cfDecon

Cells shed DNA into the bloodstream when they die, and that cell-free DNA carries the methylation marks of the tissue it came from. Working out which cell types contributed to a plasma sample, and in what proportion, is the basis of a whole class of non-invasive tests: detecting a tumour by the appearance of hepatocyte-derived DNA, monitoring transplant rejection, or tracking neurodegeneration through neuron-derived fragments. The computational problem — deconvolving a mixture against a reference atlas of cell-type methylation signatures — is hard because the mixture is dominated by blood, the contributions of interest are often below one percent, and the reference atlas may not contain every cell type that is actually present.

cfDecon, from the Chinese University of Hong Kong with collaborators at Soochow University, is a deep learning framework that performs this deconvolution at the resolution of individual sequencing reads rather than at individual CpG sites. Read-level methylation is more informative than a site-averaged ratio: a single read spanning several CpGs records whether they were methylated together, which distinguishes cell types that site-level summaries blur. cfDecon builds on that read-resolution formulation with a learned model in place of the expectation-maximisation used by earlier read-level methods.

The architecture pairs a multichannel autoencoder with an iterative refinement loop, and outputs two things: the cell-type proportions, and condition-aware cell-type-specific methylation profiles that make the estimate interpretable. The work appeared as a bioRxiv preprint in February 2025 and was published in the RECOMB 2025 proceedings.

#Key Features

• Read-level resolution: Deconvolution operates on whole reads, preserving the joint methylation pattern across neighbouring CpGs that a site-wise average discards.
• Multichannel encoding: The encoder processes the read-level input through five parallel channels, and a learned gating network weights and compresses them into a single latent representation before decoding.
• Iterative refinement: Proportion estimates are refined across passes rather than produced in one shot, which is what keeps rare cell types and cell types absent from the reference atlas from being absorbed into the dominant components.
• Interpretable outputs: Alongside the proportions, cfDecon generates cell-type-specific methylation profiles conditioned on the sample's disease state, so differentially methylated CpGs can be pulled out and passed to enrichment analysis.
• Train once, apply repeatedly: The model is trained on a reference atlas plus simulated mixtures and then applied to new clinical samples without per-sample refitting.

#Technical Details

The core module is an autoencoder whose encoder progressively compresses the input through a stack of reduction layers while a mirrored decoder reconstructs it. The encoder operates over five parallel channels; a gating network models the interaction between them, and a compressor aggregates the five channel outputs into a unified latent vector using softmax weighting before decoding. Only the compressed representation passes to the decoder, forcing the model to keep what is shared across channels.

Evaluation covers simulated mixtures under three regimes — normal cellular composition, rare cell types, and cell types missing from the reference — where cfDecon outperforms prior deconvolution methods, and an atlas-level reference dataset on which it raises Lin's concordance correlation coefficient by over 33% under strict data separation controlling for leakage. On two real clinical cohorts held out from training, it raises disease detection performance from 0.53 to 0.79 for amyotrophic lateral sclerosis and from 0.55 to 0.77 for hepatocellular carcinoma relative to existing methods. The codebase on GitHub provides explicit train and test modes plus data-processing and evaluation scripts; it ships no licence file, and the preprint is CC-BY-NC-ND, so the terms of reuse are restrictive.

#Applications

The direct application is non-invasive diagnostics from a plasma methylation sequencing run: estimating which tissues are dying and at what rate. The ALS and HCC results show the two ends of that spectrum — a neurodegenerative condition where the signal of interest is a small neuronal fraction, and a cancer where tumour-derived DNA is the readout. Because cfDecon also returns condition-specific methylation profiles, it supports the biology as well as the classification, letting researchers ask which CpGs in which cell type shift with disease and feed those into functional enrichment.

#Impact

cfDecon's contribution is showing that a learned, read-resolution model with an explicit refinement loop handles the two situations that break atlas-based deconvolution in practice: cell types that are rare, and cell types the atlas never had. The large jumps in disease-detection performance on the ALS and HCC cohorts are the strongest evidence, though they are retrospective analyses of existing datasets rather than a prospective clinical evaluation. No pretrained checkpoint is distributed — the repository ships training and inference code but not weights — and the absence of a licence file leaves the terms of reuse for the code unresolved.