Awesome sequence-to-function

A collection of models and methods that predict functional genomic readouts directly from DNA sequence.

Sequence-to-expression models

• AbExp: predicts tissue-specific aberrant gene expression from DNA sequence variants
• AlphaGenome: unified model predicting many regulatory modalities across 1 Mb of sequence at base resolution
• Basenji: dilated convolutional model predicting CAGE and epigenomic tracks from sequence
• Borzoi: predicts RNA-seq coverage from sequence, including splicing and polyadenylation effects
• Decima: single-cell resolution expression prediction from sequence
• Enformer: transformer model extending the receptive field for gene expression prediction
• enformer-pytorch: community PyTorch implementation with pretrained weights
• EPInformer: gene expression prediction combining sequence, epigenomic signal, and enhancer-promoter contacts
• ExPecto: predicts tissue-specific expression and variant effects from sequence
• LegNet: predicts gene expression and variant effects from short regulatory DNA sequences
• ProCapNet: predicts base-resolution transcription initiation profiles from sequence
• Sei: predicts sequence regulatory activity and assigns it to regulatory classes
• Xpresso: predicts steady-state mRNA levels from promoter sequence

Chromatin accessibility and TF binding

• Basset: learns the regulatory code of accessible DNA with convolutional networks
• basepairmodels: training and interpretation code for base-resolution profile models
• BPNet: base-resolution models of TF binding profiles
• bpnet-lite: lightweight PyTorch reimplementation of BPNet and ChromBPNet
• chromBPNet: bias-factorized, base-resolution models of chromatin accessibility
• CREsted: training and interpreting sequence models of cell-type-specific accessibility
• DanQ: hybrid convolutional and recurrent model of noncoding function
• DeepSEA: early deep model predicting chromatin effects of noncoding variants
• DeepSTARR: predicts enhancer activity measured by STARR-seq
• gkmExplain: efficient importance scores for gapped k-mer SVMs
• lsgkm: large-scale gapped k-mer SVM for regulatory sequence classification
• maxATAC: TF binding prediction from ATAC-seq signal and sequence
• scBasset: sequence-based modeling of single-cell chromatin accessibility
• scPrinter: multi-scale footprinting and TF binding inference from accessibility data
• Selene: PyTorch library for training and applying sequence models (DeepSEA successor framework)

Splicing

• AbSplice: predicts tissue-specific aberrant splicing from DNA sequence variants
• MMSplice / MTSplice: modular models of splicing and tissue-specific splicing effects
• Pangolin: splice site usage prediction across tissues and species
• SpliceAI: predicts splice junctions from primary sequence
• SpliceBERT: pre-trained RNA language model predicting splicing features and variant effects from precursor RNA sequence

RNA stability, UTRs, and translation

• APARENT: predicts alternative polyadenylation from sequence
• APARENT2: residual network version of APARENT for polyadenylation variant effects
• DeepRiPe: predicts RNA-binding protein binding from sequence
• Optimus 5-Prime: 5' UTR design and variant effect prediction on translation
• Orthrus: contrastive mature RNA model for functional RNA property prediction
• Saluki: predicts mRNA half-life from sequence
• UTR-LM: language model of 5' UTRs for translation and expression prediction

3D genome from sequence

• Akita: predicts 3D genome architecture and Hi-C contact maps directly from DNA sequence
• deepC: predicts Hi-C chromatin interactions from sequence
• Orca: predicts multiscale 3D genome folding from sequence
• Puffin: interpretable model of transcription initiation from promoter sequence

Enhancer-gene linking

• ABC model: activity-by-contact enhancer-gene prediction
• Cicero: cis-regulatory co-accessibility links from single-cell accessibility
• GraphReg: chromatin-interaction-aware gene regulation model
• scE2G: enhancer-gene prediction from single-cell multiome data
• SCENIC+: single-cell multiomic inference of enhancer-driven regulatory networks
• SCENT: single-cell enhancer target gene mapping
• TargetFinder: predicts enhancer-promoter interactions from genomic features

Variant effect prediction

• AlphaMissense: proteome-wide missense variant pathogenicity prediction
• CADD: integrative deleteriousness score for variants across the genome
• GPN: genomic pretrained network, including GPN-MSA for variant effect scoring

Model interpretation

• DeepLIFT: importance scores by backpropagating activation differences
• fastISM: fast in-silico mutagenesis for convolutional sequence models
• gopher: evaluation and interpretation of quantitative regulatory sequence models
• SHAP: unified framework for feature attribution
• tangermeme: toolkit for attribution, marginalization, and motif analysis on sequence models
• TF-MoDISco: discovers motifs from attribution scores
• tfmodisco-lite: faster, leaner reimplementation of TF-MoDISco

Regulatory sequence design

• boda2 / Malinois: deep learning design of cell-type-specific regulatory elements
• DDSM: Dirichlet diffusion score model for generating regulatory sequence
• Ledidi: turns trained sequence models into sequence editors
• regLM: language-model-based design of regulatory DNA

Frameworks and tooling

• EUGENe: end-to-end framework for building and evaluating sequence models
• EvoAug: evolution-inspired data augmentation for regulatory sequence models
• gReLU: comprehensive framework for training, interpreting, and designing with sequence models
• Kipoi: model zoo and standardized API for genomics models
• ML4GLand: collection of libraries for sequence-based machine learning in genomics

Benchmarks and evaluation

• BEND: benchmark of DNA language models on realistic genomic tasks
• CAGI: community assessment of genome interpretation methods
• DART-Eval: benchmark of DNA models on regulatory sequence tasks
• Genomic Benchmarks: datasets and baselines for genomic sequence classification

Reviews

• Deep learning: new computational modelling techniques for genomics
• Obtaining genetics insights from deep learning via explainable artificial intelligence
• Current sequence-based models capture gene expression determinants in promoters but mostly ignore distal enhancers
• Leveraging genomic deep learning models for non-coding variant effect prediction
• Advancing regulatory variant effect prediction with AlphaGenome
• AlphaGenome, a Swiss-army knife for exploring non-coding DNA