1053 - Explainable Machine Learning to Identify Chronic Lymphocytic Leukemia Based on Gut Microbiome Data

Background
Recent studies have shown extensive crosstalk between our immune system and gut microbiome (GM). The host immune system plays a vital role in the maintenance of GM homeostasis by establishing a balance between eliminating invading pathogens and promoting the growth of beneficial microbes, whereas GM can modulate the immune system by cytokine production. As chronic lymphocytic leukemia (CLL) is, among others, characterized by high rate of infectious complications and an altered immune system, it is warranted to uncover the pattern of microbiome composition distinguishing CLL from other hematological malignancies and healthy cohorts. Thus, we aimed to investigate whether machine learning model would be able to classify patients with CLL based on microbiome and demographic data.

Methods
A Danish cohort including a total of 438 feces samples collected from different patient groups and healthy donors were analyzed. 87 samples were collected from patients with CLL, 161 from patients with hematological malignancies such as acute myeloid leukemia (AML), myelodysplastic syndrome (MDS), or severe aplastic anemia, 121 from patients undergoing major thoracic surgery, 58 from healthy individuals, and 11 from kidney donors. Feces samples of healthy individuals were chosen to match the CLL population with respect to age, demographic data, sample collection method, and sequencing platform. Taxonomical profiling was done using an in-house bioinformatics pipeline and the resulting dataset was randomly split into a training set with 80% used as an input for machine learning algorithms and 20% of individuals stratified by CLL in the blinded test set for which results are reported.

Results
Preliminary results suggested that age, Coprococcus comes and Costridium boltae were the 3 most informative features to identify CLL (Fig 1). Age had the highest contribution to the identification of CLL and was positively correlated with its SHAP values, which translates to increased chance of identifying CLL as age increases. Coprococcus comes was the second most contributing feature with its values being negatively correlated with SHAP values. In other words, the chance of identifying CLL decreases as the relative abundance of Coprococcus comes increases. Costridium boltae as the third most contributing feature followed a similar pattern seen in Coprococcus comes, only with less difference in relative abundances. The predictions were retrieved from the best performing model (LightGBM) achieving the area under the receiver operating characteristics curve, AUC = 0.71 with specificity = 0.92, sensitivity = 0.44, positive predictive value = 0.62, and negative predictive value = 0.86 on the test set.

Conclusion
The presented results are preliminary data and thus we plan to further investigate the interaction between age and bacterial abundances in the included cohorts to avoid spurious correlations and predictions – updated analyses will be presented. Nonetheless, the results suggest that there exists a strong potential in using microbiome data for identifying disease states such as CLL. Incorporating cohorts with diverse hematological disease severities (AML, MDS, and aplastic anemia) alongside a healthy cohort, we aimed to elucidate the genuine microbiome pattern associated with CLL. By also encompassing other pathologies and considering the microbiome's potential reflection of health status, we believe we have accounted for a comprehensive representation. We observed several bacterial species being both negatively and positively correlated with identification of CLL, which offers a highly encouraging opportunity for exploring and implementing interventions aimed at modulating the microbiome through introduction of beneficial bacteria and/or elimination of pathogens. This knowledge could potentially lead to targeted interventions aimed at modulating the composition and function of the microbiota to improve disease outcomes for CLL patients.