Enformer
The transformer that reads 200 kilobases of DNA to predict gene expression and variant effects.
About
Enformer (Avsec et al., Nature Methods 2021), built by DeepMind with Alphabet's Calico, predicts gene expression and chromatin states directly from DNA sequence. It takes 196,608 base pairs of input and predicts 5,313 human and 1,643 mouse genomic tracks at 128-base-pair resolution, covering CAGE, DNase, and ChIP-seq assays. By using transformer modules instead of dilated convolutions, it sees regulatory elements up to 100 kilobases away, five times the receptive field of its predecessor Basenji2.
It became the reference model for a wave of regulatory-genomics work and variant-effect prediction, and its weights are openly available under Apache 2.0. Newer models like Borzoi have since improved on it for variant-effect tasks, but Enformer remains the canonical baseline everything is compared against. There is no hosted API; you run the weights yourself.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (Avsec et al., Nature Methods 2021): outperformed Basenji2 on gene expression prediction across genes and CAGE experiments; 5x larger receptive field (100 kb vs 20 kb); statistically significant gains across all 5,313 human tracks (paired Wilcoxon p < 1e-38)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: Nature Methods 2021 paper (DOI 10.1038/s41592-021-01252-x), DeepMind's research blog, and the deepmind-research repo confirm the architecture, the 200 kb input window, and the Apache 2.0 weights. Strengths: foundational regulatory-genomics model, permissive license, open weights, the field's standard baseline for years. Docked for: (1) aging architecture with stronger successors (Borzoi) for variant-effect tasks; (2) heavyweight at 236M parameters for routine use; (3) no hosted API or maintained product surface. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Foundational regulatory-genomics model; the field's standard baseline for years.
- · Aging architecture with stronger successors (Borzoi) for variant-effect tasks.
- · No hosted API or maintained product surface.