Systematic review and meta-analysis of genome-wide pooled CRISPR screens to identify host factors involved in influenza A virus infection

Systematic review of the literature

A systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to identify pooled genome-wide CRISPR screens to identify host factors involved in influenza virus infection (Fig. S1). The literature search yielded 304 studies, from which 290 were excluded because they were duplicates or found to be irrelevant based on their title and abstract. The entire articles of the remaining 14 studies were retrieved and further evaluated for eligibility. Two of those studies did not meet the inclusion criteria because of experimental settings or the unavailability of data in the form of raw sequencing reads or a gene list (22, 23). Finally, a total of 12 studies that met our eligibility criteria were identified, which allowed the extraction of a total of 19 gene lists, of which 15 were proviral and 4 antiviral. The three screens reported by Song et al. (24) were considered to be unranked, as the reported gene lists did not originate from a gene-ranking analysis. The gene list of the screen performed by Li et al. is based on a genome-wide screen followed by a targeted screen (18). MAIC input parameters and experimental conditions for each screen can be found in Table 1. A more detailed overview of the methodology and study design of the included screens can be found in Table S1.

We performed an analysis of the overlapping hits as determined by the gene-ranking algorithm used in the original research paper. The resulting upset plots display the intersection between indicated data sets (at the bottom) in bars (at the top) (Fig. 1A and B; Table S2). Of note, although Zhou et al. (28) performed two screens, they only provided one list of top hits based on both screens. Overlap between screens from the same authors was generally larger than the overlap between screens from different authors. Genes shared by a high number of CRISPR screens mainly pertained to well-known pathways such as the sialylation and vacuolar acidification pathways. However, since authors use diverse algorithms and thresholds to determine hits (Table S1) and a mere overlap does not consider gene ranking, meta-analysis such as MAIC may be better suited to integrate and fully interpret the valuable data from individual screens.

MAIC analysis of included screens

MAIC of the 15 proviral gene lists (n = 12,613 unique genes) identified 5,561 proviral genes that were present in the data sets of two or more screens, whereas one gene (SLC35A1) was present in the data sets from nine screens, one gene (TP53) from eight screens, and two genes (SIRT2 and MMS22L) from seven screens. MAIC on the antiviral data sets (n = 4,856 unique genes) revealed that 448 genes were present in the data sets of two or more screens and nine genes of three screens (ALYREF, FOCAD, HIST1H1C, TAF1L, KRTAP6-2, KCNG1, LARP4B, TSGA10IP, and ANXA8). The number of genes present in data sets from more than two, three, four, and so on screens is presented in Fig. 2A and B, while the corresponding gene names are listed in Table S3. The complete proviral and antiviral ranked gene lists generated by MAIC, accompanied by each gene’s MAIC score, can be found in Table S4.

Hierarchical clustering of the gene lists revealed that screens described within the same studies clustered together, except for the Tran3 screen, which clustered with the screen performed by Li et al. (Fig. 3A) (18, 26). This is likely due to the differences in virus strains used between the screens Tran1 and Tran2 (performed with wild-type A/WSN/1933 and A/WSN/1933 pseudotyped with VSV-G) on the one hand and Tran3 (performed with A/seal/Massachusetts/1-SC35M/1980-GFP) on the other hand. The hierarchical clustering of antiviral screens showed that the screen reported by Gao et al. (33), containing only 150 genes, was the most distant from the other antiviral screens (Fig. 3B).

The study by Song et al. (24) contributed the least to the overall MAIC score of proviral screens (lowest sum of all MAIC scores) due to the limited gene list lengths, whereas the studies by Tran and Ledwith (26) and Zhou et al. (28) contributed the most (highest sum of all MAIC scores). Antiviral gene MAIC scores were dominated by the studies from Sharon et al. (32,34) and Trimarco et al. (34,34), as these gene lists were maximized (2,000 genes). This indicates a potential for the MAIC analysis to be biased toward certain screens due to differences in the numbers of host factors identified by each screen. These contributions of all screens are highlighted by the circos plots demonstrating the number of overlapping genes between the proviral and between the antiviral studies (Fig. 3C and D). Of note, individual screens with longer gene lists contributed more to the top 100 MAIC (Fig. S2).

The top 50 proviral and antiviral genes include well-known host dependency and restriction genes in influenza virus infection (indicated by a red asterisk) (Fig. 3E and F). These include the proviral sialic acid transporters SLC35A1 and SLC35A2 (25, 35), and vacuolar ATPase subunits (ATP6V1B2, ATP6V0B, ATP6V0C, ATP6V1H, ATP6AP1, and ATP6V0D1) (36). Well-known antiviral genes include interferon (IFN) alpha-1 (IFNA1) (37), IFN-stimulated genes [IFN-induced transmembrane protein 2 (IFITM2), 2′-5′ oligoadenylate synthetase 1 (OAS1), OAS3, MX1, PNPT1, and RELA], and B4GALNT2 (38–42). Interestingly, several poorly characterized host dependency factors with evidence from several screens were present in the top 50 proviral MAIC output, including Sirtuin2 (SIRT2), GTPase activator TBC1D20, vascular cell adhesion protein 1 (VCAM1), the serine/threonine kinase WEE1, the geranylgeranyltransferase RABGGTA, and the two pore segment channel 1 (TPCN1) (Fig. 3E). Surprisingly, ALYREF (also known as THO complex subunit 4), a component of the transcription-export (TREX) complex typically considered a host dependency factor during viral replication, ranked first in the top 50 antiviral MAIC output (43, 44) (Fig. 3F). Other less well-characterized host restriction factors constitute ANXA11, OR4F21, FAM177B, and TGM2.

Over-representation analysis of Gene Ontology annotations on MAIC results

To identify enriched pathways and gene sets in the proviral and antiviral gene list acquired through MAIC, ORA with Cytoscape interpretation was performed on the top host dependency (n = 500, evidence from >4 screens) and host restriction (n = 448, evidence from >2 screens) factors ranked by decreasing MAIC score. Enriched pathways (Gene Ontology terms) were further categorized into larger subgroups representing different host cell function categories (Fig. 4A and B). Enriched pathways and additional information (i.e., GO term, P.adj., genes pertaining to individual pathways) can be found in Table S5.

As a proof-of-concept, ORA detected several known host dependency pathways related to influenza virus infection, such as sialylation (n = 12 GO terms), endocytosis (n = 17 GO terms), vacuolar acidification (n = 43 GO terms), nuclear trafficking (n = 20 GO terms), translation (n = 21 GO terms), vesicle-mediated transport (n = 33), and proteasome (n = 12).

ORA on the proviral data set revealed 525 enriched (P.adj. < 0.05) host pathways involved in early and late stages of virus replication (Fig. 4A). Less well-characterized proviral pathways are those involved in the biogenesis of exosomes (n = 12 GO terms, e.g., GO:0099501: exocytic vesicle membrane, GO:0140112: extracellular vesicle biogenesis, and GO:0006904: vesicle docking involved in exocytosis), cell cycle and apoptosis (n = 65 GO terms, e.g., GO:2000134: negative regulation of G1/S transition of mitotic cell cycle, GO:0000819: sister chromatid segregation, GO:1900119: positive regulation of execution phase of apoptosis, GO:0000819: sister chromatid segregation, and GO:0035563: positive regulation of chromatin binding), lipid and glucose metabolism (n = 26 GO terms, e.g., GO:0004169: dolichyl-phosphate-mannose-protein mannosyltransferase activity, GO:0032936: SREBP-SCAP complex, GO:0046889: positive regulation of lipid biosynthetic process, GO:0034375: high-density lipoprotein particle remodeling, and GO:0008410: CoA-transferase activity), and microRNAs (n = 17 GO terms, e.g., GO:0060964: regulation of miRNA-mediated gene silencing and GO:2000631: regulation of pre-miRNA processing).

Interestingly, next to host dependency factors related to the viral replication cycle, ORA on the proviral MAIC gene set revealed the enrichment of numerous pathways involved in the negative regulation of the innate immune system (n = 73, e.g., GO:0039532: negative regulation of viral-induced cytoplasmic pattern recognition receptor signaling pathway, GO:0005135: interleukin-3 receptor binding, GO:0035722: interleukin-12-mediated signaling pathway, GO:1903556: negative regulation of tumor necrosis factor superfamily cytokine production, GO:0045824: negative regulation of innate immune response, and GO:0060339: negative regulation of type I interferon-mediated signaling pathway).

ORA on the antiviral MAIC resulted in the enrichment of 154 pathways and mainly highlighted well-known pathways implicated in the immune system, such as pattern recognition receptors and IFN responses (Fig. 4B). The function of these host pathways in influenza virus infection is well-described. It also revealed pathways related to the different stages in the viral replication cycle (n = 22 GO terms, e.g., GO:0060700: regulation of ribonuclease activity, GO:0006909: phagocytosis, and GO:0051170: import into the nucleus).

In summary, the top 50 MAIC genes and ORA highlighted several underexplored host factors that were present in multiple screens, which merit further research. Several of those poorly characterized host factors are presented in Table 2 and Fig. 5A and B. In addition to ORA analysis, we performed Gene Set Enrichment Analysis (GSEA) on the entire ranked proviral (n = 12,613) and antiviral (n = 4,856) gene lists generated from MAIC. This revealed a similar, albeit less elaborate, list of enriched pathways compared to the g:Profiler analysis, given the fact that GSEA is optimized for use on genome-wide data sets (Table S6; Fig. S3).

Overlap of CRISPR meta-analysis by MAIC with previous RNAi meta-analysis

To identify whether CRISPR or RNAi screening methods may be more suitable to study host dependency factors that belong to specific stages in influenza virus replication, we performed a crossover of our meta-analysis on proviral CRISPR screens with high-evidence proviral genes (n = 153) acquired through a meta-analysis on RNAi screens with influenza A virus (IAV) by others (17). These genes (Table S7) were identified by at least two RNAi screens or were derived from a meta-analysis performed by the authors. We focused on their proviral gene list, as the antiviral gene list contained only two genes with evidence from multiple RNAi screens (17).

First, we plotted the RNAi-based genes and their corresponding rank in the MAIC-ranked gene list from CRISPR proviral screens (Fig. 6). An enrichment of RNAi genes was seen within the first 500 MAIC ranks (bins 1–250 and 251–500), suggesting that there are several genes that were commonly identified by both CRISPR and RNAi screening techniques, corroborating that the MAIC top 500 genes are enriched for genes involved in IAV replication. However, many of the high-evidence proviral genes from the RNAi gene list were located at the lower end of the MAIC gene ranking list (76/153 genes ranked 1,500 or higher) or were not present in it (unranked, 40/153 genes), indicating that RNAi screens may be more likely to identify these host dependency genes.

Twenty-four proviral genes were shared by the RNAi gene list and the top 500 MAIC, many of which were associated with vacuolar acidification, ribosomes, and nuclear trafficking (Table S8). However, the precise involvement in influenza virus replication of several of these common host dependency factors remains to be elucidated (Table 2). Several host factors are specific to either CRISPR or RNAi, which are categorized by stages in viral replication. Examples of host factors involved in late-stage influenza virus replication specifically found in the RNAi screens include EIF3F and EIF3A (translation), COPA and ARCN1 (vesicle-mediated transport), and BUB1B, PSMD14, PSMA1, PSMD1, and PSMD2 (proteasome), as indicated in Table S9.

To increase our understanding of the overlap between CRISPR and RNAi on a broader level, we compared the ORA-enriched pathways in the top 500 MAIC proviral gene list to those enriched in the RNAi list (Fig. 7A; Table S10). Generally, the proviral hits from both RNAi and CRISPR are enriched for pathways covering early and late steps of the viral replication cycle. We find that both techniques may be better suited to study specific stages of viral replication. For instance, RNAi genes are highly enriched for pathways related to mRNA splicing and DNA-3/7 methylguanine glycosylase, while this is not so much the case for CRISPR. On the other hand, sialylation, miRNA, and histone-related pathways were only identified by CRISPR. Pathways related to the replication and transcription, immune system, and cytoskeleton were predominantly discovered by CRISPR. Both techniques appeared to be equally suited for the detection of pathways associated with endocytosis, vacuolar acidification, nuclear transport, translation, ribosomes, vesicle-mediated transport, autophagy, exosomes, kinases, cell cycle and apoptosis, and lipid and glucose metabolism.

Re-analysis of the overlapping pathways between a reduced list of the top CRISPR genes (top 153 instead of 500) and the RNAi gene list revealed a higher number of enriched pathways in the RNAi gene list, despite equal numbers in both gene lists (Fig. 7B; Table S10). Compared to RNAi, the top 153 gene list of CRISPR was mostly enriched in pathways related to early-replication events such as sialylation and vacuolar acidification. This suggests that the “top” hits identified in CRISPR screens are biased toward early events in virus replication, in accordance with the fact that pooled CRISPR screens can be biased toward host factors affecting viral entry-replication (45). This bias constitutes a potential source of disagreement between the RNAi and CRISPR data sets. Interactive plots displaying gene count, enrichment score, pathway name, GO term, and involved genes can be found in Fig. S4 and S5.

Essential genes: RNAi versus CRISPR

In loss-of-function (LOF) screens, genes that result in cell death upon knockdown or knockout are commonly known as essential genes. To investigate whether the results from CRISPR and RNAi contained essential genes, we compared the RNAi gene list and our high-confidence proviral genes from the MAIC analysis of CRISPR screens (top 500) to a previously established gold standard of essential genes (n = 217) (Table S11) (46). RNAi showed a significantly increased proportion of essential genes in the total high-confidence genes compared to the top 500 MAIC [Fig. 7C, Fisher’s exact Test: P < 0.0001, 18.3% (28/153) versus 4.2% (21/500)] and the top 153 MAIC [Fig. 7D, Fisher’s exact Test: P < 0.0001, 18.3% (28/153) versus 1.3% (2/153)]. This suggests that CRISPR may be more suited to identify host targets that likely do not result in reduced cell viability when inhibited.

Data eligibility

Screens were included or excluded based on pre-set inclusion and exclusion criteria as outlined in Table 3. We included pooled genome-wide CRISPR-Cas screens using any engineered or wild-type influenza A or B viruses, except for pseudotyped viruses that largely consisted of a virus other than the influenza virus since the gene list for such pseudotyped viruses is expected to primarily reveal genes that are involved in the replication of the pseudotype virus backbone. Screens using an arrayed setup were excluded because such screens require a different readout, which could lead to bias in this meta-analysis. In addition, screens focusing on a subset of genes were excluded to reduce bias toward those genes. Screens to identify antiviral or proviral genes using cells originating from a vertebrate, conducted in a LOF or gain-of-function (GOF) CRISPR-Cas setting, were included. In case gene lists of at least 150 genes or raw data (including relevant samples and sgRNA barcodes) were not available, the screens were excluded. Studies in peer-reviewed and preprint publications were included.

Literature search and study selection

A systematic review based on the guidelines of PRISMA was performed using PubMed, Scopus, and Web of Science. Searches were performed on 14 March 2023 and updated weekly until 28 June 2023. The following search keywords were used, without date or language restrictions: (((genome) OR (genom*)) AND (crispr)) AND (influenza). All returned papers were first assessed for duplication and relevance based on title and abstract. Subsequently, full-length texts were obtained to confirm eligibility according to Table 3.

Gene list extraction and preparation for MAIC

Data relevant to the MAIC analysis and other screen-specific parameters were extracted from eligible studies, and the screens were assigned to the antiviral or proviral group, depending on their experimental setup. Screens were considered to be ranked when the gene list originated from a gene-ranking analysis, or otherwise unranked. MicroRNAs were removed from the gene lists, whereas non-human genes were mapped to their human homologs using BioMart (Ensembl) (97) or excluded from the analysis if no human homolog could be identified. To this end, the Sscrofa11.1 database was used to convert porcine cell lines originating from Sus scrofa (NCBI: txid9823), while the pteVam1 database was used for the bat cell lines originating from Pteropus alecto (NCBI: txid9402). Per MAIC requirements, gene lists were limited to the first 2,000 genes.

Meta-analysis by information content

MAIC allows the combination of data sets that are expected to have genes in common retrieved from heterogenous sources with unknown quality (18). In our case, these data sets are gene lists {L₁, L₂, L₃,… L_m} that contain n genes or entities {e₁, e₂, e₃,… e_n} and are assigned to one of K categories {C₁, C₂, C₃,… C_K}. In the original paper on MAIC (18), categories are defined as a particular experimental data type, such as protein-protein networks or results from RNAi or CRISPR screening to account for technological bias. Instead, we used the different studies (often consisting of multiple screens) as categories, assuming that multiple CRISPR screens performed in the same study are more likely to identify similar genes due to similar experimental conditions. Each gene was assigned a score based on overlap with other gene lists and, if available, its gene rank. A gene represented in different categories will receive a higher score than a gene represented in different gene lists within a category. A weighting score w is calculated for each gene list L by summing the scores of each gene within the gene list. This results in higher weighting scores given to gene lists that have a higher proportion of genes with supporting evidence from multiple studies. Entity scores are iteratively recalculated using the weighting score, and conversely, the weighting scores are recalculated by summing the updated gene scores until convergence occurs. The final output is generated by combining the evidence from all lists, resulting in a gene list with the genes having the highest MAIC score ranked first. Algorithmic details and evaluation on MAIC can be found in the original paper (18).

Clustering and visualization of MAIC results

Gene lists from proviral and antiviral screens that were used for the MAIC analysis (i.e., the first 2,000 ranked genes, as listed in the MAIC output) were hierarchically clustered based on the distance matrix computed based on the Jaccard similarity coefficient. The gene lists were regarded as bits, where non-zero elements mean the relevant gene is present in the gene list, while zero elements mean they are absent. To visualize overlapping genes between studies, circos plots were generated using the circlize package in R (version 4.3.0) (98, 99). All R coding that was used to generate the figures in this paper can be found at https://github.com/annabelmaes (repository: MAIC-FluA-CRISPR-meta-analysis).

Over-representation analysis and Cytoscape interpretation

Significantly (P.adj < 0.05) enriched pathways were identified through functional pathway enrichment analysis by g:Profiler using the GO pathway databases for biological processes, molecular functions, and cellular components, followed by Cytoscape to interpret and illustrate their relationship. Fisher’s exact test or modified Fisher’s exact test was performed on unranked (RNAi gene list, top 153 MAIC gene list) and ranked gene lists (top 500 proviral MAIC genes, evidence from >4 screens or top 448 antiviral MAIC genes, evidence from >2 screens), respectively. Visualization of the gene overlap among the enriched pathways was created by the EnrichmentMap application in Cytoscape (100). The AutoAnnotate Cytoscape application was used to automatically define and summarize clusters, which were further categorized into larger subgroups representing the different stages in viral replication and related pathways (101). This workflow was described elsewhere (102).

Gene Set Enrichment Analysis

GSEA was performed on all genes from the ranked proviral (n = 12,613) and antiviral (n = 4,856) MAIC output, ranked by their MAIC score, using the WEB-based Gene SeT AnaLysis Toolkit with the estimation of P-values using an empirical probability distribution based on 10⁴ permutations (103). The following databases were interrogated, KEGG 2021 database, Gene Ontology 2021 (biological processes, cellular components, and molecular function), and WikiPathways 2021. Each gene set received a normalized enrichment score, with higher values correlating with a tendency of genes to rank at the top of the input gene list. The Benjamini-Hochberg procedure was used to adjust for multiple testing and to calculate the false discovery rate (P.adj).