Seqweaver

A single substitution in a 3' UTR can delete the site an RNA-binding protein uses to find its transcript, changing how long that mRNA survives, where it is carried, or which exons reach the final message. A chromatin model reads the same base and has nothing to say: the regulation it describes is finished by the time the transcript exists. Seqweaver covers the post-transcriptional layer directly: it takes 1,000 bp of RNA sequence encoded as a 4 × 1,000 matrix and predicts the probability that the centre position lies inside an RBP target site, jointly across 232 CLIP-derived models spanning 88 distinct proteins. A variant is scored by running the frozen network twice, on reference and alternate alleles, and differencing the two output vectors. Because RNA has an orientation genomic coordinates do not, a submission must name the strand — the one input the lab's DNA-side models never ask for.

The network was built in Olga Troyanskaya's group at the Flatiron Institute and Princeton University, with Robert Darnell at The Rockefeller University and the Howard Hughes Medical Institute contributing the CLIP expertise behind the training labels. It first appeared in 2019 as the RNA arm of a two-model whole-genome analysis of autism risk, paired with a chromatin network in the DeepSEA line; it acquired the name Seqweaver, a public server and a genome-wide release with the 2021 Nature Genetics study of RBP dysregulation in psychiatric disease. No variant data enters training at any stage, so a rare or previously unobserved allele is scored on exactly the same footing as a common one.

#Key Features

• Calibrated allele-pair effect sizes: Raw probability differences are converted to per-feature e-values against the empirical distribution over gnomAD variants, to sign-preserving z-scores, and to a disease impact score from an L2-regularised logistic regression over the predicted effects.
• 232 tasks in one shared representation: Replicate CLIP experiments across 88 RBPs — splicing factors, stability regulators, cap-binding proteins — are predicted jointly, so a feature learned for one protein is available to the others.
• Applied frozen, never refit: The identical 232-model set was run over more than 21 million gnomAD SNPs "without any modifications", and the same checkpoints are reused unchanged as a feature extractor inside Otari.
• Three ways in: VCF scores variants, FASTA scores arbitrary 1,000 bp transcript sequence, BED scores reference positions — all strand-specified.

#Technical Details

Three convolutional layers of 160, 320 and 480 kernels, each of width 8, with max-pooling of width and stride 4 after the first two and dropout throughout, flatten to a 25,440-dimensional vector that a fully connected layer maps onto the task head, followed by a second dense layer and a sigmoid — the DeepSEA design carried over to RNA. Nearly all of its roughly seven million weights sit in that first dense layer. The released PyTorch definition instantiates the same network at two output widths, 217 for the human CLIP tasks and 43 for mouse. Training labels came from in vivo RBP–RNA interaction profiles: 231 CLIP binding profiles covering 82 unique RBPs, from ENCODE and previously published experiments, restricted to genic regions, with whole chromosomes held out for evaluation. The 2021 release applied that fixed model set to 21,513,861 noncoding SNPs from gnomAD 2.1, across 5' UTRs, introns and 3' UTRs.

#Applications

Users reach the model through the HumanBase server, which takes VCF, FASTA or BED input and returns per-RBP probabilities, differences, z-scores and e-values, or through a standalone bundle for jobs past the server's roughly 10,000-record limit. Precomputed resources skip the compute entirely: the full gnomAD 2.1 dysregulation scores ship as a 37.6 GB download and the 1000 Genomes RBP LD scores as 33.5 GB, which brings partitioned-heritability analysis over an existing GWAS within reach without a GPU. The scores also serve as features for other models — Otari concatenates Seqweaver RBP affinities into its transcript-graph node attributes.

#Impact

The genome-wide application established RBP target site dysregulation as a substantial and largely unclaimed component of psychiatric heritability. Across ADHD, autism spectrum disorder, bipolar disorder, major depression and schizophrenia, 304 RBP–disorder pairs reached FDR < 0.05, with the collective RBP contribution to schizophrenia exceeding GTEx splicing QTLs more than fourfold and surviving conditioning on conservation elements and on molecular QTL annotations. One result carried through to mechanism: rs6981405 at the 8p12 schizophrenia locus was predicted to disrupt QKI binding in the DDHD2 3' UTR, which ENCODE eCLIP and QKI knockdown RNA-seq confirmed. Three limits bound the artifact: training was restricted to genic regions, so intergenic variants are out of scope; CLIP coverage decides which proteins and contexts are represented at all; and distribution is S3 tarballs plus a hosted server rather than a source repository, with the released code making predictions rather than reproducing training, under no stated license.