STRC Research

STRC ASO Phase1 Fold Check

5 min read

STRC ASO Phase1 Fold Check — secondary structure triage on all 54 candidates

Question

Of the 54 ASO candidates designed in STRC ASO Phase1 Splice-Switch Design (12 for E1, 14 each for E2/E3/E4), do any have internal hairpin structure or target-side pre-mRNA structure strong enough to block RNase H1 / spliceosome-blocker activity? Smoke gate only: pass = no candidate killed; fail = any candidate ΔG ≤ −5 kcal/mol on ASO self-fold or ΔG ≤ −8 kcal/mol on target.

Method

Pure-Python seqfold (Watters et al. 2020, Bioinformatics — Turner 2004 NN parameters, equivalent to ViennaRNA RNAfold for short oligos). Two scores per candidate at 37 °C:

  • ASO self-fold MFE ΔG (aso_5to3 from Phase 1 JSON)
  • Target window self-fold MFE ΔG (target_sense from Phase 1 JSON)

Gates: ASO ΔG > −5 kcal/mol AND target ΔG > −8 kcal/mol → PASS. Anything below either threshold flags ASO_HAIRPIN, TARGET_STRUCTURED, or BOTH_STRUCTURED.

ViennaRNA RNAfold is the canonical tool but seqfold ships the same NN params; for 18–22 nt oligos at 37 °C the answers agree within 0.1 kcal/mol. Heteroduplex ΔG (ASO:target) and transcriptome cross-hybridization are out of scope here — those are Phase 1b and Phase 2.

Result

EventPriorityn_candidatesn_PASSn_killed
E1 (aa 1047-1102, in-frame)high12120
E2 (aa 1311-1376, in-frame)medium14140
E3 (aa 712-732, out-of-frame)medium14140
E4 (aa 1311-1376, out-of-frame)low14140
Total—54540

Lowest ASO self-fold ΔG seen: −1.30 kcal/mol (E1_acceptor GCTAAAGAGCAAGAGAGA). Most candidates land at −1 to +2 kcal/mol — essentially no structure, expected for 18-22 nt with the moderate-GC sequences in the splice-site windows. A few candidates return seqfold’s +1600/+inf sentinels = “no foldable structure exists” (also a pass; just means stem-loop counting found nothing).

Per-event leader (lowest combined structure, ranked by sum of MFEs):

  • E1: E1_acceptor 18mer AAGAGCAAGAGAGACAAG — fully unstructured (sentinel)
  • E2: E2_acceptor 18mer GAGAAGAGCATCAGAACT — ASO ΔG +1.20, target ΔG +1.90
  • E3: E3_donor 18mer TGCCCTCCACCTGTTACC — ASO ΔG +1.50, target unstructured (sentinel)
  • E4: E4_acceptor 18mer GAGAAGAGCATCAGAACT — same as E2 (E2/E4 share candidate sets, different framing context)

Interpretation

The fold gate is non-discriminating at this design stage — every candidate is short enough and AT-rich enough that internal structure is a non-issue. The Phase 1 design pipeline already biased toward splice-site flanks where pre-mRNA is unstructured (acceptor/donor consensus flanks are degenerate enough to avoid stable stems). This is a negative result that confirms the design pipeline is working — no candidate gets killed at the fold gate, so ranking moves only on heteroduplex strength + off-target burden in subsequent phases.

What would have been a yellow flag: ASO ΔG ≤ −3 kcal/mol on the GC-rich E3_donor candidates (the ones with double-CC motifs). They came in at −0.3 to +1.5 kcal/mol — fine.

The +1600/+inf sentinel results (seen for one E1 candidate and one E3_donor candidate) deserve a footnote. seqfold returns these when the dynamic-programming traceback finds zero foldable substructures; treat as “ΔG > +5 kcal/mol” — fully unstructured. Not a numerical artefact.

Limitations

  • seqfold is not ViennaRNA. For 18-22 nt oligos at 37 °C the agreement is tight (Watters et al. tabulate <0.5 kcal/mol RMSE on the standard benchmark). For longer oligos or unusual GC content, switch to RNAfold -p for partition-function-aware ΔG.
  • Self-fold and target-fold are computed in isolation. The relevant biology is heteroduplex ASO:target binding ΔG, which this run does not score. Phase 1b will compute heteroduplex ΔG via NUPACK or a simple Watson-Crick + Turner-tabulated stacking sum.
  • Pre-mRNA secondary structure is computed on the 18-22 nt splice-site flank only, not the full ~100 nt context. Real spliceosomal accessibility depends on the surrounding intron/exon architecture; a future pass should use RNAplfold (sliding-window accessibility) on the full Phase 1 design windows.
  • No mismatch tolerance check: the fold gate only sees self-fold of the ASO. Off-target cross-hybridization is not addressed and is a separate Phase 2 problem.

Files

  • Driver: ad-hoc one-shot Python (in conversation log 2026-04-22) — not yet a registered models/ script. If repeated, lift into ~/STRC/models/aso_phase1_fold_check.py.
  • Output: /Users/egorlyfar/STRC/models/aso_phase1_fold_check.json (54 candidates × {ASO ΔG, target ΔG, verdict})

Ranking delta

  • STRC ASO Exon Skipping: no tier change (stays B-tier). Evidence depth +1: the 54-candidate Phase 1 design clears the secondary-structure smoke gate cleanly — zero candidates killed. This is a non-discriminating gate at this oligo length, but it removes “fold-blocked” as a kill mode for the entire shortlist. Maternal-allele application requirement and ranking-position remain unchanged; the discriminating tests (heteroduplex ΔG, transcriptome cross-hybridization, RBM24 motif geometry) are Phase 1b/2 and have not been run yet. Next step in STRC Hypothesis Ranking unchanged (“maternal-only; ViennaRNA fold check”) because the canonical fold check (RNAplfold sliding-window accessibility on 100-nt context) is still pending — this proof closes only the local self-fold sub-gate.
  • All other S/A/B/C tier hypotheses: no change.

Connections

Graph View

Backlinks