Biology Atlas

Map / Assays

assay · deep · partly written

Genetic interaction and epistasis screens

A genetic interaction screen breaks two genes at once and asks whether the double mutant behaves differently from what the two single mutants predict, which reveals functional relationships without touching any protein.

Also called: epistasis screen, synthetic genetic array, SGA, double-mutant screen, genetic interaction map

The short version

A genetic interaction is a surprise: you break gene A and gene B together and the double mutant grows differently from what the two single mutants predict. The gap is the epsilon score. Genes with similar rows of scores work in the same pathway or complex, but this measures functional relatedness, never contact: only 30 of 4,039 genetically interacting yeast pairs encoded physically interacting proteins. 12

Reported, not yet replicated

Gear 1: the double mutant more

Yeast synthetic genetic array crosses a query mutant into an array of about 4,700 deletion strains and reads colony size as fitness. In human cells, dual-guide CRISPRi knocks down two genes at once in a pooled culture; Horlbeck built 1,044,484 guide pairs targeting 222,784 gene pairs (472 x 472 genes) in K562 cells, repeated in Jurkat. 13

Reported, not yet replicated

Gear 2: the expected phenotype and epsilon more

Interaction means deviation from a null model: in yeast a digenic interaction is a significant deviation from the expected multiplicative effect of combining two single mutants; negative means sicker than expected (the extreme is synthetic lethal), positive means healthier. Illustration (our arithmetic, not a data point): single fitnesses 0.8 and 0.5 predict a double of 0.40; an observed 0.20 gives epsilon about -0.20. The choice of null model has 'profound consequences' (Mani 2008, abstract), and the human map used a different one. Thresholds are set against controls: a 5% FDR from gene-versus-negative-control pairs. 243

Reported, not yet replicated

Gears 3 and 4: from profile similarity to pathways more

Each gene's row of epsilon scores is its fingerprint; correlated rows suggest the same pathway or complex even when the proteins never touch. Parallel or buffering pathways link by negative interactions; the 2010 map found interactions between different pathways and complexes were often 'monochromatic', almost all negative or all positive. Classical epistasis ordering compares a double mutant with each single mutant under a signal and assumes a signal, a strict upstream and downstream hierarchy and null alleles; quantitative screens drop the signal and find parallel routes, not order. 1256

Reported, not yet replicated

Where it breaks more

Redundant paralogues hide pairwise interactions (in the 2004 map, CLN1 and CLN2 showed 34 interactions that were all triple-mutant effects). The null model changes the call. Very sick pairs cannot be scored accurately. Reproducibility varies: replicate correlation of guide profiles was R = 0.75 in K562 but 0.44 in Jurkat, and the MICOS-ATP synthase buffering seen in Jurkat was not seen in K562. Only about 20% of attempted yeast queries gave usable data in 2004. 13

Reported, not yet replicated

Question answered

Do these two genes work in the same pathway or complex, or in parallel routes to the same essential job? It measures functional relatedness from the fitness of double mutants; it does not say the proteins touch. 12

Reported, not yet replicated

Readout physical

Fitness of double mutants against the product of the single-mutant fitnesses. In yeast, colony size on an array of deletion strains after robotic mating and meiosis (synthetic genetic array); in human cells, pooled CRISPRi with dual guide RNAs, reading guide-pair abundance at the end of the screen against the start. 13

Reported, not yet replicated

Not measured

Contact, direction, mechanism, or whether the relation holds in another cell type. Genes in one complex score alike, but in yeast only 30 of 4,039 genetically interacting pairs encoded proteins that physically interact. The phenotype is usually one readout (growth). 13

Reported, not yet replicated

Setting

Where in cascade

Sample prep

Yeast: a query mutant is crossed into an array of about 4,700 deletion strains. Human: a pooled library of dual-guide constructs targeting gene pairs, introduced into cells expressing a catalytically dead Cas9 fused to a repressor (CRISPRi). 13

Reported, not yet replicated

Conditions

Calling a pair an interaction needs both size and significance: the 2010 yeast map used |epsilon| > 0.08 and P < 0.05. The human screen ran for about 10 population doublings and called 'strong' interactions at +/-3, which covered 2.2% of measured pairs. The null model is a choice: Mani and colleagues compared four (product, additive, log, min) and found 52% of known synergistic interactions were inferred under min, though product and log were better for finding functional relationships. 234

Reported, not yet replicated

Throughput

Genome-scale in yeast: about 23 million double mutants, with about 550,000 negative and 350,000 positive interactions and about 90% of yeast genes covered (Costanzo 2016, supplement page). Human: 1,044,484 guide pairs targeting 222,784 gene pairs (472 x 472 genes) in K562 cells, repeated in Jurkat. 53

Reported, not yet replicated

Engineerable variants

Synthetic genetic array (yeast)

A query mutant is crossed into an array of about 4,700 deletion strains by robotic mating and meiosis, and colony size is read as fitness. 1

Reported, not yet replicated

Dual-guide CRISPRi (human cells)

A dead Cas9 repressor knocks down two genes at once without cutting DNA; guide-pair abundance is the fitness readout. 3

Reported, not yet replicated

Profile similarity

Each gene's row of epsilon scores against every partner is its fingerprint; Pearson correlation of rows groups genes into pathways and complexes. Costanzo 2016 used cut-offs of 0.05 (compartment), 0.2 (bioprocess) and 0.4 (pathway or complex). 5

Reported, not yet replicated

Different null models

Product, additive, log and min definitions change which pairs count as interacting; the human map used a quadratic fit of single against pair phenotypes. 43

Reported, not yet replicated

Use cases

Assigning function to an uncharacterised yeast gene (YMR299c)

In the 2004 yeast map the uncharacterised YMR299c clustered with the dynein-dynactin spindle pathway, and its deletion showed matching phenotypes. 1

Reported, not yet replicated

TMEM261 in the human map

TMEM261 correlated with Complex I genes in the human map, and its repression decreased respiration-linked readouts. 3

Reported, not yet replicated

Genetic and physical networks compared (yeast)

Only 30 of 4,039 genetically interacting yeast pairs encoded proteins that interact physically; pairs sharing more than 16 genetic neighbours interacted physically in 28 of 333 cases (an 11-fold enrichment); the genetic network was at least four times denser than the physical network. In the human map, 315 of 390 top-correlated pairs (r > 0.6) had no STRING physical annotation, while 79.3% of the highest-confidence STRING interactions were recovered at correlation above 0.1. 13

Reported, not yet replicated

Data availability

Yeast genetic interaction map (thecellmap.org and the authors' supplement) public

The Costanzo 2016 yeast matrices are public through the authors' supplement page and thecellmap.org; a Dryad record exists under doi 10.5061/dryad.4291s. 5

Caveats. The licence of the Dryad copy was not confirmed in a cited source, and the supplement page does not mention that DOI.

Reported, not yet replicated

Human CRISPRi genetic interaction map (Horlbeck 2018) public

The paper cites its data as deposited on Mendeley Data (doi 10.17632/rdzk59n6j4.1). 3

Caveats. The licence was not checked, and the map covers only 472 x 472 genes enriched for essential genes, so 2.2% strong is not a genome-wide rate.

Reported, not yet replicated

No price or service found cro

No core-facility or CRO price for a genetic interaction screen was found in any opened source. 3

Caveats. Absence from what was searched, not proof that none exists.

Unverified: from our own report, source not opened

Teach this

In one sentence

A genetic interaction screen breaks two genes at once and scores how far the double mutant departs from the expectation set by the single mutants, which maps function without touching a protein. 12

Reported, not yet replicated

The five-minute version

Knock out gene A, gene B, then both; if the double mutant grows worse or better than the product of the singles, the pair interacts genetically (epsilon). Do this at scale: about 23 million yeast double mutants, or about 223,000 human gene pairs by CRISPRi. Each gene's row of scores is a fingerprint, so genes in the same complex or pathway cluster. It is a sensitive but non-specific shadow of physical interaction (30 of 4,039 genetically interacting yeast pairs encode proteins that bind), so use it as an orthogonal check on interaction edges, not as an interaction assay. 153

Reported, not yet replicated

What a listener will ask

Does a genetic interaction mean the proteins bind?

No: in yeast only 30 of 4,039 genetically interacting pairs encoded physically interacting proteins. 125346

Reported, not yet replicated

Can it tell me which gene is upstream?

Not in the quantitative-screen sense: it finds parallel routes and clusters, and classical epistasis ordering needs a signal and a strict hierarchy.

Do two labs get the same map?

Not necessarily: the null model changes the call and replicate correlation varied from 0.75 to 0.44 across two human cell lines. 125346

Reported, not yet replicated

The common wrong idea

Genes with a genetic interaction are in physical contact.

Genetic interaction measures functional relatedness; only 30 of 4,039 interacting yeast pairs encoded physically interacting proteins. 125346

Reported, not yet replicated

The one citation to hand over: 1

Still open here

Not yet written: Time per run, Cost, Before after, Capital, ML role.

Linked from

Sources

  1. Tong, A. H. Y., et al.. Global mapping of the yeast genetic interaction network. Science (2004). paper; opened 2026-11-17 all records citing this
  2. Costanzo M, Baryshnikova A, Bellay J, et al.. The genetic landscape of a cell. Science (2010). paper; opened 2026-11-17 all records citing this
  3. Horlbeck MA, Xu A, Wang M, et al.. Mapping the Genetic Landscape of Human Cells. Cell (2018). paper; opened 2026-11-17 all records citing this
  4. Mani R, St Onge RP, Hartman JL, Giaever G, Roth FP. Defining genetic interaction. Proceedings of the National Academy of Sciences (2008). paper; opened 2026-11-17 all records citing this
  5. Costanzo M, VanderSluis B, Koch EN, et al.. A global genetic interaction network maps a wiring diagram of cellular function: supplementary data page. Boone Lab, University of Toronto (2016). organisation-page, not peer reviewed; opened 2026-11-17 all records citing this
  6. Azpeitia E, Benitez M, Padilla-Longoria P, Espinosa-Soto C, Alvarez-Buylla ER. Dynamic network-based epistasis analysis: boolean examples. Frontiers in Plant Science (2011). paper; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (1)

This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Throughput class
ultra-high
Emerging
no
Status
partly written
Last updated
2026-11-17
Record
content/assays/genetic-interaction-screens.json