In silico Alu Mutagenesis Pinpoints the F8 Intron 18 Branch Point

A 331 bp Alu retrotransposon insertion in intron 18 of the Factor VIII gene causes severe hemophilia A by triggering complete exon 19 skipping. The original report (Ganguly et al. 2003) proposed that the insertion disrupts a branch point at position −85, but that location was inferred, not experimentally mapped. In a companion study, we showed that this Alu is uniquely splice-silent — it creates zero cryptic splice sites, acting as a pure spacer. That property makes it an ideal positional probe: by inserting the same element at varying distances from the exon and asking two deep learning models where splicing breaks, we can map the true branch point with single-nucleotide resolution.

Background — The Assumed Branch Point

In a companion study, we showed that both SpliceAI and AlphaGenome correctly predict that a 331 bp Alu Yb9 insertion 19 bp upstream of exon 19 in the F8 gene abolishes splicing — consistent with the complete exon skipping and severe hemophilia A reported by Ganguly et al. (2003). The original report by Ganguly et al. (2003) proposed that the mechanism involved displacement of a putative branch point at position −85, breaking lariat formation.

The companion study also showed that the intact Alu is uniquely splice-silent: it creates zero cryptic splice sites, while nearly half of 5,000 composition-matched shuffles do. The Alu acts as a pure spacer — adding distance between the 3′ splice site and the branch point without introducing competing signals. That property makes it an ideal tool for mapping the branch point: if we insert the same splice-silent element at varying distances from the exon, the position where splicing breaks should mark the branch point location.

What is a branch point?

The branch point is a conserved adenine within the intron whose 2′-OH group attacks the 5′ splice site, forming the lariat intermediate that initiates splicing. In humans, the branch point is typically 18–40 nt upstream of the 3′ splice site.

Why exon 19?

Wang et al. (2020) noted that the natural 3′ss of F8 exon 19 is inefficiently recognized; NNSPLICE and NetGene2 did not predict it at default cutoffs. We scored all 26 F8 exon boundaries with MaxEntScan and found that exon 19 ranks 6th weakest out of 25 F8 splice sites for both its acceptor (7.10) and donor (6.97), both below the >8 threshold typically associated with strong splice sites (Yeo & Burge 2004). This leaves the exon potentially more dependent on proper branch point function than most.

Experimental Design — Position Scan

We inserted the identical Alu Yb9 element (antisense, disease orientation) at 24 positions across intron 18, ranging from −19 (the disease position) to −200. From −19 to −30, we used single-nucleotide resolution to precisely map the disruption boundary. Beyond −30, we used coarser spacing.

Each insertion was scored by two deep learning models:

• SpliceAI — predicts the probability that each position is a splice acceptor or donor (0–1)
• AlphaGenome — predicts splice site usage and junction-level PSI (percent spliced in, 0–1)

Every insertion from −19 through −29 produces near-zero acceptor scores and zero PSI. But at position −30, splicing abruptly recovers: SpliceAI jumps from 0.008 to 0.543, AlphaGenome from 0.012 to 0.512, and PSI from 0.000 to 0.877.

This razor-sharp transition, a single base pair between −29 and −30, is the signature of a branch point. An Alu inserted at position −29 lands within the branch point motif (AAATAAT, positions −32 to −26), splitting it and disrupting lariat formation. An Alu inserted at −30 lands just upstream of the motif, leaving it intact. Both models agree independently, placing the branch point at position −27.

An Alu insertion at position −85, the putative branch point from Ganguly et al., shows no disruption whatsoever. Splicing is fully preserved with scores matching wild type, ruling out −85 as the functional branch point.

Key Numbers

Independent Confirmation — Wang et al. (2020)

Our in silico mapping of the branch point to position −27 is independently corroborated by a clinical genetics study. Wang et al. (2020) reported two unrelated patients with moderate hemophilia A (FVIII:C 1.2% and 2%) caused by the intronic variant c.5999-27A>G — a single nucleotide change at the exact position we identified as the branch point. Sanger sequencing confirmed exon 19 skipping, and SVM-BPfinder predicted branch point disruption. Wang et al. found that four standard splice site prediction algorithms showed no impact — this is a variant class invisible to traditional splice site predictors, yet both deep learning models flag it, consistent with having implicitly learned branch point requirements.

Three unrelated patients now demonstrate exon 19 skipping caused by disruption at position −27, through two entirely different mutational mechanisms: point mutation (Wang et al.) and Alu insertion (Ganguly et al.).

In Silico Validation — BP A>G Mutation

To further confirm position −27, we directly tested the branch point A>G point mutation (c.5999-27A>G) — the exact variant found in Wang et al.'s patients — with both models.

A single nucleotide change at −27 causes a larger splice disruption than Alu insertions at many distant positions (e.g., −150 or −200). This is strong evidence that −27 is the functional branch point: even a point mutation here substantially impairs splicing, while a 331 bp Alu insertion far upstream has no effect.

Systematic Adenine Mutagenesis — Is −27 Really THE Branch Point?

The branch point motif at this locus is AAATAAT (positions −32 to −26), which contains five adenines. Is position −27 uniquely the functional branch point, or could another adenine in the motif serve as the lariat-forming nucleotide?

We tested 10 point mutations in three groups:

1. All three substitutions at −27 (A>G, A>C, A>T) — if −27 is the BP, all should disrupt splicing
2. A>G at other motif adenines (−32, −31, −30, −28) — should not disrupt if −27 is the BP
3. A>G outside the motif (−37, −36, −35, −34) — negative controls

The results unambiguously identify position −27 as the branch point:

• Position −27 is uniquely sensitive — all three substitutions cause disruption, with A>G being the most severe
• A>G > A>C > A>T in disruption severity — consistent with the biochemistry of lariat formation: guanine lacks the 2′-OH geometry required for nucleophilic attack, while pyrimidines (C, T) may retain partial capacity. The fact that both models reproduce this severity ordering — without explicit training on branch point biochemistry — suggests they have learned genuine sequence-function relationships, not merely memorized common variants
• No other adenine in the motif is a branch point — A>G at positions −32, −31, −30, and −28 show less than 10% effect on both models
• Outside controls confirm specificity — A>G mutations at −37, −36, −35, and −34 have negligible effects

The Branch Point Motif — GC Replacement

The branch point consensus is YNYYRAY (7 nt), where Y = pyrimidine, R = purine, and A = branch point adenine. At F8 intron 18, positions −32 to −26, the sequence is AAATAAT, where the underlined A at −27 is the branch point adenine.

To test whether the entire motif is required (not just the branch point adenine), we replaced all 7 nucleotides with a GC-rich sequence:

The GC replacement produces the strongest acceptor score reduction of any perturbation tested, consistent with complete branch point destruction. PSI is not computable because AlphaGenome's junction model does not predict the E18→E19 splice when the branch point motif is absent — the 3′ splice site is effectively non-functional.

However, AlphaGenome's junction predictions reveal that an alternative junction appears with an acceptor 39 bp upstream of the normal position — extending exon 19 into the intron. This alternative acceptor likely uses a different branch point further upstream. The E19 donor retains a relatively high score (SpliceAI: 0.57, AlphaGenome: 0.84) because it still has a paired acceptor via exon definition, just not the canonical one.

Is there a plausible branch point at this upstream position? The intron 18 sequence at approximately −66 contains AACTAAC — a strong match to the YNYYRAY consensus (matching at all 7 positions). This suggests the model may have learned that when the primary branch point is destroyed, a secondary YNYYRAY-matching sequence further upstream can serve as a backup. Whether this alternative branch point is genuinely functional or a model artifact remains an open question.

Summary

Three independent lines of evidence converge on position −27 as the branch point in F8 intron 18:

1. Position scan: The disruption boundary falls between −29 and −30 (24 Alu insertions, 2 models)
2. Point mutagenesis: Only position −27 is sensitive to substitution (10 mutations, 3 substitution types)
3. Clinical genetics: c.5999-27A>G causes hemophilia A via exon 19 skipping (Wang et al. 2020)