Skip to content

In Silico Saturation Mutagenesis and Splice Variant Validation

A planned application for systematically predicting splice-altering positions across entire genes and validating predictions against experimental evidence from SpliceVarDB and GTEx RNA-seq junction data.

Status: Planned (Phase 3 deliverable) Prerequisites: M4 Phase 1A+1B (complete), Phase 2 ClinVar integration

1. Motivation

Clinical sequencing frequently identifies variants of uncertain significance (VUS) in disease genes. A key question for each VUS: does this variant disrupt splicing?

Currently, this is answered case-by-case: run the variant through SpliceAI or our meta-layer, check the delta score, consult ClinVar. But for a gene like MYBPC3 (35 exons, ~21K coding positions), there are ~63K possible SNVs — each potentially splice-altering.

In silico saturation mutagenesis systematically scans every position, predicting which mutations would disrupt splicing, where the cryptic splice sites would appear, and what the consequence would be. The result is a splice vulnerability map for the entire gene.

2. Workflow

Step 1: Saturation Scan

For every position in the gene (or a focused region), introduce all possible single-nucleotide substitutions and run the variant effect pipeline:

For each position p in [gene_start, gene_end]:
    ref = FASTA[chrom][p]
    For each alt in {A, C, G, T} - {ref}:
        delta = VariantRunner.run(chrom, p, ref, alt, gene, strand)
        If max(|delta|) > threshold:
            consequence = SpliceEventDetector.analyze(delta)
            Store: (p, ref, alt, delta_scores, consequence, cryptic_positions)

Step 2: Build Splice Vulnerability Map

Aggregate results into a per-position vulnerability score:

vulnerability[p] = max over all alt alleles of max(|delta|)

Positions with high vulnerability are "splice-sensitive" — any mutation there is predicted to disrupt splicing. The map reveals:

  • Invariant splice sites (GT/AG): vulnerability ~1.0 (expected)
  • Extended splice motifs (-3 to +6 for donors, -20 to +3 for acceptors): vulnerability 0.3-0.8
  • Exonic splice enhancers/silencers (ESE/ESS): vulnerability 0.1-0.5 (sequence motifs that regulate splice site selection)
  • Deep intronic positions: rare but occasionally high vulnerability (positions that create cryptic exons when mutated)

Step 3: Cross-Validate with SpliceVarDB

SpliceVarDB is a curated database of experimentally validated splice variants classified as "splice-altering" or "non-splice-altering" based on RNA-seq, minigene assays, or clinical evidence.

For each variant in SpliceVarDB for this gene:
    Look up our saturation scan prediction at (chrom, pos, ref, alt)
    Compare:
        Our prediction: splice-altering (delta > 0.2)?
        SpliceVarDB:    splice-altering?

    Agreement types:
        True positive:  both say splice-altering
        True negative:  both say non-splice-altering
        False positive: we predict splice-altering, SpliceVarDB says no
        False negative: we miss it, SpliceVarDB says splice-altering

This gives us precision, recall, and AUROC against experimental ground truth — not just at the gene level, but at the per-position level.

Step 4: Validate Cryptic Site Positions with GTEx Junctions

For each position where we predict a cryptic splice site gain:

For each predicted cryptic site (gain event at position X):
    Query GTEx junction parquet for junctions involving position X
    If junction reads found:
        → Biological validation: the cryptic site is used in vivo
        Record: tissue(s), read count, PSI
    If no junction reads:
        → Either our prediction is wrong, or the variant doesn't
          exist in the GTEx cohort (most likely for rare variants)

This leverages the 353K GTEx junctions across 54 tissues already wired into our pipeline via the junction modality.

3. Expected Outputs

Per-Gene Splice Vulnerability Map

Gene: MYBPC3 (chr11:47,331,406-47,352,702, minus strand)
Total positions scanned: 21,296
Splice-sensitive positions (Δ > 0.2): 847 (4.0%)
Splice-destroying positions (Δ > 0.5): 312 (1.5%)

Hotspots:
  Exon 25 donor boundary:  15 sensitive positions in 20bp window
  Exon 13 acceptor region: 22 sensitive positions in 30bp window
  Intron 11 (c.1227-13):   Deep intronic cryptic acceptor

SpliceVarDB Concordance

Gene: MYBPC3
SpliceVarDB variants evaluated: 47
  True positives:  31 (splice-altering, correctly predicted)
  True negatives:  12 (non-splice-altering, correctly classified)
  False positives:  2 (we over-predict)
  False negatives:  2 (we miss)

Precision: 0.94
Recall:    0.94
AUROC:     0.97

GTEx Junction Validation

Predicted cryptic sites with GTEx junction support:
  Position 47343155 (cryptic acceptor): 27 junction reads in heart tissue
  Position 47333227 (cryptic donor):    4 junction reads in muscle
  Position 47332576 (shifted donor):    12 junction reads in 3 tissues

Validation rate: 3/5 predicted cryptic sites have GTEx support (60%)

4. Compute Requirements

Gene size Positions SNVs (×3) Time (CPU) Time (GPU)
Small (5K bp) 5,000 15,000 ~4 hours ~12 min
Medium (20K bp) 20,000 60,000 ~17 hours ~50 min
Large (100K bp) 100,000 300,000 ~3.5 days ~4 hours

Optimization opportunities: - Batched base model inference: Run multiple variants through the base model simultaneously (current pipeline processes one at a time) - Sparse storage: Only store positions with Δ > threshold (typically <5% of all positions), reducing storage 20x - Focused scanning: Scan only exon-proximal regions (±500bp from exon boundaries) for the first pass, then deep intronic for high-value genes

5. Relationship to Other Phases

Phase 1A (VariantRunner)          ← foundation, complete
Phase 1B (SpliceEventDetector)    ← consequence classification, complete
Phase 2  (ClinVar integration)    ← pathogenicity labels for benchmarking
Phase 3  (Saturation mutagenesis) ← THIS APPLICATION
Phase 5  (Agentic interpretation) ← automated PubMed validation

Phase 3 builds directly on Phase 1A+1B (which provide the per-variant delta and consequence prediction) and Phase 2 (which provides the ClinVar pathogenicity labels for benchmarking alongside SpliceVarDB).

6. Existing Infrastructure

Component Status Location
VariantRunner (single variant) Complete meta_layer/inference/variant_runner.py
SpliceEventDetector Complete meta_layer/inference/splice_event_detector.py
SpliceVarDB loader Complete meta_layer/data/splicevardb_loader.py
GTEx junction data Complete features/modalities/junction.py
ClinVar loader Planned (Phase 2) meta_layer/data/clinvar_loader.py
Batched variant inference Not started Needed for compute efficiency

7. Clinical Impact

VUS Reclassification

For a patient with a MYBPC3 VUS at position X: 1. Look up position X in the pre-computed vulnerability map 2. If vulnerability > 0.5: strong evidence for splice disruption 3. Cross-reference with SpliceVarDB: has this position been experimentally validated? 4. Check GTEx: does RNA-seq show junction evidence at the predicted cryptic site? 5. Combine evidence for ACMG classification (PS3/BS3 criteria for functional studies)

Drug Target Discovery (ASO Design)

Antisense oligonucleotides (ASOs) can modulate splicing by blocking splice sites or regulatory elements. The vulnerability map identifies: - Which splice sites to target (high-vulnerability donors/acceptors) - Which positions to avoid (mutations at these positions would disrupt the ASO's intended effect) - Cryptic sites that could be blocked to restore normal splicing

References

  • Jaganathan et al. (2019). "Predicting Splicing from Primary Sequence with Deep Learning." Cell, 176(3), 535-548. (Saturation mutagenesis in Figure 1D)
  • Anna & Bhatt (2018). "Splicing mutations in human genetic disorders: examples, detection, and confirmation." J Appl Genetics, 59, 253-268.
  • SpliceVarDB — curated database of experimentally validated splice variants with pathogenicity classification
  • Findlay et al. (2018). "Accurate classification of BRCA1 variants with saturation genome editing." Nature, 562, 217-222. (Wet-lab saturation mutagenesis of BRCA1)