ExPecto

An eQTL study can only speak about alleles it has already seen. A promoter variant carried by a single family has no expression association to look up, and the noncoding mutation space is far too large for population data ever to cover. ExPecto sidesteps the problem by never learning from variants at all. It predicts how highly a gene is expressed in a given tissue from the DNA sequence around its transcription start site, then scores a variant by running the same fixed model twice — once on the reference sequence, once on the alternate — and differencing the two predictions. An allele nobody has observed is scored exactly the way a common one is.

The prediction runs in three stages. A deep convolutional network first turns sequence into chromatin: it scans the 40 kb window centred on a gene's representative TSS in 200 bp steps and, at each step, predicts 2,002 histone-mark, transcription-factor-binding and DNase-accessibility profiles. That network — released alongside ExPecto as Beluga — extends the earlier DeepSEA chromatin model with double the convolution depth, a wider input window and twice the feature space. A spatial transformation collapses the 200 positional bins into distance-weighted summaries, and a per-tissue regularized linear model converts those into a predicted expression level.

Jian Zhou, Chandra Theesfeld, Kevin Yao, Kathleen Chen, Aaron Wong and Olga Troyanskaya built ExPecto at the Lewis-Sigler Institute for Integrative Genomics at Princeton University and at the Flatiron Institute, publishing it in Nature Genetics in 2018.

#Key Features

• Ab initio variant scoring: No matched genotype-expression data is used in training, so rare, private and entirely hypothetical alleles get an effect on the same footing as common ones.
• Forty kilobases of regulatory context: Predictions draw on 20 kb upstream and 20 kb downstream of the representative TSS, reaching regulatory signal that a promoter-window model never sees.
• Distance-weighted spatial transformation: Ten prespecified exponential basis functions summarize each chromatin feature's spatial profile, cutting the input dimension twenty-fold while preserving where along the sequence a signal falls.
• Tissue-resolved output: One model per tissue, 218 in all, each predicting log RPKM, so an effect comes back as a direction and magnitude per tissue rather than one pathogenicity score.
• Genome-scale saturation mutagenesis: Cheap enough per variant to have been run in silico over more than 140 million promoter-proximal mutations across human RNA polymerase II-transcribed genes.
• Precomputed effect browser: A public server holds predictions for every substitution within 1 kb of a representative TSS, plus 1000 Genomes variants above a 0.3 log fold-change threshold in any tissue.

#Technical Details

The chromatin component predicts 2,002 profiles from ENCODE and Roadmap Epigenomics covering more than 200 cell types, calling the central 200 bp of a 2 kb input window. Scanning ±20 kb around the TSS in 200 bp steps yields 200 spatial bins per feature. The spatial transformation weights those bins by exponential decay in distance to the TSS — decay constants of 0.01, 0.02, 0.05, 0.1 and 0.2, applied separately upstream and downstream — reducing the representation to 20,020 features. Coefficients are shared across bins, so the fitted model has twenty times fewer parameters than an unpooled one. The final layer is 218 L2-regularized linear models fitted by gradient boosting on expression profiles from GTEx, Roadmap Epigenomics and ENCODE, with representative transcription start sites selected from FANTOM CAGE data. The released implementation runs on hg19 coordinates, ships the per-tissue models in the repository, and includes a training script for fitting additional tissue models on a user's own expression profile.

#Applications

The most direct use is narrowing a GWAS locus: given a block of variants in linkage disequilibrium, each gets a tissue-specific predicted expression change, prioritizing those with mechanistic support over ones merely tagged by association. The original work did this across all publicly available GWAS at the time and validated predictions experimentally for four immune-related diseases. The same scoring applies to rare noncoding variants from clinical or family sequencing, where no population-level association will ever exist, and the saturation mutagenesis output supports questions about how strongly selection constrains expression at each position around a promoter.

#Impact

ExPecto established a composition that regulatory genomics has reused since: freeze a sequence-to-chromatin network, treat its outputs as a learned representation, and fit a light downstream model per tissue or task on top. The same lab's Otari later applied that pattern at isoform resolution, stacking chromatin, splicing and RNA-binding-protein predictors over transcript graphs. Practical constraints are worth knowing: it is tied to hg19, the provided tissue models are in a legacy gradient-boosting format that the training script does not reproduce, the hosted browser serves a precomputed variant set rather than arbitrary input, and the code and weights carry a Princeton academic-use agreement limiting them to non-commercial research.