E. Moradi, J. Vuorinen, T. Hoikka, A. Hartewig, L. Helin, A. Rodriguez-Martinez, M. Pekkarinen, M. Vulli, S. Lehtipuro, S. Ampuja, A. Makinen, V. Fey, F. Tabaro, A. De Koker, R. V. Paemel, B. De Wilde, N. Callewaert, M. E. L. Kuusisto, H. R. Teppo, O. Kuittinen, H. Haapasalo, K. Nordfors, M. Nykter, J. Haapasalo, J. Kesseli, K. J. Rautajoki
Background DNA methylation-based classification has become an integral component of central nervous system (CNS) tumor diagnostics in neuro-oncology. However, current classifiers would benefit from improved interpretability, cross-platform generalizability, and scalability in routine clinical practice. Methods We developed a hybrid feature selection and machine-learning framework to derive compact, biologically relevant DNA methylation feature sets for CNS tumor classification. Variance-based filtering, intra- and inter-class consistency assessment, and elastic-net logistic regression were combined to identify informative CpG regions. A linear support vector machine (SVM) classifier was trained to distinguish methylation classes using microarray data and applied for sequencing data after imputing missing values. Selected tumor classes were differentiated with a few informative classifying features. Results The framework identified 1,003 informative genomic regions enriched for enhancer elements and neurodevelopmental pathways. In large external validation cohorts profiled by DNA methylation arrays (n = 1,993), the classifier achieved an accuracy of 0.96. The low-dimensional feature sets supported the investigation of diagnostically challenging cases and improved differentiation of histologically similar tumor entities, like embryonal tumors, using as few as two discriminative CpG features. Robust classification was preserved with bisulfite-equivalent targeted methylation sequencing and untargeted Nanopore-sequencing data. MGMT promoter methylation state and off-target read-derived genome-wide DNA copy number profiles provided supportive information. Conclusions Consistent DNA methylation patterns combined with SVM enable accurate and interpretable CNS tumor classification across sequencing platforms and provide a clinically scalable framework for next-generation neuro-oncology diagnostics.