Machine learning-based classification of chronic lymphocytic leukemia signalotypes to allows biological and clinical insights into novel CLL subgroups

Abstract Introduction: The mutational landscape of chronic lymphocytic leukemia (CLL) is characterized by a broad spectrum of recurrent genetic alterations, which can generally be mapped onto a limited set of key signaling pathways. However, the genetic programming within CLL cells alone is insufficient explain their survival and proliferation capabilities. Instead, interactions with the microenvironment are critical drivers of disease initiation, progression, and maintenance. Therefore, gaining a comprehensive understanding of the complex interplay between genetic abnormalities and microenvironmental signals holds promise for improving prognostic accuracy and tailoring therapeutic strategies. Signaling patterns derived from RNA expression analyses represent an integrative approach to capture the combined influence of both intrinsic genetic factors and extrinsic microenvironmental cues. Despite this potential, the individual variability in CLL cellular responsiveness at the signaling level has not been extensively explored to date. Methods: We utilized a large publicly available dataset generated by Knisbacher et al., employing a custom-built bioinformatics pipeline with rigorous quality control measures to identify differential signaling patterns across CLL samples. This approach enabled us to define so-called “signalotypes,” distinct signaling phenotypes characterized by unique pathway activation profiles. Subsequent analyses linked these signalotypes to clinical annotations within the dataset. We then applied advanced machine learning algorithms to capture the molecular signatures underlying each signalotype, allowing accurate classification of fully annotated CLL samples from our own biobank into the established signalotype clusters. We then conducted extensive in vitro toxicity screens using these biobank samples. We tested a broad range of compounds in solo culture conditions and selected subsets of substances under more complex coculture setups, simulating microenvironmental interactions. Results: Our bioinformatics analysis of RNA-seq from 661 chemo-naïve CLL patients from the large public dataset yielded a robust definition of 11 signalotypes, supported by multiple layers of validation. These groups exhibited marked differences in the activation status of signaling pathways and metabolic activity, highlighting underlying biological heterogeneity. Notably, the clusters demonstrated substantial divergence in clinical outcomes, including time to first treatment and overall survival, suggesting clinically relevant biological distinctions. Moreover, established molecular markers such as IgHV mutation status, cytogenetic abnormalities, and mutations in recognized CLL driver genes were non-randomly distributed among the signalotypes. We trained an ensemble of machine learning models on the molecular patterns of each signalotype, enabling the mapping of 101 individual patient samples from our biobank to their corresponding clusters with high confidence and performed high-throughput drug screening using 117 substances. We found considerable heterogeneity in drug response profiles across different signalotypes. Our findings suggests that our screening approach can effectively differentiate between broadly effective agents (like Venetoclax) and those with more selective activity tied to specific biological subgroups. More detailed analysis of 20 substances in coculture settings validated the high heterogeneity between signalotypes. Conclusions: Our study demonstrates that signaling patterns derived from RNA sequencing data can be used to stratify CLL patients into subgroups with distinct biological characteristics, clinical trajectories, and treatment responses. Machine learning techniques enable the extraction and recognition of core molecular features defining each signalotype, allowing accurate patient classification based on these signatures, also allowing to extract a minimum number of features that are essential to define the molecular core of a specific signalotype. By integrating these insights with functional drug screening, we uncover substantial variability in therapeutic susceptibility across CLL subgroups, which could inform more personalized treatment approaches.