CNN-CR core regulators: data to reproduce the figures of 'A compact motif vocabulary is differentially interpreted across human cell types'

Data accompanying A compact motif vocabulary is differentially interpreted across human cell types (Shen S, Wang W, 2026). Seven archives, each read by a named notebook of the code repository github.com/Wang-lab-UCSD/CNNCR_WangLab. Three archives larger than 1 GB are deposited as pieces named <archive>.tar.gz.part00, .part01, ...: download all pieces and join them with cat <archive>.tar.gz.part* > <archive>.tar.gz; CHECKSUMS.md5 lists the md5 of every whole archive. Unpack each archive with tar -xzf <archive> -C $ENHANCER_ROOT as described in the repository's PATHS.md; every archive contains a README.txt. | archive | contents | GB | notebooks | |---|---|---|---| | `fig1_training_inputs_all116.tar.gz` | the CNN-CR training input of every biosample: `{BSSID}.windows.fa.gz` (the 400-bp promoter and enhancer windows of the 18,356 genes, header `gene_index|ENSG|slot|type|chr:start-end|strand`) and `{BSSID}.expression.tsv` (log2 FPKM); `code/cnncr_model/cnncr_io.py` turns a pair into the one-hot tensor the models were trained on. The K562 pair ships in the repository | 2.6 | 01, 02, 05 (other biosamples) | | `fig1_motif_sets_all116.tar.gz` | the non-redundant promoter and enhancer motif vocabulary of every biosample (notebook 02 output) with the median contextual weights, and the TOMTOM tables of every set against JASPAR 2022, CIS-BP 2.0 and the two Epimotifs catalogues | 0.1 | 02 (Fig. 1e,f) | | `fig2_core_motif_derivation.tar.gz` | the input and every intermediate of the core-motif derivation: the ten models' promoter PWMs and contextual weights of every biosample, the expression vectors, the IC-filtered sets, the self- and JASPAR-TOMTOM tables, the representative motif sets, the pooled 8,441-motif set with its self-TOMTOM table, and the released 43 motifs | 0.6 | 04 | | `fig3_conservation_inputs.tar.gz` | phyloP100way (hg38) restricted to the promoter windows of the 116 biosamples (176 MB; identical values to the full UCSC file inside the windows), the K562 FIMO sites of the core and control-pool motifs with the K562 window coordinates, the per-biosample phyloP summaries of all 116 biosamples, UniProt Swiss-Prot human and the gnomAD v2.1.1 regional missense constraint table | 0.34 | 06 | | `fig4_eqtl_inputs.tar.gz` | the GTEx fine-mapped eQTL benchmark as released with Enformer (Avsec et al. 2021; causal PIP > 0.9 and matched non-causal variants, 49 tissues), the tissue-to-biosample map, the window coordinates of the 26 eQTL biosamples and their FIMO sites of the 43 core motifs and of the ten control sets | 1.4 | 07 | | `fig4_fig5_motif_sites_on_peaks.tar.gz` | FIMO sites of the 43 core motifs and the control pool in the TSS-proximal ENCODE TF ChIP-seq peaks of K562, HepG2, A549, GM12878 and MCF-7, and of the core and published control motifs in K562 and HEK293 peaks; the published Fig. 4d tables | 3.0 | 08, 11 | | `fig5_reader_switch_inputs.tar.gz` | the core-motif loci of the five ChIP-profiled lines and the loci shared by all five; the quantile-normalised knockout-effect vectors of the K562 (Replogle et al. 2022) and HEK293T (X-Atlas/Orion) Perturb-seq screens on their 8,236 shared genes; the reader-switch feasibility table; the published site-level reader tables and Fig. 5d-g tables | 0.92 | 10, 11 |