Jeroen de Ridder

UMC Utrecht

In short, Jeroen de Ridder's research entails the following:

MaLMeC: Machine-Learned DNA Methylation Classification to enable tumor subtyping from liquid biopsies

Central Nervous System tumors are a rare and often lethal disease where accurate diagnosis plays a major role in making the right treatment choice. The precise molecular subtype of CNS tumors can be determined based on DNA methylation profiling. This will indicate the origin and mechanism of the particular tumor, guiding the treatment strategy.

However, accessing the tumor is often impossible without surgery, as these tumors usually reside inside the skull or spinal cord. This poses a challenge for neurosurgeons, as they have to operate on a tumor without knowing its specific type necessitating intraoperative diagnostics in which surgical samples are rapidly processed, and results are conveyed back to the operating room while the surgery is still ongoing.

We have recently developed ‘Sturgeon’, and AI application to sequence and classify tumor samples within 90 minutes, allowing exact molecular classification during surgery. Promisingly, tumor DNA fragments can also be obtained from cerebrospinal fluid. Our aim is to develop a variation of Sturgeon that is able to leverage the information from such circulating tumor DNA fragments, and to enable tumor classification from liquid biopsies (ie. Lumbar punctions or ventricular drain fluid).

A major advantage of this approach is that pre-operative tumor classification will be possible,allowing the surgical plan to be planned accordingly. The major challenge to be overcome is that the CSF contains a mixture of tumor and normal DNA. By leveraging real and simulated data we will train AI models to account for this mixture. In the four-year duration of this project we will both retrospectively and prospectively sequence at least 200 liquid biopsies from CNS tumor patients and strive for clinical implementation of this method.

Short summary of progress/results May 2026

The goal of this project is to create MaLMeC: a machine learned method for CNS tumor type classification from nanopore-based methylation sequencing of cfDNA and tissue biopsies. An important limitation in the treatment of CNS tumors is getting an accurate subtype diagnosis. Subtype diagnosis is currently achieved through a biopsy or surgical resection. This process increases the risk of comorbidity and limits the possibilities for neoadjuvant treatment. A liquid biopsy approach would alleviate these problems by providing a pre-operative diagnosis. However, these samples are known to have an extremely low tumor fraction (cerebrospinal fluid (CSF) or blood plasma).

Therefore, we aim to develop a deconvolution method that can handle this low purity to provide an accurate pre-operative subtype diagnosis. The initial results confirm that these nanopore samples have very dynamic levels of healthy cfDNA present. We have developed Tarpon, a classifier specifically designed to classify CSF samples. On our initial cohort of > 70 samples, Tarpon is currently able to provide a correct diagnosis in over half of the samples, while very rarely misdiagnosing samples. Current efforts are aimed at increasing the number of samples with a confident correct diagnosis and optimising sample collection procedures.

Short summary of progress/results May 2025

The goal of this project is to create MaLMeC: a machine learned method for CNS tumor type classification from nanopore-based methylation sequencing of cfDNA and tissue biopsies. An important limitation in the  treatment of CNS tumors is getting an accurate subtype diagnosis. Subtype diagnosis is currently achieved through a biopsy or surgical resection. This process increases the risk of comorbidity and limits the possibilities for neoadjuvant treatment. A liquid biopsy approach would alleviate these problems by providing a pre-operative diagnosis. However, these samples are known to have an extremely low tumor fraction (cerebrospinal fluid (CSF) or blood plasma).

Therefore, we aim to develop a deconvolution method that can handle this low purity to provide an accurate pre-operative subtype diagnosis. The initial results confirm that these nanopore samples have very dynamic levels of healthy cfDNA present. We leveraged information to create a comprehensive framework to simulate the cfDNA signal we get from CSF samples. The simulated samples were used to train and test both deep-learning and classical machine learning approaches for deconvolution. We already see a 30% increase in classification accuracy when compared to classification methods developed for tissue biopsy samples.