Ultra-fast deep-learned classification algorithms for diagnosing pediatric CNS and solid tumors

Biography

Dr. L. A. Lennart Kester has been a clinical molecular biologist in pathology (KMBP) in the diagnostic laboratory at the Princess Máxima Center since February 2020, where he focuses on the analysis of RNA sequencing and whole-exome sequencing as part of routine diagnostic care. Besides this, he is responsible for the development and implementation of novel techniques in routine diagnostics. His current focus is on the development of an RNA expression-based paediatric cancer classifier that can differentiate between 162 different paediatric cancer types.

Upon completion of his training as a clinical molecular biologist in pathology, Lennart started his own research group within the Princess Máxima Center focusing on the development of novel diagnostic assays in paediatric oncology and is now co-supervising two PhD students.

Before 2020, he completed a PhD in the laboratory of Prof. Dr A. van Oudenaarden at the Hubrecht Institute, where he developed single-cell RNA sequencing and single-cell whole-genome sequencing techniques (graduated cum laude). After this, he worked as a postdoctoral fellow at the Netherlands Cancer Institute, developing novel machine learning approaches to select the most optimal chemotherapy regimen based on bulk RNA sequencing data from breast cancer.

Abstract

Cancer is the leading disease-related cause of death in children and young adults. The five-year survival rate for these tumors varies significantly based on tumor type and location, ranging from over 90% for Wilms tumors to less than 40% for malignant rhabdoid tumors. A correct and timely diagnosis is essential in order to start the optimal treatment. The initial diagnosis is established on the basis of morphology and immunohistochemistry. However, an increasing number of tumor entities require molecular profiling of the tumor to reach a definitive diagnosis, based on which a treatment protocol is chosen. This molecular profiling can take up to 14 days, thereby delaying the start of treatment. We previously developed and published Sturgeon, a deep-learned algorithm, that allows for ultra-fast central nervous system (CNS) tumor classification during surgery based on shallow nanopore whole-genome sequencing (Vermeulen et al. 2023). This method provides an accurate molecular diagnosis for patients with a brain tumor within two hours of acquisition of the material, while the patient is still in surgery. This allows the surgeon to adapt the surgical strategy based on the molecular diagnosis of the tumor ensuring an optimal balance between the extent of resection and risk of neurological damage and comorbidity. Using the same methodology we now developed Tucan, a deep-learned algorithm that allows for the classification of pediatric solid tumors. This algorithm not only allows for the classification of 80 pediatric solid tumor types (including over 40 types of sarcoma), but also generates an accurate copy number profile, including focal amplifications, within hours. This allows for immediate start of the optimal treatment protocol for the 10–20% of the pediatric patients for which immediate action is warranted. Both Sturgeon and Tucan rely on methylation calls from nanopore sequencing data and require only 0.05x to 0.1x coverage of the genome. This not only makes Sturgeon and Tucan extremely fast, it also makes them very cost-effective methods for tumor classification, highlighting their potential for use in low- and middle-income countries. We have now validated Sturgeon both retrospectively and prospectively and found that it provides the correct classification in over 95% of samples within two hours of acquisition of the material from the patient. Sturgeon is now implemented in standard clinical care and used routinely to provide a molecular diagnosis during surgery. Tucan performance is similar, with over 95% of samples receiving the correct classification. Combined, Tucan and Sturgeon ca...