(iii) A lentiMPRA experiment measuring the effects of the same PPIF promoter variants assayed with Variant-EFFECTS. This revealed systematic differences between effects measured in an endogenous versus reporter context, highlighting the strengths of Variant-EFFECTS.

(ii) A mutagenesis screen targeting the IL2RA promoter & measuring effects using an alternative assay, which demonstrated how Variant-EFFECTS can be generalized to study the quantitative function of regulatory sequences that control various cellular phenotypes, including RNA & protein expression.

(i) A validation screen of 10 significant variants from the original PPIF enhancer screen, which demonstrated the reproducibility and robustness of Variant-EFFECTS measurements across screens performed with varied parameters (MOI, coverage, guide, representation).

In total, 78% of edits designed to increase or decrease expression in either cell type significantly affected PPIF expression in the intended direction, although to varying degrees.

The range of effect sizes induced by these ML-designed edits was striking. We found that small changes (4-10bp) to the PPIF promoter sequence, delivered via CRISPR prime editing, can turn PPIF gene expression up by more than 2-fold (+140%) or down by more than 7-fold (-86%)!

CRISPR edits to endogenous DNA that tune expression to a desired level (e.g., up 2-fold) could enable new therapeutic applications for many diseases. To test if we could achieve such changes with CRISPR prime editing, we used an ML-guided approach to rationally design edits to the PPIF promoter.

Additionally, none of the models appeared to correctly interpret the effects of edits to the distal enhancer—either due to limited context in local models or failure of long-range models to learn its importance. However, our results suggest an alternative way to capture long-range variant effects.

Through this analysis, we found that chromatin accessibility models outperformed the corresponding expression models for predicting effects on gene expression, similar to previous analysis of data from plasmid reporter assays and natural genetic variation.

To illustrate how gold-standard experimental measurements of variant effects on expression in endogenous context can be leveraged for quantitative evaluation of predictive models, we benchmarked the accuracy of recent deep learning models of gene regulation against our new Variant-EFFECTS dataset.

Unexpectedly, the effects of these TF motif insertions were highly correlated between cell types & states, but differed by a constant scaling factor. This suggests that endogenous/cellular context can affect the contributions of genomic variation & TF binding sites to expression in multiple ways.

To demonstrate how Variant-EFFECTS will enable new studies of how TF motif instances encode cell type-specific gene expression in an endogenous context, we inserted TF motif instances into the PPIF promoter and compared effects across three distinct cell types and states.

We found that mutagenizing even a single TF motif instance often led to large changes in expression. For example, disrupting the central ETS motif in the enhancer accounted for a significant portion of the decrease in PPIF gene expression observed when perturbing the enhancer with CRISPRi.

To demonstrate how Variant-EFFECTS can accelerate study of endogenous regulatory DNA, we performed high-throughput tiled mutagenesis of the PPIF enhancer & promoter in THP-1 cells, providing a comprehensive map of sequence effects on expression & highlighting key features of regulatory architecture.

We performed this comparison for a total of 12 variant-gene-cell-type combinations throughout this work. Mean effects from pooled Variant-EFFECTS screens were consistently concordant with those measured by qPCR in clonal cell lines, confirming the quantitative accuracy & reproducibility of our tech.

First, to demonstrate the Variant-EFFECTS protocol, we introduced edits disrupting a 5’ splice site of the PPIF gene in THP-1 cells. As expected, these edits strongly reduced expression, and their effects were consistent with those measured via qPCR in clonal cell lines carrying the same edits.

To address this, we developed Variant-EFFECTS (Variant Effects From Flow-sorting Experiments with CRISPR Targeting Screens), by combining pooled prime editing with fluorescence-based cell sorting to directly measure effects on target gene expression.

Here, we applied Variant-EFFECTS to map the logic of regulatory elements in the genome, to benchmark the quantitative accuracy of sequence models with our gold-standard dataset, and to combine prime editing with computational design strategies to screen for CRISPR reagents that reprogram expression.