One framework for variant analysis¶

Altar connects genomic variants to model scores and reference evidence without tying an analysis to one model, database, or compute provider.

Get started Browse integrations Build an extension

What Altar provides¶

Canonical variants

Read and validate VCF or TSV variants once. Model bindings and evidence sources consume the same assembly-aware representation.
Comparable model results

Each model declares its score fields and interpretation. Results remain attributable to the exact model instance that produced them.
Reference evidence

Join precomputed scores, annotations, and variant-to-gene relations without pretending every dataset is an executable model.
Portable execution

A model binding creates backend-neutral work. Run it locally, on Kubernetes, on Modal, or with your own execution implementation.

The analysis Altar represents¶

variants
   ├── executable models ──► scores ───────────────┐
   ├── annotation sources ─► reference fields ────┼─► variant results
   └── relation sources ───► variant–gene links ──┘

A variant result preserves the contributions of every model and evidence source. A configurable rule may also mark the variant as prioritized. That flag means a declared rule matched; it does not mean the variant is pathogenic, causal, clinically significant, or statistically significant.

Framework, bindings, and runtimes¶

Altar is maintained as a monorepo with independently installable parts:

Part	Responsibility	Examples
`altar`	Stable interfaces, canonical variant processing, orchestration primitives, reference implementations	model plans, score stores, annotation sources, conformance suites
bindings	Scientific interpretation for one model, dataset, or service	Cherimoya, GPN-Star, SpliceAI, AlphaMissense, Open Targets E2G
runtimes	Heavy or conflicting execution environments	TensorFlow ChromBPNet, PyTorch Cherimoya, containerized variant processing

This separation lets a binding remain lightweight while its runtime carries model frameworks, weights, and GPU requirements.

Choose your path¶

Understand the scientific data model if you consume or interpret results.
Run the real scoring tutorial to build a multi-model, annotated result progressively.
Browse supported integrations before planning an analysis.
Choose an extension interface if you are adding a model, dataset, backend, or store.

Project status

Altar is pre-1.0. The documented façade modules are compatibility-controlled; experimental namespaces and implementation paths may change. See the compatibility policy.