BGPT: Paper Review: RNA G-quadruplexes mediate cooperativity in HNRNPH binding and splicing regulation

Explore by Goal

Quick Explanation Copied

Core claim: RNA G-quadruplexes (rG4s) “unfolded” by HNRNPH enable indirect cooperativity, producing switch-like (high Hill) control of alternative splicing across hundreds of exons.

Skeptical highlight: The strongest causal chain is for specific minigene constructs and in vitro biophysics; the patient evidence is correlational (expression + subtype clustering + variant association), so generalizing mechanism-to-clinic needs orthogonal validation.

Long Explanation

Paper review (visual, skeptical, evidence-based)

Title: RNA G-quadruplexes mediate cooperativity in HNRNPH binding and splicing regulation
Paper DOI: 10.64898/2026.03.03.709289
Paper date: March 03, 2026

One-sentence claim

rG4-forming HNRNPH binding sites enable cooperative HNRNPH binding (via rG4 unfolding/multivalent backfolding), translating to switch-like splicing regulation across hundreds of exons, and linking rG4 disruption + HNRNPH2 dynamics to breast-cancer splicing patterns.

Best-supported mechanism link

Necessity/sufficiency: AKR1A1 minigene shows downstream rG4 overlap is needed; non-rG4 binding site doesn’t rescue regulation; inserting an rG4-binding site restores cooperativity.
Biophysical causality: FRET and binding titrations show rG4s are unfolded by HNRNPH and binding is cooperative only when rG4 formation is allowed (GTP vs 7-deaza-GTP).

Transcriptome-wide HNRNPH titration: regulated splicing event landscape

Counts of significant alternative splicing events across HNRNPH knockdown/overexpression titration (MCF7).

Event counts are from the paper’s RNA-seq titration analysis.

Cooperativity prevalence (Hill nH≥2) is high

Fraction of regulated events classified as cooperative using dose-response Hill fits.

Cooperative fractions are explicitly stated for CE/IR/ALE classifications in the paper.

Illustrative Hill coefficients (steepness)

Examples highlighted in the paper include very steep cooperative repression/enhancement.

Example nH values are reported in the paper text.

RTstop profiling validates rG4 formation—and shows prediction error

Quantitative comparison of in silico rG4 predictions vs experimental detection in the adapted RTstop IVT library.

The paper reports: BS+rG4 regions with at least one experimentally assigned rG4 = 687/982 (70.0%); BS-only constructs with measurable rG4 = 250/996 (25.1%).

Patient evidence sources (cohort scale)

PMBB for variant association and TCGA BRCA for subtype clustering and expression correlation.

The paper states PMBB had 13,473 rG4-disrupting variants tested (after filtering) and TCGA BRCA analysis was on 1,177 samples (1,069 tumor; 108 normal).

Mechanistic reconstruction (what the paper demonstrates)

1) Cooperativity is common at the splicing level

The study uses a titration series of HNRNPH KD/OE in MCF7, followed by RNA-seq and alternative splicing quantification (MAJIQ/VOILA). It then fits Hill equations for regulated events (with ΔPSI-based significance and pseudo R² filters), finding that a large majority of regulated cassette exons, IR, and ALE are classified as cooperative (nH≥2).

2) rG4 enrichment links structure to binding sites and to steepness

iCLIP2 in MCF7 maps high-confidence HNRNPH binding sites; sequence context shows G-rich patterns (G enrichment within binding sites and adjacent G pileups). When predicted rG4s (pqsfinder) are overlapped with binding centers, predicted rG4s are enriched at cooperatively repressed exons, and the number of predicted rG4s correlates with steeper dose-response cooperativity.

3) RTstop profiling validates rG4 formation experimentally—and quantifies prediction error

The adapted RTstop profiling uses a 200-nt IVT oligo library spanning predicted binding-site regions, with conditions designed to favor rG4 folding (KCl) vs disfavor (NaCl) and a control to prevent rG4 formation (7-deaza-GTP). The paper finds experimentally detectable rG4s in many predicted BS+rG4 regions, but only ~25% of predicted BS-only regions show measurable rG4 formation, and it quantifies false negatives/false positives for the prediction tool. This matters for skepticism: mechanistic conclusions rely on imperfect computational rG4 calling. Note: the paper’s mechanistic chain is strongest where experimental rG4 formation and functional readouts co-vary, rather than where it depends purely on prediction overlap.

4) Biophysical mechanism: rG4 unfolding enables cooperative multivalent binding

The paper uses FRET to show HNRNPH1 addition increases FAM emission when rG4 is KCl-stabilized, consistent with unfolding. Then, in vitro binding titrations show stronger and cooperative HNRNPH1 binding for rG4-forming RNAs, with cooperativity abolished when rG4 formation is prevented using 7-deaza-GTP during IVT. A quantitative thermodynamic model considers rG4 folding energy, binding free energy, and an entropic backfolding penalty for multi-qRRM binding to exposed G-runs.

5) Translation to splicing: cascade amplification explains higher Hill at output

The in vivo splicing Hill coefficients can be much larger than in vitro binding nH, so the paper uses a cascade kinetic model: step 1 (cooperative HNRNPH binding) → step 2 (modulation of spliceosome assembly with multiple sigmoidal levels) → step 3 (co-transcriptional splice decision). In such architectures, modest cooperativity at each step can multiply to yield highly switch-like output.

Skeptical note: cascade models can fit wide behaviors; the key test is whether independent perturbations change the intermediate steps as predicted (the paper provides strong binding and minigene evidence, but fewer direct intermediate-step mechanistic measurements).

6) Clinical relevance: variants + subtype-specific splicing programs

The paper screens PMBB breast-cancer cohorts for rG4-disrupting variants using G4mer and logistic regression adjusted for ancestry PCs. It reports 8 significant variants associated with breast cancer, and then uses TCGA BRCA data to correlate CE PSI with HNRNPH1/2 expression, with cooperatively enhanced/repressed exons showing directional correlation with HNRNPH2 rather than HNRNPH1. It then clusters samples based on inclusion frequencies of cooperatively regulated CE, separating normal from tumor and stratifying basal subtype severity.

Key skeptical constraint: association in patient cohorts cannot by itself distinguish whether rG4 disruption is causal vs linked to other variants/regulatory contexts. Orthogonal validation could include locus-specific perturbations of predicted rG4 formation in relevant patient-derived models.

Critical appraisal (limitations, blind spots, and what would change the conclusion)

1) rG4 identification is imperfect (prediction error)

The paper explicitly quantifies false negatives and false positives in pqsfinder predictions relative to RTstop-derived rG4 assignments. Therefore, any transcriptome overlap logic that uses predictions alone will be diluted, and mechanistic conclusions should be weighted toward experiments where rG4 formation is manipulated/validated.

2) MCF7-centric biology and synthetic IVTs may not fully recapitulate chromatin/temporal folding dynamics

Binding and splicing cooperativity are demonstrated in MCF7 cells and in vitro-transcribed RNA contexts. While this provides strong mechanistic support, tissue-specific RBPs, helicases, ionic milieu, and co-transcriptional folding could alter rG4 lifetimes and access. The study’s broader relevance would be strengthened by validation in additional cell types and by showing that perturbations of rG4 formation at endogenous loci reproduce the same Hill/PSI shifts.

3) Cooperativity ≠ uniqueness of mechanism

High Hill coefficients can arise from multi-step kinetic amplification generally, not only from rG4 unfolding. The paper provides rG4-dependent cooperativity evidence (GTP vs 7dGTP in vitro; rG4 overlap necessity in AKR1A1 minigene), but other structure classes or RNA-modification-dependent accessibility effects could also contribute at additional loci. Discriminating these would require locus-by-locus structural perturbations with orthogonal probing.

4) Patient associations are not direct causal tests

Variant enrichment and correlation with HNRNPH2 expression suggest a clinical link. However, the study does not (within the provided text) directly measure rG4 topology in patient tumors, nor does it perturb rG4 formation and show the predicted splicing reversal in patient-relevant models. Thus, the clinical mechanism is plausible but still needs orthogonal causal validation.

Most useful takeaways for a mechanistic systems view

Structure–function coupling can create ultrasensitivity: the paper’s binding cooperativity (Hill≈2) can be amplified by a multi-step splicing cascade to produce larger output Hill coefficients in vivo.
rG4-dependent cooperativity is experimentally separable: using 7-deaza-GTP to prevent rG4 formation switches binding cooperativity on/off in vitro, providing a strong causal handle.
Prediction alone is insufficient: rG4 motif prediction has substantial false negatives/positives, so mechanistic conclusions should emphasize experimentally validated rG4 formation or perturbation outcomes.

Further BGPT actions (bespoke exploration)

Run AI Scientist analysis

Author reviews to explore

Feedback:

Updated: April 18, 2026