CLIPNET

Most human promoters carry no recognizable TATA box and no strong initiator element, yet RNA polymerase II still begins transcription at particular nucleotides within them. A model that returns one expression value per region cannot speak to this: the question is not how much transcription a sequence supports but exactly where, base by base, it starts. CLIPNET (Convolutionally Learned, Initiation-Predicting NETwork) predicts strand-specific PRO-cap coverage across a 500 bp window at single-nucleotide resolution, split into two heads — a profile head for the shape of initiation and a quantity head for its total magnitude. Because the heads are separate, attribution can ask which motifs position initiation and which set its abundance as two distinct questions.

Adam Y. He and Charles G. Danko at Cornell University released CLIPNET as a bioRxiv preprint in March 2024. It departs from the usual practice of training genomic models on a single haploid reference genome: it was fit on matched PRO-cap and individualized genome sequence from 58 genetically distinct lymphoblastoid cell lines, so variation between people supplies direct evidence about which nucleotides matter. Interpreting the trained ensemble with DeepSHAP recovered a regulatory syntax spanning five positions between −200 and +50 bp of the transcription start site, in which transcriptional activator motifs drive initiation quantity while core promoter motifs — degenerate DPR sequences and AT-rich TBP binding sites in TATA-less promoters among them — set initiation position.

CLIPNET sits beside Puffin, which also models initiation at nucleotide resolution but predicts a cell-type-agnostic profile, and inherits the profile-plus-quantity output decomposition of base-resolution models such as ChromBPNet. It is deliberately narrower than the long-context regulatory models Enformer and Borzoi: 1 kb of input, one assay, one cellular context.

#Key Features

• Base-resolution initiation profiles: Takes 1,000 bp of DNA and returns predicted PRO-cap coverage for the central 500 bp on both strands, resolving individual transcription start sites rather than aggregate promoter activity.
• Separated position and quantity: The two outputs are optimized by a multiscale loss and attributed independently, which is what lets activator and core promoter motifs be told apart by the role they play.
• Diploid sequence input: A two-hot encoding sums both alleles, so unphased heterozygous SNPs enter as IUPAC ambiguity codes instead of collapsing to a reference base.
• Ensembled across genome folds: Nine structurally identical replicates, each trained with a different holdout fold, are averaged; a further fold is withheld from all of them, so the ensemble can be benchmarked on sequence no replicate has seen.
• Interpretation shipped with the model: The clipnet package exposes DeepSHAP attribution, Deep Feature Interaction Map epistasis analysis, and in silico mutagenesis scans from the command line.

#Technical Details

Each replicate is a convolutional network: two convolutional layers (64 filters of width 8, then 128 of width 4) feed a tower of nine exponentially dilated convolutions (64 filters, width 3, dilation 1 to 512) joined by skip connections. A dense layer emits the length-1000 profile vector (500 bp plus strand concatenated with 500 bp minus strand); global average pooling followed by a dense layer emits the scalar quantity. Training minimizes negative cosine similarity on the profile plus a mean squared logarithmic error on the quantity, weighted at λ = 1/500, in TensorFlow 2.13.

Training used 67 PRO-cap libraries — 58 individuals plus 9 replicates — from GEO accession GSE110638, lifted to hg38 and matched to 1000 Genomes SNP genotypes. On the fully withheld fold (chromosomes 9, 13, 20 and 21; 4,901 peaks) the ensemble reached a median Pearson's r of 0.790 against observed PRO-cap tracks, against 0.570 for Puffin and 0.213 for a naive average-profile predictor, with individual replicates at 0.674–0.710. Predicted main TSS position correlated at r = 0.815 and log10 quantity at r = 0.669. Predicted allelic differences tracked measured ones at r = 0.48 across 2,057 initiation-quantity QTLs and r = 0.54 across 1,207 directionality QTLs.

#Applications

CLIPNET scores the effect of non-coding variants — eQTLs, GWAS hits, candidate regulatory SNPs — on where and how strongly a cis-regulatory element initiates transcription, and annotates the motif grammar of promoters and enhancers genome-wide through DeepSHAP and in silico mutagenesis. Its quantity head also supports prioritizing massively parallel reporter assay constructs, a task the authors note Puffin was not built for. Because the training data come from lymphoblastoid cell lines, predictions describe initiation in that cellular context.

#Impact

CLIPNET showed that transcription initiation is not adequately explained by an additive combination of independent motifs: the nonlinear model substantially outperforms a linear one, and the QTL analyses supply concrete loci where one motif's importance depends on another's. It also demonstrates the value of training on personalized rather than reference genomes, a design the same group extended in later work. The model remains a preprint that has not completed peer review. Code, weights and processed training data are released under the MIT license, with weights on Zenodo and Hugging Face and the package installable from PyPI.