Regulatory variant effect prediction from DNA sequence, trained only on nervous-tissue epigenomic assays to score non-coding SNPs in brain disorders.
No providers recorded yet. Browse all providers
Most variants that genome-wide association studies turn up for neurological and psychiatric conditions sit outside genes, where their consequence is regulatory rather than structural. Association alone cannot separate a causal regulatory variant from a neighbour in tight linkage disequilibrium, so a locus can light up without anyone knowing which variant does anything — and reading a regulatory consequence off a non-coding change means knowing what the surrounding sequence does in the relevant tissue.
Neur-Ally, from Anil Prakash and Moinak Banerjee at the Rajiv Gandhi Centre for Biotechnology in Kerala, takes the sequence-to-epigenome approach and narrows it deliberately. Where general models such as Enformer span dozens of tissues, Neur-Ally trains exclusively on nervous-system data — brain tissues, neural cell lines, in vitro differentiated cells, primary cells, and organoids — on the argument that a model spread across the whole body cannot learn grammar specific to one. From DNA sequence alone it predicts whether a genomic window carries chromatin accessibility, particular histone modifications, or binding by particular transcription factors.
Scoring then uses the trained model unchanged: in silico mutagenesis substitutes the alternative allele, the model is run again, and the shift in predicted signal becomes the variant's score. Published in NAR Genomics and Bioinformatics in 2025 after a January preprint.
Input is a 2,000 bp window — a 200 bp genomic bin plus 1,800 bp of flanking context — encoded as integers over A, C, G, T, N and embedded at 128 dimensions with sinusoidal positional encoding. Two rounds of 1D convolution (kernel size 10, 64 then 32 filters) and max pooling at stride 3 extract motif-scale features, followed by multi-head attention layers for long-range dependencies and a dense layer with sigmoid activation producing independent multi-label probabilities. Training data is 9 million bins over hg38, labelled positive for an assay when a narrowPeak call covers more than half the bin; chromosomes 7 and 8 were held out for validation and testing. The TensorFlow/Keras implementation trained for 39 epochs under binary cross-entropy.
Performance is reported per assay family: chromatin accessibility labels reach mean AUROC 0.93 with PR-AUC 0.23 against a 0.01 baseline, histone modifications 0.84 with PR-AUC 0.29 against 0.03, and transcription factor binding 0.87 with PR-AUC 0.22 against 0.01 — the gap between the two measures reflecting how rare positives are per label. At an E-value threshold of 1e-05 the model flagged 48 of 7,663 non-coding neurological GWAS variants, 169 brain eQTL variants, and 4 of 92 autism-associated SNPs. Of eight independently reported probable brain regulatory SNPs, it called rs7364180 and rs12411216 significant, matching their published mechanisms: CTCF-mediated chromatin looping and altered accessibility in prefrontal cortex.
The model is aimed at post-GWAS triage: a study returns a locus, and a researcher needs to know which variant to take to a reporter assay or CRISPR edit. Neur-Ally ranks candidates and names the assay and sample where an effect is predicted, turning an abstract score into a testable hypothesis about a specific mark in a specific brain region. It suits prioritising brain eQTL variants and sifting candidates from Alzheimer's, Parkinson's, autism, schizophrenia, and epilepsy studies. Running it requires substantial external reference data — the hg38 genome, ENCODE exclusion regions, GTEx eQTL files, dbSNP, the GWAS Catalog, and conversion-unstable position lists — and the preprocessing scripts accept narrowPeak files from any tissue, so the architecture can be retargeted beyond the nervous system.
Neur-Ally is a focused instrument rather than a general foundation model, and that is its argument: for disorders concentrated in one organ, a model trained on that organ's regulatory landscape should outrank a broader one diluted across tissues. The paper does not benchmark that claim head-to-head against pan-tissue predictors, so the case rests on construction rather than comparison, and the low PR-AUC values are a reminder that individual regulatory labels stay hard to predict even when ranking is good. Its practical value lies in the E-value machinery, which converts raw prediction shifts into calibrated significance against a matched null — the step most often left to arbitrary thresholds. Code, pipeline, and trained weights are MIT-licensed on GitHub and archived on Zenodo, and the work is peer-reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.