Graham McVicker

This is the preprint write up of my sabbatical work with Dave Kelley’s group at Calico. We tried out several transformer replacements for multi-task learning in functional genomics (i.e. what Borzoi does). Mamba, in particular, seems to outperform a mini version of Borzoi, especially when “striped”.

Comprehensive dissection of cis-regulatory elements in a 2.8 Mb topologically associated domain in six human cancers - Nature Communications The oncogene MYC plays a key role in cancer initiation and progression. Using thousands of CRISPR perturbations, the authors investigate regulators of MYC in six different cancers. These tumor-specifi...

Comprehensive dissection of cis-regulatory elements in a 2.8 Mb topologically associated domain in six human cancers

www.nature.com/articles/s41...

Contribution of autosomal rare and de novo variants to sex differences in autism Autism is four times more prevalent in males than females. To study whether this reflects a difference in genetic predisposition attributed to autosom…

Our new paper about rare variant contributions to sex differences in autism is out at AJHG, led by Mahmoud Koko with @vw1234.bsky.social and Kyle Satterstrom. Biggest analysis of exome data in autism to date including SPARK and ASC. www.sciencedirect.com/science/arti...

Can DNA sequence models predict mutations affecting human traits?

We introduce TraitGym, a curated benchmark of causal regulatory variants for 113 Mendelian & 83 complex traits, and evaluate functional genomics and DNA language models. Joint work w/ Gökcen Eraslan and @yun-s-song.bsky.social 🧵👇

On-target edit events detected for different cell types (K562, primary T cells, GM12878, Jurkats) and guides using TIDE. Most tested guides resulted in efficient insertions of the 34bp donor sequence containing the T7 promoter.

Superb-seq is also EASY TO PERFORM. Protocol uses standard kits, equipment, and NO virus. We also provide free software, SHERIFF, to analyze Superb-seq data. We hope this will enable wide-spread adoption and use in diverse cell types!

Superb-seq could be used to readily distinguish benign from potentially risky Cas9 off-target edit profiles for therapeutic applications. Perturb-seq and Guide-seq lack this capability, since they do not jointly detect edit sites and transcriptomes at single cell resolution.

Guide presence is an unreliable indicator of on-target guide editing. Other methods that use guide-only read outs may be confounded by off-target edits that are sometimes even more frequent than the on-target edit.

Edit events colored by the corresponding guide. Only one guide had no detected off target events. Y-axis shows homology between genome and guide spacer. X/I/O indicates whether edit is exonic, intronic, or intergenic. Most off-target edits are intronic.

All off-targets were non-coding, so paired scRNA-seq was crucial to determine their functional effects using differential expression, as shown for the USP9X intronic edit above.

Venn diagram showing overlap between off-target edit sites identified with Superb-seq and those predicted by in silico tools Cas-OFFinder, COSMID and E-CRISP. The overlap is poor and 11 off-targets were not predicted by any computational tool.

Many of the off-target edit sites (total = 36) were not predicted by popular in silico tools. These tools appear to underestimate the tolerance of guide bulges and mismatches with off-target edit sites, yet also predict many off-target edits that do not occur.

Pairwise alignments of SMARCA4 g22 sequence and the genome site at the edit. All off-target sites have PAMS, but have various mismatches to the guide spacer sequence.

Off-target sites had clear guide homology and PAMs. There was however a surprising tolerance of guide bulges and mismatches with off-targets, including mismatches in the “seed” region.

Detected edits for SMARCA4 guide 22. 13 edits were detected. Y axis is the frequency of the edit (number of cells) and X axis is the homology between the genome and the guide spacer sequence. Some off target events are more frequent than the intended on-target.

One guide had an off-target edit that was 34x more frequent than the on-target edit! A total of 5 off-target sites were observed in more cells than the on-target for this particular guide.

Sheriff counts edit alleles per cell at each edit site. An off-target edit within the first intron of USP9X is associated with differential expression of USP9X and >100 downstream genes.

We quantified Cas9 edits per cell (‘edited alleles’), and associated edit alleles with gene expression. One intronic off-target edit perturbed the expression of USP9X and >100 downstream genes!

Genome positions and frequencies (number of cells) of on and off-target events detected with Superb-seq.

SURPRISING RESULT. We applied Superb-seq to 10k K562 cells and detected pervasive off-target Cas9 edits, with an average of 6 off-target sites per guide, ranging in frequency from 0.03-18.6% of cells!

A schematic of Superb-seq showing 3 steps: (1) edit labeling with donor sequence containing T7 promoter; (2) In situ transcription with T7 polymerase; (3) combinatorial single-cell RNA-seq + computational analysis with Sheriff.

Superb-seq detects edits and cell RNA by labelling nuclease cleavage sites with homology-free insertion of a T7 promoter, followed by “Zombie” in situ transcription with T7-pol to generate edit-site marking barcoded RNA. T7 and cell RNAs are jointly read out with scRNA-seq.

Unlike single-cell Perturb/CROP-seq methods that read guide RNAs as a proxy of edits, Superb-seq captures edit events directly. This enables functional assessment of edits including off-target events that confound analyses and raise gene therapy risk.

Joint single-cell profiling of CRISPR-Cas9 edits and transcriptomes reveals widespread off-target events and their effects on gene expression A longstanding barrier in genome engineering with CRISPR-Cas9 has been the inability to measure Cas9 edit outcomes and their functional effects at single-cell resolution. Here we present Superb-seq , ...

Excited to announce our preprint describing SUPERB-SEQ 🦸, a new method to measure Cas9 edits and their effects on gene expression in single cells. Led by @micklorenzini.bsky.social and @bradbalderson.bsky.social www.biorxiv.org/content/10.1...

The distribution of highly deleterious variants across human ancestry groups A major focus of human genetics is to map severe disease mutations. Increasingly that goal is understood as requiring huge numbers of people to be sequenced from every broadly-defined genetic ancestry...

“The distribution of highly deleterious variants across human ancestry groups”. Preprint with Anastasia Stolyarova and @gcbias.bsky.social: www.biorxiv.org/content/10.1...

Efficient count-based models improve power and robustness for large-scale single-cell eQTL mapping Population-scale single-cell transcriptomic technologies (scRNA-seq) enable characterizing variant effects on gene regulation at the cellular level (e.g., single-cell eQTLs; sc-eQTLs). However, existi...

Excited to present our work on developing jaxQTL, a fast single-cell eQTL mapping tool that improves power and robustness in identifying sc-eQTLs using count-based models. See details in threads 🧵
www.medrxiv.org/content/10.1...

Leveraging functional annotations to map rare variants associated with Alzheimer's disease with gruyere The increasing availability of whole-genome sequencing (WGS) has begun to elucidate the contribution of rare variants (RVs), both coding and non-coding, to complex disease. Multiple RV association tes...

Excited to share our first foray into (noncoding) rare variant association testing: a probabilistic model that learns functional annotation importance and finds associations missed by existing methods. Anjali did a fantastic job with model assessment and scaling! www.medrxiv.org/content/10.1...