Review Papers Claim by Claim

 Long Explanation Copied

MethylAI: what the paper claims and how well it supports them

This preprint introduces MethylAI, a deep learning model that predicts single-CpG methylation from 18,432-bp sequence after cross-species pretraining on 1,900 WGBS methylomes (12 species) and human fine-tuning on 1,574 methylomes spanning 52 tissues MethylAI preprint.

Headline performance: mean PCC 0.903 (range 0.741–0.962) on an independent test chromosome (chr11), and 0.875 vs 0.813 (INTERACT) and 0.828 (HyenaDNA) across 1,574 test samples under matched model sizes Benchmark metrics.

Attribution findings: 19.5% of examined TF motifs were more informative than the CpG dinucleotide itself; gene-body CGI motifs (IKZF1, CTCF, BATF::JUN, GATA2) showed 9.6–16.8× enrichment conserved across ten tissues Motif attribution results.

Validation is the strongest feature: CTCF knockout in HEK-293T gave WGBS shifts concordant with in silico perturbation (PCC 0.859–0.861), and CRISPR deletion of a TRMU CTCF motif raised local methylation; mQTL direction accuracy was 87–90% in active motifs vs ~53–58% elsewhere Experimental and mQTL validation.

Critical caveats

• Wet-lab validation is narrow: only CTCF in one cell line (HEK-293T); claims about REST, GATA, AP-1, and ZNF motifs remain correlational/attribution-based.
• Interpretability caveat: as the sister Melody paper explicitly notes, motif-attribution analyses establish model-derived associations, not definitive causal mechanisms Melody interpretability limitation — the same caveat applies to MethylAI's MAS/attribution framework.
• Independent competition is thin: only two baselines (INTERACT, HyenaDNA); the concurrent Melody framework reports different metrics (Spearman 0.645–0.723 across 39 tissues) making head-to-head comparison impossible without re-benchmarking Melody benchmark.
• Data lineage: heavy reliance on Loyfer et al. 2023 atlas and ENCODE; tissue representation is uneven (blood dominates); chromosome-level splits mitigate leakage but the same CpGs appear across tasks.
• GWAS claims are enrichment-based (FDR < 0.1) on 1,689 active-motif SNPs; the ZNF convergence ("nearly half") is suggestive but does not establish mechanism.

Bottom line: a technically impressive, well-documented sequence-to-methylome model with genuine experimental validation; its mechanistic “methylation grammar” narrative outruns what one cell line and attribution maps can prove. Reproducibility is aided by public code (github.com/Yu-Lab-Genomics/MethylAI) Code availability.

Author reviews: