Review Claim by Claim

• Figure A — Model vs baseline performance (CNN vs Ridge) on cross–cell-type endogenous expression.
• Figure B — Predicted vs observed fold-change ranking for dCas9-p300 perturbations across genes (rank Spearman ≈0.8 reported).
• Interpretation panels and targeted recommendations follow the figures.

Concise evidence-driven critique (evidence inline)

• Training data and baseline performance: The authors train CNNs on ENCODE histone PTM ChIP-seq (-log10 p-value tracks) and RNA-seq across 13 human cell types and show CNNs outperform ridge regression on held-out chromosomes and held-out cell types (median cross-cell-type Spearman ρ ≈0.53 vs 0.39) — this supports that non-linear spatial patterns around TSS contribute predictive signal Batra et al., model training and cross-cell validation
• Normalization & data hygiene: They adapted S3norm to correct batch-like differences across ChIP-seq tracks — appropriate and necessary for comparability; method grounded in Xiang et al. (S3norm) Xiang et al., S3norm method
• Perturbation model assumptions — strength and Achilles’ heel: The in silico perturbation models dCas9-p300 as (i) steric hindrance at guide site via MNase occupancy, (ii) local Gaussian H3K27ac deposition (σ kernel), and (iii) acetylation only at nucleosome-occupied positions (m_i × Gaussian × exp(-5 m_j) × λ). This is a reasonable, testable simplification consistent with prior work showing dCas9-p300 primarily increases H3K27ac nearby, but it ignores (a) p300 promiscuity and non-histone acetylation, (b) cross-talk with other PTMs (H3K4me3 etc.), and (c) potential recruitment of transcription factors or chromatin remodelers by dCas9 complexes — all of which the authors acknowledge as possible causes of lower within-gene ranking accuracy Batra et al., perturbation model and limitations
• Validation & generalization: The model ranks cross-gene fold-changes in HEK293T dCas9-p300 experiments with Spearman ≈0.8 (strong); but within-gene gRNA ranking is weaker and the Perturb-seq K562 validation shows moderate correlation (~0.47), indicating generalization depends on gene- and locus-level context and that gRNA efficacy variance and unmodeled biology remain important Batra et al., validation statistics
• Reproducibility and openness: The paper provides code and data (GitHub repository, SRA/GEO accessions for MNase-seq and Perturb-seq) which greatly helps reproducibility; computational methods (architectures, hyperparameters, ensemble of 100 replicates) and experimental protocols (CUT&RUN, MNase-seq, qPCR) are described — reproducibility rating is high, though full re-run requires ENCODE data download and system resources Batra et al., data & code availability

Where the claims are strongest

1. The models reliably learn established spatial relationships between marks and expression (e.g., promoter H3K4me3/H3K27ac positive, H3K27me3/H3K9me3 negative) — consistent with prior literature on histone marks predicting expression Batra et al., learned histone-feature patterns
2. Cross-gene ranking of dCas9-p300 efficacy is well predicted — useful for selecting which genes are likely to be responsive to acetyltransferase-based activation.

Key limitations, blindspots & risks

• Mechanistic simplification: modeling dCas9-p300 as only boosting H3K27ac at nucleosome-occupied positions omits non-histone acetylation, TF recruitment, and downstream chromatin remodeling that can vary across neighboring guides (cites p300 promiscuity) Weinert et al., p300 acetylome breadth
• gRNA efficacy variability: observed large differences between neighboring gRNAs (even within 50 bp) indicate that sequence- and locus-specific factors (PAM context, nucleosome breathing, local TF occupancy) strongly modulate efficacy and are not fully captured by current scoring metrics — authors note this and call for screens to map local efficacy Horlbeck et al., nucleosome effect on Cas9
• Training–perturbation mismatch: Endogenous PTM distributions used for training may not span extreme acetylation introduced by a strong dCas9-p300 perturbation (extrapolation risk for neural nets) — neural networks can fail to extrapolate beyond training ranges Xu et al., NN extrapolation limits

Actionable suggestions (experiments & modeling) to improve predictive power

1. Generate a systematic paired dataset: For several genes, tile gRNAs in dense windows (e.g., every 5–10 bp across ±500 bp), measure both local CUT&RUN or ChIP-seq for H3K27ac/H3K4me3 after each guide and RNA output (qPCR or Perturb-seq). This would create direct training data mapping guide → PTM change → expression change and break the current two-step assumption.
   Why: removes reliance on modeled perturbations and will let ML learn true mapping from guide to PTM perturbation patterns.
2. Measure non-histone acetylation/trans effects after dCas9-p300 (proteomics time-course) to quantify p300 promiscuity and potential trans-acting changes (Weinert et al. style) Weinert et al., CBP/p300 acetylome
3. Refine gRNA efficacy models for epigenome editors: combine sequence-based features, MNase occupancy, local TF ChIP signal, DNA methylation, and experimentally-measured on-target CUT&RUN H3K27ac to create an epigenome-specific guide scoring function.
4. Train models to predict PTM change (guide → delta PTM tracks) first (supervised using CUT&RUN), then map PTM-change → expression with a separate expression model; compare to current direct perturbation-through-simulation pipeline.

Short checklist for readers who want to re-run or extend the study

• Download ENCODE ChIP-seq & RNA-seq tracks (13 cell types listed in paper) and Avocado imputations used for missing marks; apply S3norm parameters as described (IMR-90 as reference) Batra et al., data & normalization instructions
• Obtain MNase-seq (PRJNA892960) and Perturb-seq (GSE255610) for exact replication.
• Start with the 10 kb TSS-centered context and CNN architecture described (5 conv blocks, 32 kernels width 5, pooling, 16-unit dense) and ensemble across replicates for stability.

Conclusions — balanced and evidence-weighted

The paper convincingly shows that histone PTM patterns across cell types are predictive of endogenous gene expression and that a model of local H3K27ac deposition plus an expression predictor can rank expected dCas9-p300 fold-changes across genes with strong agreement (Spearman ≈0.8). However, predicting within-gene gRNA-level efficacy remains challenging because the mechanistic link between a single-guide–induced histone change and transcription likely involves additional variables (non-histone acetylation, TF recruitment, nucleosome breathing, off-target acetylation) not fully captured in the perturbation model. Overall: high-quality, transparent, and useful advance; next steps require richer paired perturbation PTM datasets to close mechanistic gaps.