-Author name in bold denotes the presenting author
-Asterisk * with author name denotes a Non-ASH member
Clinically Relevant Abstract denotes an abstract that is clinically relevant.

PhD Trainee denotes that this is a recommended PHD Trainee Session.

Ticketed Session denotes that this is a ticketed session.

3336 Identifying 5-Hydroxymethylcytosine Signatures in Circulating Cell-Free DNA and Treatment Response in Multiple Myeloma with a Multi-Step Machine Learning Approach

Program: Oral and Poster Abstracts
Session: 653. Multiple Myeloma: Clinical and Epidemiological: Poster II
Hematology Disease Topics & Pathways:
Adult, Technology and Procedures, Study Population, Human, Machine learning, Omics technologies
Sunday, December 8, 2024, 6:00 PM-8:00 PM

Bei Wang1*, Benjamin A Derman, MD2, Zhou Zhang3*, Krissana Kowitwanich4*, John F. Cursio, PhD5*, Daniel Appelbaum6*, Chuan He, Ph.D.7*, Wei Zhang3*, Andrzej J Jakubowiak, MD, PhD2 and Brian Chiu, PhD1

1Department of Public Health Sciences, University of Chicago, Chicago, IL
2Section of Hematology/Oncology, Department of Medicine, The University of Chicago, Chicago, IL
3Northwestern University Feinberg School of Medicine, Chicago, IL
4Department of Chemistry, University of Chicago, Chicago, IL
5University of Chicago Biological Sciences Division, Chicago, IL
6Department of Radiology, University of Chicago, Chicago, IL
7Department of Chemistry, Department of Biochemistry and Molecular Biology, and Institute for Biophysical Dynamics, The University of Chicago, Chicago, IL

Background

Despite recent therapeutic advancements in multiple myeloma (MM), many patients do not achieve complete response after induction therapy, and currently there is no reliable way to predict treatment response. Robust 5-hydroxymethylcytosine (5hmC) markers can be sensitively detected in circulating cell-free DNA (cfDNA) in the peripheral blood, offering a promising approach to capture tumor heterogeneity without invasive bone marrow sampling. The 5hmC-modified marker genes have been associated with overall survival in MM and predict treatment response in diffuse large B cell lymphoma. To date, no study has investigated genome-wide 5hmC profiles in cfDNA underlying distinct response to induction therapy in MM.

Methods

CfDNA was extracted from plasma samples of 308 patients with newly diagnosed MM collected between 2010-17 at the time of study enrollment in the UChicago Myeloma Epidemiology Study. Patients were followed through 28 February 2022. We collected baseline demographic, clinical, laboratory, and treatment data from electronic medical records. We identified the best response achieved at the end of induction therapy (i.e., frontline therapy after diagnosis) according to the International Myeloma Working Group (IMWG) response criteria. We profiled genome-wide 5hmC in cfDNA using the 5hmC-Seal and next-generation sequencing. The 5hmC-Seal data were mapped to the human genome reference (hg19) and annotated to gene bodies. Data on response was available for 290 MM patients and were included in this analysis. The patients were randomly divided into a training set (n=203) and a testing set (n=87). Differentially 5hmC-modified genes were identified (p-value<0.05, log2 fold change>0.1) in the training set, followed by logistic regression with the elastic net regularization for feature selection through 10-fold cross validation. The optimal “alpha” and “lambda” were identified to maximize model performance measured by the area under the receiver operating characteristics (ROC-AUC). This selection step was repeated for 100 times in the training set, and genes appearing in at least 95% of iterations were identified as signature genes. These genes were used to fit a final model in the training set, and the coefficients were used along with enrichment levels of 5hmC to compute a weighted predictive 5hmC score (wp-score) of response in the testing set. Lastly, we assessed the association between the wp-score and response in the testing set.

Results

Of the 290 patients, 71 (24.48%) were classified as responders, achieving a complete response (CR) or better (CR, n=38; stringent CR, n=33). The remaining 219 patients (75.52%) were classified as non-responders, including those with very good partial response (n=82), partial response (n=99), stable disease (n=20), and progressive disease (n=18). Most clinical prognostic factors at diagnosis, such as stage and lactate dehydrogenase levels, were similar between the two groups, except that a greater proportion of the non-responders had an elevated serum free light chain (sFLC) ratio (54.25% vs. 33.82%; p<0.01). Compared with non-responders, responders had better overall survival (age-adjusted hazard ratio [aHR]: 0.49; 95% confidence interval [CI]: 0.30-0.81) and progression-free survival (aHR:0.41; 95% CI: 0.29-0.61) over a median of 70 months of follow-up. The wp-score was associated with response (i.e., CR or better) to induction treatment in both training (odds ratio [OR]: 11.59, 95% CI: [5.84-26.92]) and testing sets (odds ratio [OR]: 1.70, 95% CI: [1.04-2.92]), after controlling for age, sex, and sFLC ratio. Analyses according to classes of induction therapy (e.g., VRd, KRd, etc) are ongoing and results will be presented at the meeting. Several biological functions/pathways from the gene ontology (GO) database, such as chronic inflammatory response and peptidase activity, were enriched among the differentially 5hmC-modified genes associated with response to induction therapy.

Conclusions

Our preliminary findings indicate that the 5hmC marker gene-based wp-score derived from cfDNA at the time of diagnosis can predict response to induction therapy in MM. These findings enhance our understanding of the epigenetic contributions to therapeutic response to induction therapy in MM and provide a basis for a non-invasive approach for managing MM using epigenetic signatures.

Disclosures: Zhang: Tempus Labs: Consultancy.

*signifies non-member of ASH