Mapping the Genetic Landscape of the
DNA Damage Response

Research Overview

Pairwise genetic interactions provide insights into the functional relationships between genes by examining their combined loss of function. In this study, we present findings from combinatorial Cas12a screens that assessed the full set of genetic interactions among 233 DNA damage response genes.

Understanding GI z-scores

Genetic interaction scores (GI z-scores) indicate the phenotypic deviations from the expected outcomes when two genes are independently inactivated compared to the phenotype observed when both genes are simultaneously targeted in the same cell.

−

Negative GI z-score Worse than expected phenotype

+

Positive GI z-score Better than expected phenotype

Genetic Interaction Study Network Visualization

Interactive Features

Clustered Heatmaps Visualize genetic interaction modules
Multiple Data Views Two time points with different filtering parameters
Gene Query Search individual genes for comprehensive summary
AlphaFold2 Integration Visualize protein structure prediction scores
Network Analysis Visualization Tool Visualize network graphs of the genetic interaction screen

Gene Knockout Experimental Setup

Heatmap Controls

Heatmaps

RPE-1 / MCF10A

RPE-1

MCF10A

View Type

Export

Gene Search

Dataset

Search

Actions

Heatmap Tip

Drag a rectangle to zoom in.
Double-click to zoom out.
Hover any cell for details.

Gene	Rank	z-score

© 2024 Ciccia Laboratory, Columbia University Medical Center

Experimental Design and Datasets

Experimental Design

Library: 699 constructs targeting 233 single DDR genes, 81,084 constructs targeting 27,028 DDR gene pairs, and controls (non-essential and essential single genes, essential gene pairs, synthetic lethal paralog pairs and corresponding single paralog genes).
In4mer platform [ref ]: each construct expresses a 4-gRNA array (2 gRNAs per gene for gene pairs) to maximize homozygous double-knockout rates. Positional effects are controlled by including 2 additional constructs with reordered gRNAs, for a total of 3 constructs per target gene.
Nuclease: enhanced Cas12a from Acidaminococcus sp. (enCas12a) — originally engineered for higher activity and broader PAM compatibility [ref ], and later optimized for combinatorial genetic screens [ref ].
Cell lines: MCF10A — spontaneously immortalized mammary epithelial cells; RPE-1 — hTERT-immortalized retinal pigment epithelial cells with TP53 deletion (to minimize p53-induced arrest from loss of key DDR genes). Both cell lines are characterized by stable near-diploid karyotype.
Time points: early (T9) and final (MCF10A_T17 and RPE-1_T18).
Replicates and representation: screens were performed in triplicate at 1000× (MCF10A) and 500× (RPE-1) library representation.

Datasets

The screens generated four individual datasets. In addition, integrated datasets were generated by combining two or more individual datasets using Stouffer's method. Data for all individual and integrated datasets are available for download using the Download tab in the top menu bar. In the web application, interactions can be filtered by dataset, GI z-score, FDR, and gene/gene pair using the Data tab in the top menu bar.

Individual datasets (4)

RPE-1_T9 — RPE-1 early time point.
RPE-1_T18 — RPE-1 final time point.
MCF10A_T9 — MCF10A early time point.
MCF10A_T17 — MCF10A final time point.

Integrated datasets (5)

RPE-1_integrated = RPE-1_T9 + RPE-1_T18.
MCF10A_integrated = MCF10A_T9 + MCF10A_T17.
RPE-1/MCF10A_T9_integrated = RPE-1_T9 + MCF10A_T9.
RPE-1/MCF10A_T18/T17_integrated = RPE-1_T18 + MCF10A_T17.
RPE-1/MCF10A_integrated = RPE-1_T9 + RPE-1_T18 + MCF10A_T9 + MCF10A_T17.

Genetic Interactions

Genetic interaction (GI) z-scores measure the strength and direction of interactions between gene pairs. They were computed with the GRAPE pipeline (Genetic interaction Regression Analysis of Pairwise Effects), developed by the Traver Hart Lab. GRAPE (1) estimates single mutant fitness (SMF) for each gene; (2) calculates expected double-knockout (dKO) log fold change (LFC) for each gene pair as a sum of the two genes' SMF values; (3) defines the raw GI score as the deviation between observed and expected dKO values; (4) filters out gene pairs with expected dKO values smaller than the minimum of the observed dKO value; (5) calculates GI z-scores by applying a local variance normalization approach. GI z-scores can be viewed as heatmaps using the Heatmaps & Search tab in the top menu bar.

Single Mutant Fitness (SMF)

SMF is the single KO fitness score for each gene, estimated by the GRAPE regression model from all pairwise measurements in the screen, under the assumption that most gene pairs do not exhibit interactions.

Genetic Interaction z-score Interpretation

The GI z-score quantifies how much the observed dKO phenotype deviates from what would be expected if the two genes acted independently. It is computed in four steps:

Expected dKO fitness (LFC_expected) — calculated from SMF scores for each gene pair, assuming additivity.
LFC_expected = SMF_g1 + SMF_g2
Observed dKO fitness (LFC_observed) — measured from the screen as the log₂ ratio of the dKO gRNA read counts (N) at the final time point vs T0.
LFC_observed = log₂( N_endpoint / N_T0 )
Raw GI score — calculated as the difference between observed and expected LFCs.
raw GI = LFC_observed − LFC_expected
Local variance normalization — the raw GI is divided by a local standard deviation estimated from the 500 nearest gene pairs (ranked by expected dKO LFCs), to account for the higher variance observed at lower expected dKO fitness values.
GI z-score = raw GI / σ_local

Sign convention: negative z-scores correspond to synthetic sick/lethal interactions (dKO fitness is worse than expected), and positive z-scores correspond to positive/buffering interactions (dKO fitness is better than expected).

Statistical significance is assessed by calculating one-sided p-values from the z-distribution (separately for synthetic sick/lethal and positive (suppressor and within-pathway) interactions) and applying the Benjamini–Hochberg correction for multiple testing.

High-confidence genetic interactions are called using an FDR threshold of < 0.1 in each individual dataset (RPE-1_T9, RPE-1_T18, MCF10A_T9, and MCF10A_T17). Across these datasets, >750 high-confidence pairwise interactions were identified.

The complete list of unique high-confidence hits across all datasets, as well as high-confidence hits for individual datasets, can be downloaded here and here , respectively.

Caveats

Limited detection of negative interactions that contain essential/strongly selective genes

Strong negative SMF effects associated with essential or strongly selective genes obscure their pairwise interaction signatures at the final screen time point. To allow potential detection of negative interactions involving these genes, we included an early time point (T9), where SMF effects may be less severe. Indeed, we were able to detect negative genetic interactions involving essential genes at T9. Accordingly, we recommend examining T9 when assessing negative interactions involving essential/strongly selective genes, although some genes are already strongly depleted at T9 and genetic interactions for these genes cannot be accessed.

Artifactual positive interactions between essential/strongly selective genes

Gene pairs involving two essential or strongly selective genes can sometimes appear as artifactual positive interactions. This occurs because expected dKO fitness values for pairs of highly deleterious genes can fall outside the empirically observable range of the CRISPR screen, resulting in positive GI scores. Furthermore, the variance of raw GI scores grows as expected dKO values become more negative. To mitigate this issue, the GRAPE pipeline implements a dynamic range filter that removes gene pairs whose expected dKO values are below the minimum observed dKO fitness effects. These gene pairs are not present in our data tables and do not have GI z-scores in the heatmaps (appear as white squares). Additionally, GRAPE applies empirical local variance normalization, which accounts for the increased variance observed among highly deleterious perturbations, further reducing the number of significant false-positive interactions. However, some artifactual positive interactions remain. Thus, positive interactions between essential/strongly selective genes should be interpreted cautiously. We recommend follow-up studies only if these interactions are supported at an earlier time point or using orthogonal validation strategies (e.g., CRISPRi).

Clustering

To identify functional gene modules, the GI z-score matrix is clustered using a multi-step approach: (1) HDBSCAN clustering with cosine similarity to assess gene-pair profile similarity, repeated across many filtered matrices; (2) conversion to a GI odds ratio that captures the co-clustering frequency of each gene pair; (3) hierarchical clustering of the odds ratio matrix to define final modules.

HDBSCAN & Genetic Interaction Profile Similarity

Each gene has a genetic interaction profile — the vector of its GI z-scores against every other gene in the library. Genes with similar profiles are likely to share function or participate in the same complex.

Profile similarity is measured using cosine similarity, and clustering is performed with HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) [ref ] [ref ] using the parameters: metric = cosine, min_cluster_size = 2, max_cluster_size = 15.

To increase robustness, clustering is repeated across many filtered versions of the input matrix, varying GI strength thresholds (z-score 1.5–4.0) and connectivity thresholds (1–30 minimum significant interactions per gene). For each gene pair, the number of times the two genes are assigned to the same cluster is then counted across all runs, avoiding reliance on any single clustering outcome.

GI Odds Ratio

For each gene pair, the co-clustering count from the HDBSCAN runs is normalized to produce the GI odds ratio:

GI odds ratio_g1,g2 = 2 · C_g1,g2 / ( C_g1 + C_g2 )

where C_g1,g2 is the number of co-clustering events between genes g1 and g2, and C_g1, C_g2 are the total clustering assignments for each gene.

Values range from 0 (never co-clustered) to 1 (always co-clustered). The GI odds ratio matrix can be viewed as a heatmap using the Heatmaps & Search tab in the top menu bar.

Functional Modules Identified

Hierarchical clustering (Ward's linkage, Euclidean distance) is then applied to the GI odds ratio matrix to define the final gene modules. The number of modules is selected by silhouette analysis across k = 2–100 clusters.

Each dataset yielded 46–69 distinct modules. The identified modules broadly align with the major DDR pathways (e.g., non-homologous end joining, Fanconi anemia, translesion synthesis, Holliday junction resolution, mitotic DNA repair).

The full module assignments for every dataset (one sheet per dataset) can be downloaded here .

AlphaFold2-multimer Predictions

To test whether module members form direct physical interactions, candidate gene pairs within each module from the integrated datasets were modeled with AlphaFold2-multimer. This analysis predicted enrichment in direct physical contacts between multiple module members, supporting a link between genetic module composition and direct physical interactions. These predictions can be viewed as heatmaps using the Heatmaps & Search tab in the top menu bar.

Prediction Scores

The quality of the AlphaFold2 models for each gene pair is scored using four parameters:

avg_ipTM — interface predicted TM-score, averaged over 5 models (higher = more confident direct interaction)
max_ipTM — best interface predicted TM-score across the 5 models
max_pTM — best overall predicted TM-score
max_confidence — overall model confidence score

Top hits among the 2,029 modeled gene pairs recover known complexes such as TREX1–TREX2, FAAP24–FANCM, SPIDR–ZSWIM7, MAD2L2–REV1, and CENPS–CENPX.

The full table (sorted by ipTM) can be downloaded here .

Dataset

Timepoint

Z-Score

Z< Z>

FDR

neg< pos<

Gene Pair

Gene