The first step of this approach consisted in identifying and extracting a list of genes known to be associated with the Adverse Outcome (AO) of interest, which in our case was microcephaly, by querying various public databases. The selected databases for this analysis were GeneCards (https://www.genecards.org/- version 5.11) and DisGeNET (https://www.disgenet.org/- version 7.0). GeneCards is an integrative database that provides information on all annotated and predicted human genes by aggregating different web sources. DisGeNET is a database that integrates a collection of genes associated with human diseases and compiles more than 20,000 unique human genes and orthologs from various sources, including human/animal databases and text mining. In both databases, the results are sorted based on a score (gene disease association—gda score for DisGeNET and Relevance score for GeneCards) that takes into account various information such as curation level, organism, the different resources and the number of publications supporting the association between the gene and the disease.

These two databases were queried using the term “Microcephaly” (corresponding to our AO of interest), and all genes associated with this query were downloaded. Since the process of neurogenesis is well conserved across vertebrate species, we decided to integrate orthologs (rat/mouse) to retrieve as much important information as possible. To facilitate data integration and avoid potential redundancy between species, all genes were mapped to their HUGO symbols using the HGNC Comparison of Orthology Predictions (HCOP) tool (https://www.genenames.org/tools/hcop/).

To improve the confidence level of our study and retain only the most relevant genes for our analysis, we applied a filter based on the number of publications and excluded any genes with fewer than 2 scientific publications associated with microcephaly referenced on GeneCards and DisGeNET. The resulting gene list after the exclusion filter served as input for the next step of text mining (Figure 1; Supplementary Table S1).

The second step aims to establish a connection between the genes extracted in the first step and potentially known stressors using contained in the abstracts of PubMed literature. In our case, we aimed to associate our list of genes associated with microcephaly with exposure to ionizing radiation (IR). To achieve this, we employed AOP-helpFinder (v1.0), a hybrid tool that combines text mining and graph theory. This tool can be accessed online at http://aop-helpfinder.u-paris-sciences.fr/index.php (accessed on 22 July 2022). This tool is specifically designed to identify and extract connections between co-occurring terms, including stressor-event or stressor-gene relationships (Gundacker et al, 2022). The program converts sentences into a graph structure, where words are depicted as nodes connected by weighted edges that reflect the proximity between words. This graph-based representation allows AOP-helpFinder to calculate scores, taking into account the distance between words (distance score) and their position in the abstract. As a result, it allows for capturing and retaining the most pertinent links.

The tool requires two inputs to operate.

1. A list of key events, which in this case corresponds to our list of genes associated with microcephaly extracted from the GeneCards and DisGeNET databases (step 1).

Using these inputs, AOP-helpFinder automatically analyzes the entire PubMed abstracts related to IR to identify connections with the list of genes associated with microcephaly. To enhance confidence in the obtained results, we configured the “Reduced search” parameter to 20%. This option, based on the position of the links in the abstract, provides fewer results compared to the default parameter (0%), but it significantly reduces the rate of false positives by excluding links found in the first 20% of abstracts, which often correspond to the introduction or literature review sections rather than results.

As output, AOP-helpFinder provides a collection of articles that support the association between genes and IR. This text mining step plays a crucial role in our strategy as it enables the identification of biological knowledge that connects IR to microcephaly, thereby providing more detailed mechanistic information.

The last step was the enrichment and analyses of all the extracted data to identify potential KEs of the AOP, from the biological information linking IR and microcephaly. A biological enrichment analysis was performed with g:Profiler enrichment tool (Raudvere et al, 2019) (version e107_e.g.,54_p17_bf42210) using the list of selected genes common to IR and microcephaly in order to identify the most impacted pathways. In this approach, biological enrichment differs from its original use, as the list of genes extracted in the previous steps is provided as input, without considering any expression level. In systems biology, this strategy has already been applied to complement and classify the extracted data (Rugard et al, 2020; Wu et al, 2021). The over representation analysis (ORA) was done by querying different databases: KEGG (Kyoto Encyclopedia of Genes and Genomes—2022/09 release) and REACTOME (2022/09 release), which contain a collection of manually curated pathways as well as Wikipathways (2022/09 release), a community platform where pathways are curated (as of 2023, January). A hypergeometric test, corrected by the Benjamini–Hochberg method was done to adjust p-values. A significance level of 0.05 for the adjusted p-values was used to select significant pathways. In addition to the enrichment results, a gene-pathway network was created using KEGG and Reactome classification. The important identified pathways may thus correspond to Key Events leading to microcephaly, and the genes classified in these pathways may provide information on the Key Event Relationships (KERs). After identifying relevant information through the enrichment and the network analyses, a manual curation by experts of the full text articles identified by AOP-helpFinder was performed to provide the mechanistic and biological information leading to radiation-induced microcephaly (Figure 1).

The screening of both the GeneCards and the DisGeNET databases, allowed the identification of 6,708 and 1,064 genes respectively, described as being associated in the progression of microcephaly. As more than 80% of these genes have no bibliographic supports, we increased the level of confidence and the robustness of this dataset by keeping only the protein-coding genes with at least two scientific publications describing an association with microcephaly (Supplementary Figure S1). This led to the identification of 348 genes from GeneCards and 117 genes from DisGeNET, corresponding to a total of 382 unique genes linked to microcephaly progression (Supplementary Table S1). It is important to note that no causative relationship between the gene and microcephaly should be expected from this screening, although this method could identify the majority of currently known primary hereditary microcephaly (MCPH) genes (Siskos et al, 2021). For example, the genes MCPH1, ASPM (MCPH5), CENPJ (MCPH6), WDR62 (MCPH2), and CDK5RAP2 (MCPH3) are among the genes with the highest gda score (DisGeNET) and Relevance score (GeneCards), indicating their significant association with microcephaly as evidenced by a large number of publications referenced in both databases (Supplementary Table S1).

The 382 unique genes were combined with the stressor “Ionizing radiation” as inputs for the text mining tool AOP-helpFinder. Through a comprehensive analysis of 131,781 abstracts related to IR, which were extracted from a pool of 34 million abstracts available in the current PubMed database, AOP-helpFinder successfully established a link between 101 genes and IR (Supplementary Table S1). We found 20 genes with more than 50 links, including many known genes associated in the progression of radiation-induced microcephaly (Figure 2). For instance, the tumor suppressor gene TP53, a major key event in the microcephaly AOP, was linked to over 400 articles on IR. Additionally, several genes involved in the DNA damage response, another key event in AOP 441, were found to have multiple links to IR, including ATM (over 1,200 links), ATR, H2AX, and BRCA1 (Jaylet et al, 2022). These genes have also been well studied in the context of microcephaly, with 9 articles describing a link between TP53 and microcephaly, and 5 articles describing a link between ATM and microcephaly according to GeneCards (Supplementary Table S1), demonstrating consistency between the results of this case study and those verified in AOP 441.

The proposed in-silico strategy allowed the identification of 101 genes associated with biological processes that can lead to the progression of radiation induced microcephaly. We performed a biological enrichment analysis on this gene list, by querying different pathway databases (KEGG; Reactome, Wikipathways), to identify the most represented biological pathways associated with radiation-induced microcephaly, and thus to identify potential novel KEs.

The most statistically significant results corresponded to pathways known to be involved in the response to IR. Particularly, a set of pathways associated with DNA damage and repair (such as double strand breaks DSB repair, DNA repair, DSB response with corresponding adjusted p-values ranging between 8e-27 and 8e-12 on Reactome), pathways associated with P53 (e.g., Regulation of TP53 activity, P53 signaling pathway with adjusted p-values lower than 1e-11 on Reactome and 1e-5 on KEGG), apoptosis (adjusted p-value = 2e-7 for KEGG) were identified (Supplementary Table S2). The current preliminary AOP for microcephaly (AOP 441) is composed of the following KEs: “DNA damage” inducing “Activation of TP53” leading to “Apoptosis” and “Premature cell differentiation” leading to “Microcephaly.” These findings confirm the efficiency of the automated strategy to identify known KE relevant for AOP development (Figure 3; Supplementary Figure S2).

We next constructed a network linking the different enriched pathways via the genes involved in these pathways. Indeed, genes can be part of different pathways. Thus, the created gene-pathway network allowed to explore the relationships between KEs and to identify genes potentially involved in KERs. For example, there is a large panel of genes involved in both the response to radiation-induced DNA damage and TP53 activation. Among the genes coding for proteins important for the DNA-damage response, the sensor genes MRE11 and Rad50 of the MRN complex were found, as well as important transducers such as ATM or ATR and ATRIP and mediators (MDC1, TP53BP1, BRCA1, RBBP8, Checkpoints Kinases and histone H2AX linked to DNA damage) (Figure 3).

Therefore, these results show that this automated strategy quickly identifies biological pathways known to be involved in radiation-induced microcephaly. In addition, it identifies genes and potential molecular mechanisms involved in the considered enriched biological process thereby providing a basis for the description of KERs. However, it is important to stress that this approach does not necessarily provide causal or mechanistic links between genes, pathways and disease. Rather, the use of AOP-helpFinder combined with these analyses can provide a rapid prioritization of the knowledge from the relevant literature to further explore potential mechanistic links leading to radiation-induced microcephaly. Based on the evidence automatically gathered here, the KER describing the relationship between the KEs “DNA damage” and “TP53 activation,” can be the activation via phosphorylation of ATM following sensing DNA double strand breaks (DSB) (Uziel et al, 2003). Indeed, the activation of ATM leads to the activation of various pathways related to TP53, including the histone H2AX, as well as various mediators such as MDC1 and TP53BP1 which will accumulate at DNA damage sites following ATM activation (Burma et al, 2001; Goldberg et al, 2003; Lou et al, 2006). All these genes, critical in the cellular response to DNA damage, were detected by AOP-helpFinder as part of the pathways induced by IR and possibly linked to radiation-induced microcephaly if damages are not efficiently repaired. This cascade of events finally leads to the activation of effector kinases (e.g., checkpoint kinases) and tumor suppressor protein TP53, which will have a key role in inducing cell cycle arrest and in the choice between DNA repair, apoptosis, senescence or differentiation depending on the degree of DNA damage (Banin et al, 1998; Turenne et al, 2001; Fei and El-Deiry, 2003).

Novel potential KEs, not described in the preliminary AOP 441, were also identified by this non-supervised strategy. Among these, several genes involved in the cell cycle (adjusted p-value = 1.6e-9 for KEGG and 1.43e-19 Reactome) or senescence (adjusted p-value = 5.48e-10 for KEGG and 5.6e-7 for Reactome) were retrieved (Figure 3). Notably, these two pathways were considered in the initial review of the AOP 441, but not included due to lack of scientific evidence. The global approach used here can thus provide additional mechanistic knowledge to the existing AOP. Many cyclin-dependent kinase genes (e.g., CDK2; CDK6; CDC6) are involved in these pathways (Figure 4). Cyclin-dependent kinases are responsible for TP53-mediated cell cycle arrest via P21 (CDKN1A) (Liu et al, 2020), but also appear to be involved in non-TP53 dependent pathways. Indeed, it was observed a downregulation of Cdk2 and Cdc6 after irradiation of TP53 null mutant mouse embryos. The downregulation of these Cyclin-dependent kinases can be explained by an upregulation of some members of the E2F transcription factors family, therefore suggesting an alternative pathway of TP53 leading to cell cycle arrest (Verheyde et al, 2006; Verheyde and Benotmane, 2007). Cell cycle arrest leads then to a decrease in cell proliferation, impacting the proper development of the brain and therefore representing an additional risk factor for microcephaly (Barkovich et al, 2005; Francis et al, 2006; Passemard et al, 2013). From these data, it is therefore possible to add the two additional KE 1505 “Cell cycle, Disrupted” and KE 1821“Decrease, cell proliferation” the existing AOPs for microcephaly (Figure 5).

Moreover, a wide range of genes present in the network are found differentially expressed in irradiated rodent animal models (Verheyde et al, 2006; Quintens et al, 2015; Mfossa et al, 2020). For example, the p53 activation following DNA damage triggered by IR leads to the induction of genes associated with differentiation and development processes, responsible for premature differentiation of neuron progenitors (Quintens et al, 2015). The key gene ASPM, known to be responsible for primary microcephaly when the protein is disrupted (Bond et al, 2003; Létard et al, 2018) was also found by our approach and linked to IR. On different human and murine cell lines as well as on fetal mouse brain, a study has shown that IR lead to a decrease in ASPM mRNA level associated with a reduction of ASPM protein (Fujimori et al, 2008). Moreover, it was observed that this downregulation of ASPM by IR was accompanied by a reduction of BRCA1, also present in our network. ASPM has many roles explaining its involvement in the process of microcephaly. This protein is involved in brain development by maintaining neural progenitor division, proliferation and differentiation (Fujimori et al, 2014; Létard et al, 2018) and is also required for the proper functioning of DNA repair. A recent study showed that ASPM facilitated DSB repair by stabilizing BRCA1 and protecting it from degradation by the proteasome, and that inhibition of ASPM expression led to an increase in BRCA1 ubiquitination leading to a decrease in its protein levels (Xu et al, 2021). Moreover, downregulation of Aspm was also observed in embryonic murine p53 null, meaning that the alteration of its expression is not or not totally due to TP53 activation (Fujimori et al, 2008).

Thus, a new KE “Alteration of gene expression” could be proposed for our AOP, which would lead to cell cycle arrest (Verheyde et al, 2006; Mfossa et al, 2020), impaired proliferation and differentiation of neural stem/progenitors cells (Verheyde et al, 2006; Fujimori et al, 2014; Quintens et al, 2015; Létard et al, 2018; Mfossa et al, 2020), as well as inadequate DNA damage repair (Xu et al, 2021) (Figure 5). Indeed, defective DNA repair is an important risk factor in microcephaly process and can amplify all the KEs of the AOP (i.e., excessive activation of TP53; apoptosis; premature cell differentiation; cell cycle arrest etc.). For example, DNA damage repair defects due to PNPK mutations (Shen et al, 2010) or DNA ligase IV deficiency (Chrzanowska et al, 2012) found in our network, result in radiosensitivity and can lead to the development of microcephaly.

The information on the newly identified potential KEs could be completed by using the KERs already present on AOP-Wiki. Thus, a two-way relationship between “Increased DNA damage” and “Inadequate DNA repair” described through KERs 1910, 1911 and 2,608 could be added, supporting the notion of an excessive accumulation of DNA damage leading to an overwhelmed DNA repair process. Relationships between “Increased DNA damage” and “Altered Gene expression” through KER 1301 or between “Increased DNA damage” and “Cell cycle, Disrupted” through KER 2790 could also be added, suggesting in particular that the modification of the expression of some key genes is not totally due to TP53 activation.