Skip to content

Novel Isoform Discovery (M3)

Goals served: isoform discovery (primary project goal)

Tier: Incubating

Last updated: 2026-04

Problem

Canonical annotations (MANE, RefSeq) capture only ~10% of biologically active splice sites. The remaining ~90% include disease-specific, tissue-specific, and variant-induced isoforms — the targets of greatest scientific and therapeutic interest. The M3 task is to identify these novel splice sites that are absent from canonical annotations but supported by orthogonal evidence (cross-species conservation, RNA-seq junction reads, chromatin features).

Critically, the M3 model must be trained with junction reads held out as a label, not a feature — otherwise it leaks information about the very sites it should discover.

User-facing functionality (planned)

  • Predict high-confidence novel splice sites genome-wide, ranked by multimodal evidence
  • Filter by tissue context (brain, immune, cancer) via GTEx junction overlap
  • Produce per-site evidence cards (conservation, RBP binding, chromatin, foundation-model embeddings)
  • Feed ranked candidates into downstream isoform assembly and agentic validation

Driving examples (planned)

  • (none yet — M3 training not started)

The existing meta_m3_novel.yaml feature config (junction excluded) provides the feature-pipeline side; the training and inference scripts are pending.

src/ surface (existing, ready for M3 training)

  • agentic_spliceai.splice_engine.features.* — already supports junction-as-label via the meta_m3_novel config profile
  • agentic_spliceai.splice_engine.meta_layer.core.feature_schema — M3 variant already scoped
  • Training / inference modules — to be added, likely extending the existing sequence_model and training/* infrastructure

Evaluation (planned)

  • Dataset: MANE-negative, annotation-rich positions with junction support held out as label
  • Baselines: base model (canonical only), M1-S v2, random forest on conservation + chromatin
  • Key metrics: recall of GTEx-supported non-canonical junctions, conservation enrichment of top-K candidates, literature-confirmed rate
  • Validation: cross-tissue RNA-seq junction verification (GTEx v8 already available at data/mane/GRCh38/junction_data/)

Maturity tier and signals

Current tier: Incubating

Signals supporting the tier:

  • Feature config (meta_m3_novel.yaml) exists with junction excluded
  • GTEx v8 junction data aggregated (353K junctions, 54 tissues)
  • M3-S training architecture scoped in meta-layer design docs
  • No scripts yet, no results yet — pre-implementation

Graduation signals

To advance to Active, the application needs:

  • First training script working end-to-end (even at small scale)
  • Baseline evaluation against random forest or M1-S-trained-on-negatives
  • Reproducible benchmark on held-out chromosome(s)
  • Initial results committed to examples/meta_layer/results/

Known limitations (anticipated)

  • Label definition is non-trivial: "novel" is a negative space (MANE-negative but evidence-positive). Label noise will dominate.
  • GTEx junction support varies by tissue and depth — low-expression novel sites will be missed
  • Adjacent application (M4 induced sites) overlaps conceptually but has a different training signal