Analysis Workflows

End-to-end, multi-stage workflows that string the individual layers of Agentic-SpliceAI into a single runnable pipeline.

Meta-Layer MLOps (M1-S / M2-S / M3)

The Meta-Layer MLOps Workflow is the golden path for the sequence-level meta-models — from raw GTF/FASTA to a promoted, evaluated checkpoint:

1. Data Preparation — GTF/FASTA → splice-site labels
2. Base Scoring — OpenSpliceAI per-nucleotide predictions
3. Feature Engineering — the 9-modality feature stack
4. Training M1-S and M2-S
5. Evaluation — meta-vs-base, canonical and alternative sites
6. Reporting & Promotion — results, provenance, canonical registry
7. GPU Pods — running it all at genome scale on RunPod

M3 (novel sites) reuses Stages 1–3, then branches into its own sub-series — novel sites have no annotation to label against, so they need separate label curation and an anti-circular evaluation:

1. M3 Label Curation — junctions, long-read, disease catalogs
2. M3 Training — recognizer (M3-S) and candidate refiner (M3-R)
3. M3 Evaluation — per-gene precision@k against independent truth

Related workflows

• Splice Prediction Guide — base-layer prediction
• Data Preparation CLI — data-processing commands
• Agency Patterns — agentic workflow patterns

Planned

• Variant impact assessment
• Agentic validation of M3 novel-site candidates (literature, expression, conservation)