Mutational Signatures and Machine Learning for Risk Stratification of Acute Myeloid Leukaemia Based on Targeted Sequencing Data

Background/Objectives: To date, no validated scoring system can accurately predict the responses of acute myeloid leukaemia (AML) patients to induction chemotherapy (CTX). Current risk assessment relies on complex cytogenetic and molecular abnormalities and focuses on mutations in genes considered fundamental to leukaemogenesis. Methods: We performed bioinformatic analysis of targeted sequencing (TS) data from 111 genes in 1552 AML patients, focusing on mutational patterns derived from single-nucleotide variant (SNV) catalogues. The SNV catalogues were analysed using non-negative matrix factorisation (NNMF), a linear dimensionality-reduction approach, to extract risk-defining recursive signatures (RSs) and to distinguish responders from resistant patients following induction CTX. To enable patient-level prediction, we complemented NNMF with a Random Forest (RF) model. Given the class imbalance between responders and resistant cases, model performance was improved by applying the Synthetic Minority Over-sampling Technique (SMOTE) and by incorporating germline variants alongside somatic mutations. Results: NNMF-derived RSs captured clinically relevant structures in patients’ mutational profiles and clustered patients by treatment response, indicating that the diagnostic targeted sequencing data contain sufficient information for risk stratification and treatment response prediction. At the single-patient level, RF models incorporating balanced data and germline variation improved predictive performance compared with unbalanced somatic-only models. Conclusions: These findings demonstrate that machine learning applied to targeted sequencing data can extract clinically informative mutational structures and improve risk stratification in AML, supporting its potential integration into precision treatment decision-making.