Review Papers Against Their Own Results

Concise critique and synthesis (evidence-based)

1. Claim 1 — Pangenomes reduce reference bias and reveal missing sequence. Evidence: T2T-CHM13 and HPRC quantify large added sequence (≈200 Mbp and ≈119 Mbp respectively), and multiple population-specific pangenomes (Tibetan, JaSaPaGe, KOREF) add hundreds of Mbp of non-reference sequence, enabling discovery of SNVs and SVs previously inaccessible to GRCh38 mapping T2T/HPRC evidence JaSaPaGe improvement
2. Claim 2 — Dramatic improvements in SV discovery and genotyping. Multiple long-read + pangenome studies (cattle, Tibetan, French cattle pangenome) report orders-of-magnitude higher SV discovery and show breed/population-specific SVs that associate with phenotypes (e.g., MATN3 deletion associated with stature in French cattle) Cattle SV evidence French cattle SV GWAS
3. Claim 3 — Clinical utility is promising but operationally hard. Graph/pangenome approaches improved somatic variant detection in controlled benchmarks (HapMap mixtures, COLO829) and reduce mapping-induced miscalls, but adoption faces obstacles: standard formats, annotation transfer across graphs, need for graph-aware clinical callers, compute/memory costs, and regulatory/validation pathways HapMap somatic benchmark Pangenome somatic detection gains
4. Practical limitations and blindspots the review should emphasize more:
   • Tool fragmentation and lack of standard graph formats and clinical-grade pipelines — multiple graph builders (Minigraph, PGGB, Minigraph-Cactus) produce different topologies with differing downstream behaviours Great Genotyper scaling & limits
   • Annotation transfer and gene models in graphs remains immature — tools like GrAnnoT help but interchromosomal/non-syntenic events and TE-rich regions are still challenging GrAnnoT annotation transfer
   • Validation & functional follow-up of SV-trait links is often incomplete; statistical association alone (imputed SV-GWAS) can be confounded without orthogonal wet-lab validation (PCR, expression, CRISPR) — many pangenome studies note this as a major next step SV interpretation challenges
5. Methodological and epistemic cautions: beware of publication and sampling biases (much pangenome work is population- or species-focused), differences between assembly/graph-building pipelines, and overinterpreting associations without functional validation; the field also faces trade-offs between graph completeness and computational tractability (PGGB vs Minigraph-style trade-offs) Population genotyping trade-offs Pangenome pipeline benchmarking

Concrete recommendations (for readers, implementers, clinicians)

• Adopt hybrid strategy: keep a stable linear backbone for clinical pipelines but augment with pangenome DAGs or donor-specific assemblies for difficult loci (LPA, CYP2D6, immunogenes) and somatic analyses — this balances stability and sensitivity (evidence: SMaHT HapMap+COLO829 benchmarks) SMaHT recommendation
• Standardize benchmarks and metrics (precision/recall in difficult regions, SV validation rates, per-gene coverage models) using community resources (SMaHT HapMap, Graph-based HapMap truth sets) to permit regulatory-grade evaluation HapMap benchmark resource
• Invest in annotation transfer and graph-native gene models (GrAnnoT, PanSel) and in training clinical pipelines on graph-aware surjection or graph-native variant callers to avoid losing information at the surjection step GrAnnoT approach PanSel conservation scoring

Confidence, falsifiability, and missing evidence

Confidence: moderate–high for the claim that pangenomes and long reads produce more complete variant catalogs (multiple independent studies show large increases in added sequence and SV counts), but lower for near-term clinical impact until graph-aware clinical callers, standards, and regulatory validation are available. What would change the conclusion: if large-scale, independent clinical benchmarks showed no improvement in diagnostic yield or produced unacceptable false-positive rates when pangenome approaches are applied; or if efficient, standardized graph-aware pipelines and annotation transfer prove impracticable in routine clinical labs.