Can one DNA methylation read tell us which of 39 human cell types it came from? We fine-tuned MethylBERT on the Loyfer et al. (2023) atlas to find out. For single reads with hard labels, the answer is mostly no.
- Our first 39 class model scored 99%, but the marker region alone gave away the label; after we removed that leak, accuracy within families of related cell types stayed near chance across 45 runs.
- Only about 2% of reads fall in a marker region of their own cell type, and 35% of test reads in one family have an identical training read with a different label, so a single hard label per read has little to fit.
- We split the atlas into lineage families to make the task smaller; it helped in some families (Excitable 0.55, chance 0.33) and barely in others (Structural 0.25, chance 0.17).
- The group that built MethylBERT then completed the goal: read classification across all 39 types, using data driven soft labels in place of hard ones (Rizdvanetskyi et al., 2026).
Full write-up: REPORT.md. All 45 runs: Results/.
| Task | Classes | Chance | Accuracy | Macro F1 | Note |
|---|---|---|---|---|---|
| Flat 39 class, first split | 39 | 0.026 | 0.991 | 0.969 | Leak: region ID alone predicts the label |
| Flat 39 class, region balanced | 39 | 0.026 | 0.520 | 0.511 | Matches what the region alone would score (about 0.50) |
| Pancreas acinar vs duct | 2 | 0.500 | 0.833 | 0.825 | Upper bound; split not saved |
| Excitable (cardiomyocyte, neuron, oligodendrocyte) | 3 | 0.333 | 0.548 | 0.542 | Upper bound; read level split |
| Digestive epithelia | 4 | 0.250 | 0.395 | 0.386 | |
| Reproductive epithelia | 4 | 0.250 | 0.378 to 0.394 | 0.372 to 0.382 | 2 runs |
| Blood | 6 | 0.167 | 0.198 to 0.430 | 0.189 to 0.348 | 19 runs; 0.430 is on an unbalanced test set |
| Surface epithelia | 6 | 0.167 | 0.276 to 0.277 | 0.267 to 0.273 | 3 runs |
| Metabolic (liver, pancreas) | 6 | 0.167 | 0.244 to 0.273 | 0.239 to 0.270 | 3 runs |
| Structural (fibroblast, muscle, fat, endothelium) | 6 | 0.167 | 0.217 to 0.246 | 0.175 to 0.228 | 6 runs |
- Accuracy is per read, from the checkpoint with the lowest eval loss.
- Pure samples, which should resolve to one class, came out near uniform or collapsed onto one class (details).
- First split (the 0.991 row): 80/20 random split of reads within each cell type (
Training/Toolbox/split_data_simple.py). - Later splits: 75/25 with
Training/Toolbox/split_data_flexible.py.- Whole samples go to one side only for cell types with 4 or more samples.
- Cell types with 3 or fewer samples are pooled and split by unique read.
- Which mode built each dataset was not recorded.
- The Excitable split behaves like a read level split. 49% of its test reads have an identical read in training (audit).
- So reads from the same donor can sit on both sides. That inflates accuracy. The near chance results stand, but 0.83 and 0.55 are upper bounds.
git clone https://github.com/EG-xry/CellClassification_MethylBERT.git
cd CellClassification_MethylBERT
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python demo/run_demo.py- The demo fine-tunes for 60 steps on 780 reads from the Excitable family (
demo/data/), then evaluates. - It downloads the pretrained
hanyangii/methylbert_hg19_2lcheckpoint on first use. - It takes about 40 seconds on an Apple M4 Max and writes metrics and plots to
demo/output/. - It shows that the pipeline runs; 60 steps is not a result.
Preprocessing/ .pat.gz reads + marker regions -> training rows
analyze_methylation_batch_optimized.py main script
Config/ sample-to-cell-type maps (206 atlas samples)
Data/ marker region lists (hg38_all_1000.csv, hg38_250.csv)
Toolbox/ format converters and helpers
Training/
run_training.py fine-tuning entry point (python Training/run_training.py -c <config>)
src/methylbert/ modified MethylBERT (see MODIFICATIONS.md and LICENSE there)
Config/ configs for each loss and family
Toolbox/ combine, split, convert, and analysis tools
Testing/ pure-sample deconvolution and CTC scripts
Results/ all 45 runs, split audits, run index
Data/ read/region agreement and cell type proportion tables
demo/ one-command demo
REPORT.md full write-up
Steps 2 to 4 ask for file names interactively (step 2 opens a small window if tkinter is available).
- Preprocess each atlas sample into rows of
DNA_seq, Methyl_seq, DMR_Label, cType, DMR_cType, Read_Count. Run from the folder holding the inputs (see Data sources):cd Preprocessing/Data python ../analyze_methylation_batch_optimized.py \ --config ../Config/cell_types_config_collective.json --dedupe-reads --parallel 4 - Combine the per-sample files of one family:
python Training/Toolbox/Combined_CSV.py. - Split into train and test:
python Training/Toolbox/split_data_flexible.py.- Mode 4 assigns whole samples by hand; use it to hold out donors.
- Convert to the 4-column training format:
python Training/Toolbox/convert_data_format_flexible.py. - Train:
python Training/run_training.py -c Training/Config/<config>.json. - Deconvolve a pure sample:
python Testing/cell_type_deconvolution.py --csv <sample>.csv --model_dir <run>/bert.model.
The CTC scripts in Testing/ (process_ctc_pat_files.py, predict_cell_type.py) were written for circulating tumour cell data but never run to results. predict_cell_type.py does not pass methylation to the model; do not use it as is.
- Methylation atlas: Loyfer et al. (2023), GEO GSE186458.
- 206 hg38
.pat.gzfiles, GSM5652176 to GSM5652382 (GSM5652260 not used). - Not included in this repository.
- 206 hg38
- Marker regions:
Preprocessing/Data/hg38_all_1000.csv(about 1,000 per cell type) andhg38_250.csv(up to 250 per type).- Both derive from the Loyfer atlas markers.
- The 250 list was lifted to hg38 with UCSC liftOver (
Preprocessing/Toolbox/hglft_genome_12e3e_1b8300.bed). - The exact source table was not recorded.
- Reference files (not included):
- hg38 genome FASTA (UCSC)
CpG.bed.gzCpG index in wgbstools format
- Pretrained model:
hanyangii/methylbert_hg19_{2,4,6}lon Hugging Face (Jeong et al., 2025).
- Python 3.11.
- Pinned versions are in
requirements.txt, checked by installing into a clean environment and running the demo. - Training used CUDA GPUs; the demo also runs on Apple MPS or CPU.
- This repository's code is MIT licensed (LICENSE).
Training/src/methylbert/is a modified copy of MethylBERT by Yunhee Jeong, MIT licensed.- The original license is kept in that folder.
- The changes are listed in MODIFICATIONS.md.
Figure_Loyfer.webpis Figure 2 of Loyfer et al., Nature 613, 355–364 (2023), reproduced unchanged under CC BY 4.0.
- Jeong, Y. et al. (2025). MethylBERT enables read-level DNA methylation pattern identification and tumour deconvolution using a Transformer-based model. Nature Communications 16, 788.
- Loyfer, N. et al. (2023). A DNA methylation atlas of normal human cell types. Nature 613, 355–364.
- Rizdvanetskyi, D., Roos, N. and Lutsik, P. (2026). Data-driven soft labeling scales DNA read classification to whole-body cell-type deconvolution. arXiv:2607.04987.