Methylation classification
One epigenetic fingerprint, read on any platform.
The cancer epigenome
A molecular fingerprint of the tumour’s life history
The cancer epigenome carries two things at once: the changes that accumulate as a tumour grows, and the marks that reveal the cell it grew from. That combination makes it possible to identify where a cancer originated even after it has spread.
We read that fingerprint and match it against reference cohorts of tumours whose identity is already established. The result is a methylation class: a reproducible molecular identity that groups tumours by shared biology rather than by appearance alone. The approach complements microscopic examination, extends it, and in some entities supersedes it.
How it works
Machine-learning classification with calibrated confidence
Our machine-learning models are trained on cohorts of tumours whose identity is already established, learning the methylation patterns that separate one class from another. A new sample is then scored against every class in the reference set.
Profile
Genome-wide methylation data are generated in the laboratory on whatever platform it already runs, then normalised through a versioned preprocessing pipeline.
Classify
The profile is scored against every class in the reference set. Each prediction carries a calibrated score, so a confident call is distinguishable from an uncertain one.
Report
Alongside the class, the run returns a copy-number variation profile and QC metrics, and for CNS samples, MGMT promoter methylation status.
Calibrated scores matter more than raw accuracy: they let a laboratory decide when a prediction is strong enough to act on and when the sample needs a second line of evidence.
Platform-agnostic
Any methylation platform, including sequencing
There is no separate model per instrument. One classifier accepts input from array and sequencing platforms alike, so a laboratory is not locked into the technology the reference cohort happened to use, and results stay comparable as platforms change.
Arrays
Illumina Infinium MethylationEPIC v1.0 and v2.0, and the earlier HumanMethylation450 platform.
Sequencing
EM-seq, TAPS+, Nanopore adaptive sampling, Illumina 5-base and biomodal 6-base sequencing.
MNP-Flex. Platform independence is built on MNP-Flex, the platform-agnostic classifier described in Patel et al., Nature Medicine 2025. It maps profiles from different technologies into a shared representation before classification, which is why one model can serve them all.
Evidence
The research behind the classifiers
Nature · 2018
Capper et al. DNA methylation-based classification of central nervous system tumours
The paper that established the approach, demonstrating the first methylation-based classification of CNS tumours across a large reference cohort.
Nature Medicine · 2025
Patel et al. MNP-Flex, platform-agnostic classification
Introduces the flexible classifier that decouples methylation classification from the measurement platform, enabling array and sequencing input through one model.
Cancer Cell · 2026
Sill et al. the next generation of the CNS classifier
Describes the expanded CNS reference set and classifier version underlying the current MNP release, with performance across a substantially broader class set.
Nature Communications · 2021
Koelsche et al. sarcoma classification by DNA methylation
Extends methylation classification beyond the CNS, establishing the reference cohort and class structure behind the sarcoma classifier.
A fuller list, including the sarcoma validation preprint by Jäger et al. (2025), is on our publications page.
Getting access
Two ways to run our methylation classifiers
Research use: hosted
The full portfolio, including beta classifiers, on app.epignostix.com. Free registration, approval usually within a few days.
Routine use: licensed locally
Validated CNS, sarcoma and skin classifiers installed in your own infrastructure, running fully offline with installation and validation support.
Dedicated product pages
Classifiers on app.epignostix.com are for research use only and are not intended for diagnostic procedures.
Collaborate with us
A tumour type we don’t cover, or a platform we haven’t seen?
New classifiers come out of collaborations, with clinical partners who bring cohorts and questions, and with laboratories running technologies we have yet to support. Tell us what you are working on.
