ProCapNet is a neural network trained to predict PRO-cap signal from DNA sequence alone. PRO-cap is a run-on assay that captures the 5' end of each nascent RNA, so it reports the exact base and strand at which RNA polymerase II started transcribing, including at enhancers and at unstable transcripts that RNA-seq and CAGE miss. Six separate models were trained, one on each of six cell lines with ENCODE PRO-cap data.
This track holds two kinds of output from those models:
Predictions are not measurements: they say what the sequence looks capable of, not what a given cell is doing. The matching experimental data is in the PRO-cap track.
Cell lines are listed in the table on this page, one row each, with a checkbox per data type. Use the Sample class facet to narrow the list, and the Group tracks by buttons to order the browser by sample or by data type.
Each predicted PRO-cap track is an overlay of the two strands: plus strand predictions are drawn upward and minus strand predictions downward. The y axis is the predicted number of PRO-cap reads at that base.
Contribution scores are drawn as a sequence logo when zoomed in far enough to show individual bases: the letter of the reference base is scaled by its score, so a run of tall letters is a motif the model relied on. At lower zoom the same values are drawn as a wiggle. Scores can be negative, meaning the base argued against initiation being placed where it was. Scores exist only in the roughly 2 kb window around each MANE Select transcription start site, about 38.7 Mb of the genome. Everywhere else the track is empty, which is not the same as a score of zero.
Read depth differs between the six PRO-cap experiments the models were trained on, and both predicted signal and contribution scores scale with it, so the y axis is not comparable between cell lines. Tracks are colored by the cell line the model was trained on:
ProCapNet adapts the BPNet architecture. It reads 2,114 bp of one-hot encoded sequence and outputs both a base-resolution profile over the central 1,000 bp, covering both strands as a single softmax so the model can learn strand asymmetry, and a scalar giving the log total number of initiation events in that window. Training used ENCODE PRO-cap read alignments in which the first read of each pair was discarded and only the single 5'-most base of the second read was kept, merged across replicates and kept separate by strand. Each model was trained on all PRO-cap peaks in its cell line plus randomly sampled DNase-hypersensitive sites from the same cell line at a 7:1 peak to background ratio, with 7-fold cross-validation split by chromosome. Bases that are not uniquely mappable by 36-mer reads were given zero loss weight during training. Full details are in Cochran et al.
Genome-wide predictions were generated by the Kundaje lab by applying the model to every 2,114 bp window at a stride of 250 bp, so each base is the average of four overlapping predictions, then averaging across the seven cross-validation models and across the forward and reverse-complemented sequence. On hg38 no prediction was made where most of a window was unresolved (N) in the reference.
Scores were computed with DeepSHAP, which estimates each base's contribution by contrasting the model's output on the real sequence against its output on a set of reference sequences, here 25 dinucleotide shuffles of the sequence being scored. Because DeepSHAP needs a single scalar to explain, the base-resolution profile output was summarized by mean-normalizing the pre-softmax logits and taking their dot product with the post-softmax profile, which weights each base's logit by its predicted probability of being used and sums over the 1,000 bp output window and both strands. This is the profile or TSS-positioning task; ProCapNet can also produce scores for its read-count task, which are not shown here. Each scored sequence was run through all seven cross-validation models and in both orientations, and the scores averaged.
The bigWig files were downloaded from the Kundaje lab server. The contribution score files are used unaltered. In the prediction files the per-base values were re-encoded from one interval per base into fixedStep sections, which cuts the file size by about a third and changes no value. Bases carrying a literal NaN in the published hg38 files, all of them in unresolved (N) reference sequence, were dropped so that the track autoscales and summarizes correctly; this removed 164,268,582 of the 3,088,269,832 bases on the primary hg38 chromosomes, leaving 2,924,001,250 bases with a prediction. The hs1 files had no such bases and all 3,117,275,501 were kept. Nothing else was altered: every retained value is bit for bit the published value. The steps are recorded in the makeDoc and the scripts are in the kent source tree.
The data can be explored interactively in table format with the Table Browser or the Data Integrator and exported from there to spreadsheet or tab-sep tables. From scripts, the data can be accessed through our API, track=proCapNet.
The Files column of the table on this page links each cell line's bigWigs directly, so a single file can be fetched without working out its path.
For automated download and analysis, the genome annotation is stored in bigWig files that can be downloaded from our download server. Predictions are under pred/ and are named for the cell line, the model and the strand, for example K562.proCapNet.pos.bw and K562.proCapNet.neg.bw. Contribution scores are under contrib/, for example K562.proCapNet-contrib.bw. The T2T-CHM13 predictions are in the same layout under /gbdb/hs1/proCapNet/pred/; there are no contribution scores for that assembly. Individual regions or the whole genome annotation can be obtained using our tool bigWigToBedGraph, which can be compiled from the source code or downloaded as a precompiled binary for your system. Instructions for downloading source code and binaries can be found here. The tool can also be used to obtain features within a given range, e.g. bigWigToBedGraph http://hgdownload.soe.ucsc.edu/gbdb/hg38/proCapNet/pred/K562.proCapNet.pos.bw -chrom=chr21 -start=0 -end=100000000 stdout
The original files can be downloaded from the Kundaje lab server. The trained models are on the ENCODE portal, linked from the Experiment column of the table on this page.
ProCapNet was developed by Kelly Cochran in the Kundaje lab at Stanford University. The genome-wide predictions and contribution scores were generated by Kelly Cochran in collaboration with the GENCODE consortium. Thanks to Kelly Cochran and Anshul Kundaje for making the data available.
Cochran K, Yin M, Mantripragada A, Schreiber J, Marinov GK, Shah SR, Yu H, Lis JT, Kundaje A. Dissecting the cis-regulatory syntax of transcription initiation with deep learning. bioRxiv. 2024 Nov 21;. -PMID: 38853896; PMC: 10.1101/2024.05.28.596138; PMID: 38853896; PMC: PMC11160661
Avsec Ž, Weilert M, Shrikumar A, Krueger S, Alexandari A, Dalal K, Fropf R, McAnany C, Gagneur J, Kundaje A et al. Base-resolution models of transcription-factor binding reveal soft motif syntax. Nat Genet. 2021 Mar;53(3):354-366. -PMID: 33603233; PMC: 10.1038/s41588-021-00782-6; PMID: 33603233; PMC: PMC8812996
Kwak H, Fuda NJ, Core LJ, Lis JT. Precise maps of RNA polymerase reveal how promoters direct initiation and pausing. Science. 2013 Feb 22;339(6122):950-3. +DOI: 10.1126/science.1229386; PMID: 23430654; PMC: PMC3974810