Classic machine learning on top of multiple position weight matrices improves genomic prediction of transcription factor binding sites

ArChIPelago Data and models of: Kravchenko P., Vorontsov I.E., Grosse I., Makeev V.J., Kulakovskiy I.V., and Penzar D.D. (2026). Classic machine learning on top of multiple position weight matrices improves genomic prediction of transcription factor binding sites. ArChIPelago combines multiple position weight matrices (PWMs) of a transcription factor (TF) into one classifier of TF binding sites. This archive holds the ChIP-Seq peak sets and sequences of 36 human-mouse TF pairs (414 human and 290 mouse GTRD experiments), their PWMs (2,275 human and 1,816 mouse mono- and dinucleotide PWMs) and the trained models. The tables and figures of the manuscript are on GitHub (below). All versions of this archive: 10.5281/zenodo.14927303. Training and test sets test set training set human (hg38) chromosomes 1, 8, 21 chromosomes 2-7, 9, 10, 13-20 mouse (mm10) chromosomes 1, 8, 19 chromosomes 2-7, 9, 10, 13-18 Chromosomes 11 and 12, human chromosome 22 and the sex chromosomes are used in neither set. In every file name, train marks the training chromosomes and test or control the test chromosomes. Up to 10,000 positives per TF and set; negatives (1:100) are peaks of TFs of unrelated families (TFClass), GC-matched with BiasAway. The ArChIPelago models are trained on the human training set and evaluated on both test sets; the baseline is the best single monoPWM of each TF, selected on the human training set. Code, tables and figures The code, the manuscript analysis (results table, Supplementary Tables 1-6, figure panels and their source data, Kravchenko_et_al_Supplementary_Materials.pdf) and the scanning tool are on GitHub: github.com/autosome-ru/ArChIPelago, tag v1.0.0 github.com/autosome-ru/ArChIPelago-TFBS-finder, tag v1.0.0 git clone --recurse-submodules --branch v1.0.0 https://github.com/autosome-ru/ArChIPelago.git fetches both, with SPRY-SARUS and seqtk. Files file content train.tar.gz Training sets: peak sets (BED) and sequences (FASTA) of 39 human and 41 mouse TFs (the 36 TFs of the manuscript among them), human (chromosomes 2-7, 9, 10, 13-20) and mouse (chromosomes 2-7, 9, 10, 13-18); train/_.PEAKS000001.macs.train.mfa, train/_train_set_.bed. test.tar.gz Test sets: peak sets (BED) and sequences (FASTA) of the same TFs, human (chromosomes 1, 8, 21) and mouse (chromosomes 1, 8, 19); same naming with test. PWMs_mono_HUMAN.tar.gz Human monoPWMs of the 36 TFs (1,495 files, .pwm, log-odds), one folder per TF. PWMs_di_HUMAN.tar.gz Human diPWMs of the 36 TFs (780 files, .dpwm). PWMs_mono_MOUSE.tar.gz Mouse monoPWMs of the 36 TFs (1,192 files). PWMs_di_MOUSE.tar.gz Mouse diPWMs of the 36 TFs (624 files). Archipelago_intermediate_files_tables.tar.gz Global tables of the pipeline: out_tab_* and headerStr_names_* (one row per sequence of each FASTA file below), the pooled peak sets all_bed_df__set_*.bed, and the result tables of notebook 2 (HUMAN_MOUSE_SLIM*.txt). All five archives unpack into Archipelago_intermediate_files/. train = training chromosomes, control / test = test chromosomes; _SLIM = 1001 bp regions for the Slim models. Archipelago_intermediate_files_fasta_HUMAN.tar.gz All positive and negative human sequences (300 bp) of the training and test sets: all_mfa_file_HUMAN_10000_[_no_NF[_no_N]].fasta. Archipelago_intermediate_files_fasta_HUMAN_SLIM.tar.gz The same human sequences as 1001 bp regions for the Slim models: all_mfa_file_HUMAN_SLIM_10000_*.fasta. Archipelago_intermediate_files_fasta_MOUSE.tar.gz All positive and negative mouse sequences (300 bp): all_mfa_file_MOUSE_10000_*.fasta. Archipelago_intermediate_files_fasta_MOUSE_SLIM.tar.gz The same mouse sequences as 1001 bp regions: all_mfa_file_MOUSE_SLIM_10000_*.fasta. Models.tar.gz Pre-trained Random Forest models (scikit-learn 1.3, joblib .sav) of each TF on all its human PWMs (monoPWMs, diPWMs, both; 36 TFs x 3 models), each with a .json feature specification (PWM order, training mean and standard deviation of each feature); used by ArChIPelago-TFBS-finder. Slim_models.tar.gz Trained Slim models of the 36 TFs (_SlimModel_/Motif_/SlimDimont_.xml; m = 0, 1, 5 for LSlim m=-5, 7 for m=-7). ANDR m=1 is the model trained by Manuscript_analysis/analysis/slim_dichipmunk/train_slim_ANDR_m1.sh, used in Fig. 4 / S4. Slim.tar.gz Slim tools: TrainAndApplySlim.jar, SlimDimont.jar, Ruby utilities of diChIPMunk and the OpenJDK 8 they need. hocomoco11.tar.gz HOCOMOCO v11 data used to select datasets and PWMs: models, benchmark (auc) tables and curation tables. macs.tar.gz GTRD MACS peak intervals of the ChIP-Seq experiments (macs/PEAKS*.interval). The md5 sums are in md5sums.txt (md5sum -c md5sums.txt). Use Pipeline (notebooks 0-5, Python package, tests): ArChIPelago_code/ of the ArChIPelago repository; set the paths to the unpacked archives in ArChIPelago_code/config.yml. Tables and figures of the manuscript: Manuscript_analysis/ of the ArChIPelago repository (its README.md maps every table and figure to its script); bash Manuscript_analysis/scripts/rebuild_tables_and_figures.sh rebuilds them. Scanning new sequences with the pre-trained models: ArChIPelago-TFBS-finder (the models and human PWMs of this archive are part of that repository). Citation Kravchenko P., Vorontsov I.E., Grosse I., Makeev V.J., Kulakovskiy I.V., and Penzar D.D. (2026). Classic machine learning on top of multiple position weight matrices improves genomic prediction of transcription factor binding sites.