Integration catalog¶
The integration type describes what Altar does during an analysis.
| Integration | Type | Output | Execution |
|---|---|---|---|
| ChromBPNet | model binding | four regulatory-effect scores | five GPU folds, then CPU summary |
| Cherimoya | model binding | counts effect and profile distance | configurable GPU folds, then CPU summary |
| Enformer | model binding | top-track SAD/SADR plus selected-track details | independently retryable variant batches |
| Borzoi | model binding | top gene/GTEx logSED plus lossless gene effects | four sequential replicates per retryable variant batch |
| LegNet | model binding | signed MPRA activity effect | CPU-friendly, independently retryable variant batches |
| Sei | model binding | top sequence class plus all 40 class effects | bounded GPU microbatches over retryable variant batches |
| AlphaGenome | hosted model binding | reduced multi-track headline scores | inline API calls |
| GPN-Star | annotation source | calibrated allele-specific LLR values + provenance | indexed lookup in prepared Parquet |
| SpliceAI | annotation source | splice delta scores and per-gene detail | indexed lookup in prepared Parquet |
| AlphaMissense | annotation source | pathogenicity summary and transcript detail | indexed lookup in prepared Parquet |
| gnomAD | annotation source | population allele frequencies, counts, and site filters | lookup in prepared Parquet, or a fused BigQuery read |
| ClinVar | annotation source | germline classification, review stars, conditions, and consequences | lookup in prepared Parquet, or a fused BigQuery read |
| Open Targets E2G | variant–gene link source | typed regulatory-element/gene relations | injected database, file, or service backend |
| ENCODE-rE2G + scE2G atlases | variant–gene link source | versioned typed element/gene evidence | generic in-memory or SQLite link store; no model execution |
Why GPN-Star and SpliceAI are sources here¶
Their published scores exist before an Altar job starts. The current bindings retrieve those scores; they do not run live model inference. This preserves the distinction between computed in this analysis and retrieved from a versioned release.
A future live-inference integration may coexist as a model binding if it defines runtime assets, resource requirements, output equivalence, and reproducible provenance. It should not silently replace lookup behavior.
Install only what you need¶
Bindings and runtimes are independent projects within the repository. Installing one evidence source does not pull in a model framework, and installing a model binding does not install TensorFlow or PyTorch into the Altar environment.