SYTO PUBLICATION DATA - FILE NAMING CONVENTION ============================================== Last updated: 2026-08-24 This file documents the naming convention used in this dataset, together with every abbreviation and code that appears in a file or folder name. 1. GENERAL RULES ---------------- - Words within a name element are joined without separators (e.g. "softlabels"); underscores (_) separate the elements themselves. - Directory names are lowercase with hyphens between words (tier0-results, rrbs-recovered-reads). Result file names use the element scheme in section 3. - Names are location-independent: a file still identifies itself after being moved out of its folder. - Version is a trailing _v. v1 is the first published release. 2. TOP-LEVEL LAYOUT ------------------- The deposit ships as archives, one per logical unit, plus a few loose files that should be readable without downloading anything large: tier0-results/ Arhives with final result for two main experiments + TabulaSapines mappings + consolidated ranking table of Syto variants and external baselines. tier1-models// Trained Syto variants, one archive per variant. tier2-pseudobulks// One archive per pseudobulk run. tier3-training-data/ Marker atlases used in Syto paper; target proportions used for each pseudobulk generation; training datasets under datasets//. tier4-source-data/ Staged and recovered reads, plus the hg19 and hg38 reference genomes. MANIFEST.csv One row per archive or standalone file: what it is, size, file count, sha256. runs.csv One row per experiment run. FILE_NAMING_CONVENTION.txt This file. Unpacking every archive restores the original tree: tier0-results/ Final tables underlying the publication. tier1-models/ Trained classifiers, deconvolvers and calibrators. tier2-pseudobulks/ Generated pseudobulk mixtures (columnar Parquet). tier3-training-data/ Atlases, target proportions and training datasets. tier4-source-data/ Staged and recovered reads, and reference genomes. Within tier1 and tier2, folders are /, where run_name is __, e.g. dismir__softlabelpooled. 3. RESULT FILE NAME SCHEME (tier0-results) ------------------------------------------ Six elements: _____.csv Example: SYTO_oodrrbs_dismir_softlabels_top156_v1.csv | | | | | | | | | | | +-- version | | | | +--------- feature set | | | +-------------------- label mode | | +--------------------------- classifier | +----------------------------------- experiment +---------------------------------------- project 4. CODES AND ABBREVIATIONS -------------------------- PROJECT SYTO The framework introduced in "data-driven soft labeling scales dna read classification to whole-body cell-type deconvolution". EXPERIMENT MAIN EXPERIMENTS: pseudobulk [pseudobulk] In-distribution evaluation on generated pseudobulks. oodrrbs [ood-rrbs] Out-of-distribution evaluation on the cfSort RRBS cohort. SUPPLEMENTARY EXPERIMENTS: oodrrbsgss [ood-rrbs-gss] Out-of-distribution RRBS using the GSS-sorted atlas. poolingsensitivity [pooling-sensitivity] Sensitivity analysis over read-pooling thresholds. mledeconvolvers [mle-deconvolvers] Maximum-likelihood deconvolver variants. CLASSIFIER dismir Syto-extended DISMIR read-level classifier (CNN-LSTM). methylbert Syto-refactored MethylBERT transformer classifier. lookup Syto-introduced 1NN-Lookup-table classifier. cancerdetector Syto-extended CancerDetector baseline classifier. baselines Not a classifier: reference deconvolution methods (Celfie, EpiDISH, Houseman CP, UXM) collected in one table. LABELMODE (how read-level training labels were assigned) hardlabels Hard labels with a background class. softlabels Data Driven Soft labels, with read pooling. softnopooling Data Driven Soft labels, without read pooling. softcanonical Soft labels produced by label smoothing (a.k.a canonical). uniformprior CancerDetector with a uniform cell-type prior. trainfreqprior CancerDetector with a training-frequency prior. FEATURESET (which probability-simplex scores were selected for deconvolution) top156 The top 156 features (see procedure in the paper). diagbackground Diagonal-background feature selection (see procedure in the paper). gssatlas These runs also use the top-156 feature set but the entire pipeline is run on Proxy-GSS-sorted atlas. VERSION v1 First published release. REFERENCE GENOME (appears in tier3/tier4 folder names) hg19, hg38 Human reference genome builds. Taken from wgbs_tools; ATLAS NAMES (tier3-training-data/atlases) U25, U250 Loyfer atlas variants (25 or 250 regions per cell type). l4 The minimum length of called CpGs in read/fragment considered when computing atlas values. GSS_SORTED Proxy Gap Specificity Score (GSS) sorted atlas. Each cell type is represented by the top 25 regions starting from the highest GSS trainonly Atlas built from the training split only. Otherwise, atlas built from the entire dataset 5. ARCHIVE NAMES ---------------- Archives follow the same element style as result files: project first, version last, underscores between elements, no spaces. __<...unit...>_.zip The middle elements identify the unit and vary by tier: tier0 one archive per result group SYTO_tier0_results_baselines_v1.zip SYTO_tier0_results_cfsortooddeconvolutionresults_v1.zip tier1 one archive per trained model (experiment + classifier + label mode) SYTO_tier1_models_oodrrbs_dismir_softlabelpooled_v1.zip tier2 one archive per pseudobulk run, same elements as tier1 (pseudobulk is build on the classifiers outputs and the mapping is one-to-one) SYTO_tier2_pseudobulks_pseudobulk_methylbert_softlabelpooled_v1.zip tier3 one archive for atlases, one for target proportions, one per dataset SYTO_tier3_trainingdata_atlases_v1.zip SYTO_tier3_trainingdata_hg38_oodalldatalabels_v1.zip tier4 one archive per source dataset, and one per reference genome SYTO_tier4_sourcedata_stagedu250_hg38_v1.zip SYTO_tier4_sourcedata_referencegenomes_hg38_v1.zip USING THE ARCHIVES ------------------ Members are stored with their path relative to the DEPOSIT ROOT. Always unpack from the root, whichever folder an archive came from: cd /path/to/syto-publishable-data find . -name 'SYTO_*.zip' -exec unzip -o -q {} ';' # everything, or unzip tier1-models/ood-rrbs/SYTO_tier1models_oodrrbs_dismir_softlabelpooled_v1.zip Unpacking an archive from inside its own folder would nest the tree a second time (tier1-models/ood-rrbs/tier1-models/...) and the configs would not resolve. The reference-genome archives contain relative symlinks (genome.fa.gz -> hg38.fa.gz) matching the wgbs_tools layout. Unpack them with a tool that preserves symlinks; unzip on Linux and macOS does. 6. NOTES ON CONTENT NOT ENCODED IN NAMES ---------------------------------------- - Missing-label strategy: every cfSort deconvolution result in tier0 was produced with the prior blending strategy (see the paper). - Deconvolver and calibrator are NOT in the file name. They vary within a file and appear as the "Deconvolver" and "Calibrator" columns instead. - All experiments were originally recorded using our local mlflow registry. For the purposes of sharing we have stripped the data from the mlflow related files, however, each element in tier1-models contains recorded metrics, parameters and tags from classifier fitting. Hence, it should be possible to create your own syntetic meta.yaml files to access these metadata via mlflow. 7. CONFIG PATHS --------------- The published YAML configs use paths relative to that same root, so they resolve once the archives holding the files they reference are unpacked. A config will not find its inputs while they are still inside a zip. The configs were automatically redacted for this data packaging purposes, meaning that original configs were referring to different paths in either logical filesystem or HPC. We didn't re-run all experiments after re-packaging configs. Therefore, shared configs must be treated as the reference points, meaning that the parameters are as in original experiments. Still, users may need to correct pointers to the paths depending on how they unpack the data. Two placeholders mark inputs this release does not redistribute: ${LOYFER_RECOVERED_READS} Loyfer recovered read tables: outputs of https://github.com/CompEpigen/wgbs_atlas_simulation ${SYTO_MLFLOW} An MLflow store, for the intermediate *_predicted.pkl inputs that are not published. Can also be any other intermediate store. The *_predicted.pkl were the original data splits enriched with the outputs of the trained classifiers that were used as inputs to pseudobulks generation. The pseudobulk pipeline is available to run in predictions_only mode, which allows producing such files. This is handy when, for example, we need to use a GPU-heavy classifier (like MethylBERT) for predictions, and we want to run this stage on a more expensive GPU node, deferring pseudobulk generation to a CPU-only node. While we did not publish those files, we have included two-stage configs where relevant so users can reproduce pseudobulks.