Deep Neural Network Classifier Identifies Pediatric Brain Tumors From Liquid Biopsies

Researchers developed the methylation-based predictive algorithm for central nervous system tumors (M-PACT), a deep neural network classifier, to identify and classify brain tumors in pediatric patients from subnanogram-input cell-free DNA of methylomes in liquid biopsies. A report of the performance of M-PACT was published in Nature Cancer.

“This is a next-generation assay and computational framework that we’ve optimized and applied across a range of pediatric [patients with] brain tumors,” said corresponding author Paul A. Northcott, PhD, Director, Center of Excellence in Neuro-Oncology Sciences (CENOS), and Member, Department of Developmental Neurobiology, St. Jude Children's Research Hospital. “M-PACT is about taking liquid biopsy to another level in pediatric neuro-oncology and applying the technology across many different clinical scenarios.”

Liquid biopsies based on cerebrospinal fluid have demonstrated significant benefit in evaluating tumor-derived cell-free DNA in patients with central nervous system tumors, yet there have been limitations to the sensitivity of these liquid biopsies.

Researchers developed M-PACT to classify tumors based on their DNA methylation pattern from cell-free DNA in cerebrospinal fluid. The classifier incorporated CpG imputation, a tumor enrichment algorithm, and methylation-based deconvolution for more accurate specificity. They then analyzed the classification performance of the deep neural network in a benchmarking (n = 79) and in a validation (n = 58) cohort.

M-PACT achieved accuracy of 92% for classifying embryonal central nervous system tumors in the benchmarking cohort and 88% within the validation cohort.

The deep neural network also allowed for methylation-based cellular deconvolution and sensitive copy-number variations to be accurately detected in the cell-free DNA methylomes. M-PACT is also able to be used during treatment and follow-up to classify tumors as primary or secondary malignancies. “If a tumor reoccurs years later, M-PACT can reliably determine whether it’s a true relapse or a second malignancy,” Dr. Northcott said.

Smith et al added that “our findings support the use of this method in the preoperative setting, where CSF-based diagnosis may enable presurgical medical intervention, identification of actionable targets and more informed surgical approaches.”

The neural network also demonstrated utility in nonembryonal central nervous system tumors and in nonmalignant cerebrospinal fluid.

Dr. Northcott and his team are also exploring the potential of M-PACT beyond just pediatric brain tumors. “Although we applied this to pediatric brain tumors, it will clearly be useful in other solid tumors and hematological malignancies as well,” Dr. Northcott said. “The informatics will need to grow to classify the full scope of cancer types diagnosed in children, but we’ve developed something quite powerful that is likely to be more broadly adopted in the community.”

Model Methods

The researchers first created a CpG correlation network and a network-based regression diffusion model for imputed methylation values, which was trained on an array reference cohort (n = 914) and achieved high accuracy. The classifier was then trained on simulated real-world samples (n = 3,195,000) with differing levels of tumor burden.

“Traditionally, methylation-based diagnostics for circulating tumor DNA use classifiers designed for tumor tissue, which have higher DNA input,” said co-first author Katie Han, a PhD student in the St. Jude Graduate School of Biomedical Sciences and Department of Developmental Neurobiology and MD candidate at University of Tennessee Health Sciences Center. “We reversed the usual flow and designed M-PACT for circulating tumor DNA itself with applicability to tissue, instead of the other way around.”

Three neural networks were integrated into a three-layer ensemble model that achieved an accuracy of more than 0.9, the highest prediction probability score of all tested models.

“We developed M-PACT by computationally mixing large reference datasets with normal cell-free DNA datasets,” explained co-first author Kyle S. Smith, PhD, of the Department of Developmental Neurobiology, St. Jude Children's Research Hospital.

The first multilayer perceptron classifier was trained on a dense methylation cohort, and the same cohort was used to create a simulated training cohort of methylation profiles for the other two classifiers. The concordance between the neural networks ensured that fewer false positives would be identified by the resulting ensemble model.

The researchers enhanced the tumor signal and classification probabilities using beta regression from nine nonmalignant cell classes’ methylation signatures. Estimating tumor burden required amplifying and separating methylation profiles into malignant and nonmalignant fractions based on samples with tumor classifications scores over 0.7.

Additionally, the model incorporated a methylation-based binary classification algorithm for discerning between primary tumor or nonmalignant samples for methylomes in cerebrospinal fluid.

DISCLOSURE: This study was supported by the Verein unser_kind, the Forschungsgesellschaft für Cerebrale Tumore, Ein Kiwi gegen Krebs, the Mark Foundation for Cancer Research, St. Baldrick’s Foundation, the Brain Tumor Funders’ Collaborative, the National Cancer Institute, the FWF der Wissenschaftsfonds, the OeNB Jubiläumsfonds, the Physician Researcher Pathway Scholarships of the Medical University of Vienna, the City of Vienna Fund for Innovative Interdisciplinary Cancer Research, the CCP Starter Grant, the Oncomine Clinical Research Grant, the Emil Aaltonen Foundation, the Competitive State Research Financing of the Expert Responsibility Area of Tampere University Hospital, the Finnish Ministry of Social Affairs and Health, the Tampere University Foundation Trust, the Väre Foundation for Pediatric Cancer Research, the Foundation for Pediatric Research, the Robert Connor Dawes Scientific Fellowship of the National Brain Tumor Charity and the American Lebanese Syrian Associated Charities. For full disclosures of the study authors, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.