Isoform Discovery: Context-Aware Splice Prediction¶

Status: Research & Planning Phase
Goal: Discover novel isoforms induced by variants, disease, stress, and other external factors

🎯 Vision¶

Problem: Current annotations (MANE, Ensembl) capture known isoforms, but miss: - Disease-specific isoforms - Variant-induced alternative splicing - Stress-response transcripts - Tissue-specific rare isoforms - Developmental stage-specific variants - Environmental condition-specific splicing

Solution: Use Meta-SpliceAI's adaptive prediction to discover novel isoforms by: 1. Meta Layer: Detect context-dependent splice sites beyond canonical annotations 2. Agentic Layer: Validate predictions with literature and experimental evidence 3. Integration: Combine with RNA-seq, clinical variants, and disease databases

📊 Current State vs Future Vision¶

Current: Canonical Isoform Prediction¶

Base Layer (MANE)
    ↓
Predict 1-2 canonical transcripts per gene
    ↓
Validated against known annotations

Coverage: ~44 splice sites per gene (BRCA1 example)

Future: Context-Aware Isoform Discovery¶

Base Layer (MANE baseline)
    ↓
Meta Layer (adaptive refinement)
    ↓
Novel Splice Site Detection
    ↓
Isoform Reconstruction
    ↓
Agentic Validation (literature + experimental)
    ↓
Novel Isoform Catalog

Potential Coverage: 1,218+ splice sites per gene (Ensembl-level + novel discoveries)

🏗️ Proposed Architecture¶

Layer 1: Meta Layer - Novel Splice Site Detection¶

Goal: Identify high-confidence splice sites beyond MANE canonical set

Components:

Delta Score Analysis
Compare meta-layer predictions to base-layer
High delta = context-dependent splice site
Flag sites with high confidence but not in MANE
Context Clustering
Group similar contexts (same variants, disease, tissue)
Identify context-specific splice patterns
Discover condition-dependent isoforms
Isoform Reconstruction
Assemble detected splice sites into full transcripts
Validate transcript structure (ORF, NMD rules)
Generate novel isoform annotations
Confidence Scoring
Multi-factor confidence score:
- Meta-layer prediction strength
- Conservation across species
- RNA-seq evidence (if available)
- Splice motif strength
- Consistency with biological rules

Output:

Novel Splice Sites:
- Position: chr17:43,067,608 (donor)
- Confidence: 0.92
- Context: BRCA1-c.68_69del variant
- Evidence: High delta score (0.45), strong motif
- Status: Candidate for validation

Layer 2: Agentic Layer - Evidence Aggregation¶

Goal: Validate novel isoforms with external evidence

Components:

Literature Research Agent
Query PubMed for reports of:
- Novel isoforms in gene of interest
- Disease-associated alternative splicing
- Functional impact of splice variants
Extract evidence from papers
Build evidence graph
Database Integration Agent
RNA-seq Data:
- GTEX (tissue-specific expression)
- Cancer atlases (tumor-specific isoforms)
- Disease cohorts
Variant Databases:
- ClinVar (pathogenic splice variants)
- gnomAD (population-level splice variants)
Isoform Databases:
- GENCODE (comprehensive annotations)
- RefSeq (curated transcripts)
Validation Workflow Agent
Propose validation experiments:
- RT-PCR primers for novel junction
- Minigene assay design
- CRISPR screens for splice modifiers
Prioritize candidates by:
- Clinical relevance
- Functional impact
- Validation feasibility
Evidence Aggregation
Combine predictions with external evidence
Score novel isoforms by evidence strength:
- RNA-seq support: High (direct evidence)
- Literature support: Medium (indirect evidence)
- Prediction only: Low (needs validation)

Output:

Novel Isoform Report:
- Gene: BRCA1
- Novel Junction: chr17:43,067,608-43,082,575
- Context: Tumor samples, BRCA1-mutant
- Meta-Layer Confidence: 0.92
- RNA-seq Evidence: 15/120 tumor samples (12.5%)
- Literature: 3 papers report similar junction
- Validation Priority: HIGH (clinical relevance)
- Proposed Experiment: RT-PCR primers designed

🔬 Data Integration Strategy¶

Primary Data Sources¶

RNA-seq Data
GTEX: Tissue-specific isoforms (53 tissues, 17,382 samples)
TCGA: Cancer-specific isoforms (33 cancer types)
GTEx + TCGA junction files: Direct evidence of novel splice junctions
Custom cohorts: Disease-specific RNA-seq
Variant Data
ClinVar: Pathogenic splice variants
gnomAD: Population-level splice variants
COSMIC: Cancer-associated mutations
Personal genomes: Individual-level predictions
Epigenetic Data
ENCODE: Histone marks, chromatin state
Roadmap Epigenomics: Tissue-specific epigenomes
ChIP-seq: Splicing factor binding
Conservation Data
PhyloP: Evolutionary conservation
PhastCons: Conserved elements
Cross-species comparison: Mouse, zebrafish homologs

Data Processing Pipeline¶

1. Load Base Predictions (MANE canonical)
   ↓
2. Load Meta Predictions (context-aware)
   ↓
3. Compute Delta Scores
   ↓
4. Filter High-Confidence Novel Sites
   ↓
5. Cross-reference with RNA-seq Junctions
   ↓
6. Query Literature for Evidence
   ↓
7. Score and Rank Candidates
   ↓
8. Generate Discovery Reports

📋 Use Cases¶

1. Disease-Specific Isoforms¶

Scenario: Identify breast cancer-specific BRCA1 isoforms

Workflow:

# Predict on tumor samples with BRCA1 variants
agentic-spliceai meta predict \
  --gene BRCA1 \
  --variants tumor_variants.vcf \
  --context tumor \
  --discover-isoforms

# Validate with TCGA RNA-seq
agentic-spliceai isoform validate \
  --predictions brca1_novel_isoforms.tsv \
  --rnaseq-cohort TCGA-BRCA \
  --output validation_report.html

Expected Output: - 10-20 high-confidence novel splice sites - 3-5 novel isoforms with RNA-seq support - Literature evidence for functional impact - Clinical relevance scores

2. Variant-Induced Splicing Changes¶

Scenario: Patient has VUS (variant of uncertain significance) near splice site

Workflow:

# Predict impact of variant on splicing
agentic-spliceai variant predict \
  --vcf patient_vus.vcf \
  --gene-list candidate_genes.txt \
  --detect-novel-junctions

# Research evidence
agentic-spliceai agentic research \
  --novel-junctions detected_junctions.tsv \
  --query "pathogenic splicing impact" \
  --sources pubmed,clinvar,splicevardb

Expected Output: - Novel splice junctions induced by variant - Predicted functional impact (NMD, truncation, etc.) - Literature evidence for similar variants - Clinical interpretation (likely pathogenic, benign, etc.)

3. Tissue-Specific Isoform Discovery¶

Scenario: Discover brain-specific isoforms relevant to neurological disease

Workflow:

# Compare across tissues
agentic-spliceai isoform discover \
  --gene-list neurological_genes.txt \
  --tissues brain,cerebellum,cortex \
  --reference-tissues heart,liver,kidney \
  --rnaseq-source GTEX

# Validate with brain RNA-seq
agentic-spliceai isoform validate \
  --candidates brain_specific_isoforms.tsv \
  --rnaseq-cohort GTEX-Brain \
  --min-samples 5

Expected Output: - Brain-specific isoforms not in MANE - Expression patterns across brain regions - Developmental trajectories - Disease associations

🛠️ Implementation Roadmap¶

Phase 5 (Meta Layer) - Foundation¶

Add to existing meta-layer work: - ✅ Delta score computation (already exists in ValidatedDeltaPredictor) - 🆕 Novel splice site detector - 🆕 Confidence scoring module - 🆕 Context clustering

Deliverables: - Detect novel splice sites with high confidence - Score sites by multiple evidence lines - Group by context (variants, disease, tissue)

Phase 8 (NEW) - Isoform Discovery & Validation¶

Dedicated phase for isoform discovery:

8.1: Meta Layer Extensions¶

Isoform reconstruction from splice sites
Transcript structure validation (ORF, NMD)
Novel isoform annotation format
Integration with base predictions

8.2: RNA-seq Integration¶

GTEX junction file parser
TCGA junction file parser
Custom cohort support
Junction-level validation

8.3: Agentic Validation¶

Isoform research agent (literature mining)
Evidence aggregation (multi-source)
Validation workflow generator
Discovery report generation

8.4: Clinical Applications¶

Variant-induced isoform prediction
Disease-specific isoform catalog
VUS interpretation workflow
Therapeutic target identification

Estimated Timeline: 4-6 weeks (after Phase 7)

📊 Success Metrics¶

Discovery Metrics¶

Novel isoforms discovered: Target 100+ across 50 genes
RNA-seq validation rate: >70% with junction support
Literature confirmation: >50% with supporting papers
Functional impact: >30% affect protein function

Validation Metrics¶

Precision: % of predictions with RNA-seq support
Recall: % of known rare isoforms detected
Clinical utility: Impact on VUS interpretation
Novel discoveries: Isoforms not in any database

🔬 Research Questions¶

Biological Questions¶

How common are context-dependent isoforms?
Proportion of genes with disease-specific isoforms
Tissue-specificity of alternative splicing
Variant-induced splicing changes
What contexts induce novel splicing?
Stress response
Developmental stages
Disease progression
Therapeutic interventions
Can we predict functional impact?
NMD escape
Protein domain loss/gain
Clinical significance

Technical Questions¶

What confidence threshold for novel sites?
Balance precision vs recall
Context-dependent thresholds
Multi-factor scoring
How to reconstruct full isoforms?
Splice site pairing
Exon inclusion/exclusion
ORF prediction
How to validate computationally?
RNA-seq evidence strength
Conservation requirements
Motif scoring

Literature¶

Disease-specific splicing:
Cancer-specific isoforms (TCGA studies)
Neurological disease splicing (autism, ALS)
Cardiac disease isoforms (cardiomyopathy)
Computational approaches:
LeafCutter (differential splicing)
MAJIQ (local splice variation)
rMATS (alternative splicing events)
SUPPA2 (isoform quantification)
Validation methods:
RT-PCR validation
Long-read sequencing (PacBio, Nanopore)
Minigene assays

Databases¶

GENCODE: Comprehensive gene annotations
RefSeq: Curated transcript sequences
APPRIS: Principal isoform annotations
IsoformAtlas: Tissue-specific isoforms
SpliceVault: Alternative splicing database

💡 Key Innovations¶

1. Context-Aware Discovery¶

Beyond static annotations: Predict isoforms specific to: - Individual genetic backgrounds - Disease states - Tissue/cell types - Environmental conditions

Combine: - Base-layer predictions (canonical) - Meta-layer predictions (adaptive) - RNA-seq data (direct evidence) - Literature (biological context) - Conservation (evolutionary support)

3. Agentic Validation¶

LLM-powered: - Hypothesis generation - Evidence synthesis - Experimental design - Report generation

🎯 Next Steps¶

Immediate (Research Phase)¶

Literature review:
Survey isoform discovery methods
Identify key use cases
Define success criteria
Data inventory:
Available RNA-seq datasets
Variant databases
Existing novel isoform catalogs
Prototype design:
Novel splice site detector
Confidence scoring
Validation workflow

Short Term (Phase 5 Integration)¶

Extend meta layer:
Add delta score analysis
Implement confidence scoring
Create novel site detector
Test on known cases:
Disease-associated isoforms
Variant-induced changes
Tissue-specific examples

Long Term (Phase 8)¶

Full implementation:
Isoform reconstruction
RNA-seq integration
Agentic validation
Clinical workflows
Validation study:
Experimental validation
Clinical cohorts
Publication

📁 Documentation Structure¶

docs/isoform_discovery/
├── README.md                          ← This file - Vision & overview
├── ARCHITECTURE.md                    ← Technical architecture
├── USE_CASES.md                       ← Detailed use cases
├── DATA_INTEGRATION.md                ← Data sources & formats
├── VALIDATION_METHODS.md              ← How to validate discoveries
└── RESEARCH_QUESTIONS.md              ← Open questions & experiments

🤝 Collaboration Opportunities¶

Academic Partners¶

Tissue-specific isoform experts
Disease cohort access
Experimental validation labs

Clinical Partners¶

VUS interpretation needs
Patient cohorts
Clinical validation

Industry Partners¶

RNA-seq datasets
Therapeutic target discovery
Diagnostic development

Status: 🔬 Research & Planning
Next: Extend meta layer (Phase 5) with novel site detection
Long-term: Dedicated isoform discovery phase (Phase 8)

Questions? Ideas? Open a GitHub issue or discussion.

Isoform Discovery: Context-Aware Splice Prediction¶

🎯 Vision¶

📊 Current State vs Future Vision¶

Current: Canonical Isoform Prediction¶

Future: Context-Aware Isoform Discovery¶

🏗️ Proposed Architecture¶

Layer 1: Meta Layer - Novel Splice Site Detection¶

Layer 2: Agentic Layer - Evidence Aggregation¶

🔬 Data Integration Strategy¶

Primary Data Sources¶

Data Processing Pipeline¶

📋 Use Cases¶

1. Disease-Specific Isoforms¶

2. Variant-Induced Splicing Changes¶

3. Tissue-Specific Isoform Discovery¶

🛠️ Implementation Roadmap¶

Phase 5 (Meta Layer) - Foundation¶

Phase 8 (NEW) - Isoform Discovery & Validation¶

8.1: Meta Layer Extensions¶

8.2: RNA-seq Integration¶

8.3: Agentic Validation¶

8.4: Clinical Applications¶

📊 Success Metrics¶

Discovery Metrics¶

Validation Metrics¶

🔬 Research Questions¶

Biological Questions¶

Technical Questions¶

📚 Related Work¶

Literature¶

Databases¶

💡 Key Innovations¶

1. Context-Aware Discovery¶

2. Multi-Modal Evidence Integration¶

3. Agentic Validation¶

🎯 Next Steps¶

Immediate (Research Phase)¶

Short Term (Phase 5 Integration)¶

Long Term (Phase 8)¶

📁 Documentation Structure¶

🤝 Collaboration Opportunities¶

Academic Partners¶

Clinical Partners¶

Industry Partners¶