AI Liquid Biopsy Tool Classifies Pediatric Brain Tumors from Spinal Fluid

A blood draw or a sample of cerebrospinal fluid is, in theory, a far gentler path to diagnosis than a brain biopsy. The problem has always been that pediatric brain tumors release only tiny fragments of their DNA into body fluids - far less than adult solid tumors - leaving existing classification algorithms with too little material to work reliably. A new tool called M-PACT changes that calculation.

The system, described in Nature Cancer, was developed by scientists at St. Jude Children's Research Hospital working with teams at the Hopp Children's Cancer Center Heidelberg (KiTZ), the German Cancer Research Center (DKFZ), and collaborators across Europe and the United States. M-PACT stands for Methylation-based Predictive Algorithm for CNS Tumors. It applies a deep neural network to circulating tumor DNA (ctDNA) extracted from cerebrospinal fluid, identifying each tumor's type by reading its DNA methylation pattern - the chemical marks that regulate gene activity and tend to become distinctively abnormal in cancer.

Designed for the signal you actually have

Previous methylation-based classifiers were built for tissue samples, where DNA is plentiful. Applying them to cerebrospinal fluid - where the tumor contributes only a fraction of the total cell-free DNA present - produced unreliable results. The St. Jude team reversed the engineering logic entirely.

"We reversed the usual flow and designed M-PACT for ctDNA itself with applicability to tissue, instead of the other way around," said Katie Han, a co-first author and PhD student in the St. Jude Graduate School of Biomedical Sciences. "Traditionally, methylation-based diagnostics for ctDNA use classifiers designed for tumor tissue, which have higher DNA input."

The training dataset comprised more than 5,000 DNA methylation profiles spanning roughly 100 distinct tumor entities. Rather than simply feeding the network clean tumor-derived data, the researchers computationally mixed large reference datasets with normal cell-free DNA datasets - simulating the diluted, noisy signal that cerebrospinal fluid actually provides.

"We trained it extensively and showed that even tiny amounts of ctDNA can be accurately classified," said co-first author Kyle Smith, PhD, of St. Jude's Department of Developmental Neurobiology.

Benchmarking results: 92% accuracy across tumor types

In a formal benchmarking evaluation, M-PACT correctly identified 92% of brain tumors from cerebrospinal fluid alone - without any tissue sample. That figure covers a wide range of pediatric central nervous system tumor types, not a single narrowly-defined cohort. The test cohort drew on clinically annotated liquid biopsy samples gathered through an international network spanning institutions in Austria, Finland, Germany, the Netherlands, and the United States.

Beyond initial diagnosis, the system demonstrated usefulness across the treatment timeline. When a child's tumor recurs years after first treatment, a critical clinical question is whether the new lesion is a true relapse of the original cancer or a second, independent malignancy - the treatment implications are different. M-PACT can distinguish between those scenarios using cerebrospinal fluid, without requiring a new surgical biopsy.

"If a tumor reoccurs years later, M-PACT can reliably determine whether it's a true relapse or a second malignancy," said corresponding author Paul Northcott, PhD, director of the Center of Excellence in Neuro-Oncology Sciences at St. Jude.

Reading the tumor's neighborhood, not just the tumor itself

One of M-PACT's more unexpected capabilities comes from what it does with the non-tumor DNA in cerebrospinal fluid. In any liquid biopsy sample, the vast majority of cell-free DNA comes from sources other than the tumor - immune cells, neurons, blood cells that have crossed the blood-brain barrier. Most previous analyses treated this material as noise to be filtered out.

M-PACT instead uses it as additional information, predicting what fraction of the sample DNA originates from T cells, B cells, and other cell types that constitute the tumor microenvironment - the network of normal cells that cancers co-opt to support their growth.

"Most DNA in cerebrospinal fluid is from something else, the 'negative space' of the tumor, which we previously ignored," Smith said. "Now we can predict what fraction comes from T cells, B cells, or other sources."

This feature has particular value during treatment, when tumor tissue sampling is rarely performed. Therapy changes the composition of the microenvironment, sometimes dramatically - tracking those shifts noninvasively could eventually guide decisions about when to change drugs or escalate care.

"Now we can start to see how both the tumor and its microenvironment change with therapeutic pressure," Han said.

Scope and current limitations

M-PACT was validated on pediatric brain tumors, and the 92% figure reflects that specific context. The algorithm currently covers roughly 100 tumor entities - a meaningful number, but not the full catalog of childhood cancers. Expanding the reference database to cover more tumor types will require additional large, annotated sample cohorts, which depends on continued international collaboration.

The study does not yet include data from randomized clinical trials or prospective diagnostic evaluations comparing M-PACT directly against existing tissue-based standards in routine clinical practice. The current evidence base, while strong, is retrospective. Regulatory approval and clinical deployment would require further prospective validation.

Northcott acknowledged both the potential and the work still ahead. "The informatics will need to grow to classify the full scope of cancer types diagnosed in children," he said. "But we've developed something quite powerful that is likely to be more broadly adopted in the community."

The team also tested M-PACT's framework on tumor types beyond the central nervous system, suggesting the underlying approach - training for dilute ctDNA rather than tissue - may transfer to other solid tumors and hematological malignancies. Those applications remain to be validated in separate studies.

International team and funding

The study required a large, clinically annotated collection of pediatric liquid biopsy samples assembled across multiple countries. Co-senior authors include Johannes Gojo of the Medical University of Vienna, and Kristian Pajtler and Kendra Maass of KiTZ/DKFZ. Co-first author Tom Fischer also works at the KiTZ/DKFZ. Participating institutions span Finland, Germany, the Netherlands, Austria, and multiple St. Jude departments.

Funding came from a wide range of sources including the National Cancer Institute (P01CA096832; R01CA270785; R01CA259372), the Mark Foundation for Cancer Research, St. Baldrick's Foundation, the Brain Tumor Funders' Collaborative, and the American Lebanese Syrian Associated Charities (ALSAC).