achieved 93% classifying TAL-like, TLX-like, NKX2-1, ETP-like and 'other' subtypes. Selecting the top 100 features from each subtype yielded 478 features. Benchmarking the performance of our model using these selected features resulted in a 96% F1 score. We then applied our test-time pipeline on the independent TARGET cohort with 264 samples. The final model achieved 83% accuracy across 5 subtypes. Performance was highest for clinically relevant subtypes: TAL-like (F1: 0.96), TLX-like (F1: 0.94), and NKX2-1 (F1: 0.89). The “other” category scored moderately (F1: 0.60), while ETP-like was absent in the validation set. Conclusion: Our RNA-seq-based model delivers robust, scalable T-ALL subtype classification by leveraging test-time compute, integrating dynamic batch correction and real-time model retraining with a curated 478-gene feature set. Strong performance on an independent cohort highlights its clinical utility as a rapid, reliable tool for precision oncology. Citation Format: Tarun Karthik Kumar Mamidi, Irina Pushel, Byunggil Yoo, Midhat S. Farooqi, Keith J. August, . Test-time compute for subtype classification in pediatric T-cell acute lymphoblastic leukemia using transcriptomics abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6906.