The association between neurodevelopmental disorders (NDDs) and certain types of cancers has been implicated in several epidemiological studies. Compared to age- and sex-matched individuals in the general population, patients with schizophrenia have a roughly 50% higher risk of dying from cancer (Nordentoft et al., 2021), with seven quantitative studies of 1,162,971 participants yielding mortality risk from breast, colon, lung, and prostate cancer (Ni et al., 2019). Individuals with bipolar disorder and their unaffected siblings under 50 years of age had a higher chance of developing breast cancer than those in the control group. The correlation between bipolar disorder and the younger population’s susceptibility to bipolar disorder and higher cancer risk could indicate a genetic overlap in the pathophysiology of neurodevelopment (McGinty et al., 2012; Peng et al., 2021; Chen et al., 2022). When comorbid intellectual disability and/or birth defects are present, those with autism spectrum disorders are more likely than those without autism spectrum disorders to have cancer in their early years. After controlling for characteristics such as sex, birth year, parental age, etc., the link persisted and was not expected to be caused by other confounding variables (Liu et al., 2022). The relationships between bipolar disorder, autism, and schizophrenia and malignancies of the breast, colon, thyroid, and lung are summarized in Figure 1A; if a particular NDD has been linked to a greater cancer incidence in a tissue, we link them together with a line.

Epidemiological cohort studies opened the quest for the molecular mechanisms underlying NDDs and cancer - pathologies with distinct clinical presentations. Subsequent findings have suggested a cellular and organismal relationship between these pathologies. The Ras network impacts both diseases by aberrant regulation of the cell cycle (Nussinov et al., 2022b; Nussinov et al., 2023c). Two major mitogen-stimulated signaling pathways feed into the cell cycle (MacCorkle and Tan, 2005; Foijer and Te Riele, 2006; Sever and Brugge, 2015; Yang et al., 2017; Min et al., 2020). The first is MAPK, which controls cell division and is the major pathway in cell proliferation (Guo et al., 2020). The second is PI3K/AKT/PDK1/mTOR, the primary pathway in cell growth, thus differentiation (Saltzman, 2004; Li and Kirschner, 2014; Ruijtenberg and van den Heuvel, 2016). Proliferation and differentiation are vital to both cancer and NDDs. The fact that the same proteins in these pathways, and even the same mutated residues are involved (Marmion et al., 2023), raises compelling questions, including (i) why the phenotypic presentations are vastly different, and (ii) how we can develop a strategy to identify and predict NDD-related mutations and distinguish them from those related to cancer even though the same residue is mutated. If we are able to develop such a strategy, it may help in early NDD diagnosis. Epidemiologic surveys for multiple NDDs unearthed a clinically intriguing and abysmal connection: individuals afflicted with NDDs may face higher probabilities of developing cancer later in life (Nordentoft et al., 2021; Liu et al., 2022). We believe that addressing the link between these two pathologies will contribute to deciphering the underlying molecular mechanisms, treatment and preventive care.

Our premise is that the differences between cancer and NDDs are largely the outcome of the perturbation of signaling levels, which depend on (i) gene expression, which in turn, depends on cell type and state, and timing window (e.g., embryonic developmental stages or throughout life); (ii) homeostatic mechanisms that can block or enhance the signal; (iii) the strength of the activating mutation; (iv) the types and locations of additional mutations, and (v) the expression levels of specific isoforms of genes and regulators of proteins in the pathway. Expression levels indicate the role of chromatin structure. The sparseness of data makes single cell transcriptomics challenging, and structural networks may not be able to distinguish between isoforms whose sequences (and structures) are highly similar yet have different functions (Nussinov et al., 2022b; Nussinov et al., 2023c).

As to why the epidemiological connection, we reason that one possible clue is the established cancer statistics: A single mutation is insufficient to elicit cancer, although exactly how many mutations are required for cancer to emerge has been a debated question (Tomlinson et al., 2002; Tomasetti et al., 2015; Martincorena et al., 2017; Liu et al., 2023). Going back to the question we posed above, why then individuals with neurodevelopmental disorders, such as autism and schizophrenia, may face higher chances of cancer emerging throughout their lifetime? NDDs frequently involve germline mutations inherited from a parent who may not show the related phenotype. Activating mutations emerging during life can couple with the pre-existing germline mutations, strengthening the proliferative signaling and the likelihood of cancer. Such a scenario resembles latent driver mutations in cancer (Nussinov and Tsai, 2015; Nussinov et al., 2019a; Nussinov et al., 2023b). Latent drivers have low frequencies; thus, their translational potential may have escaped detection. However, when paired with another mutation in the same allele, they can drive cancer. Our comprehensive pan-cancer statistical analysis observed a significant occurrence of the double mutation pairs (Yavuz et al., 2023b). This scenario may manifest an otherwise unobserved NDD pathology in the parent. The computational challenge is to identify the impact of such mutations within phenotype (Nourbakhsh et al., 2024).

Here, we aim to enlist computations to discover how same pathway and same-gene mutations, and overexpression can preferentially lead to cancer and NDDs phenotypes, with the overarching goal of applying the learned knowledge to identify likely culpable genes/mutations (Figure 1B). Our hypothesis-driven proposal is consistent with early experimental indications (Marshall, 1995; Murphy et al., 2002; Murphy et al., 2004; Ben-Ari et al., 2010). Our premise is that under physiological conditions, mitogen-induced signaling that propagates down the pathway into the cell cycle is strong, but the bursts are of short duration (Figure 1C). In cancer, constitutive activating mutations and overexpression induce strong signaling that does not abate over time. A long duration of potent, too strong signaling can elicit oncogene induced senescence. In NDDs signal strength is moderate and long. Strong signaling bursts is expected to promote cell proliferation, weaker signaling to foster differentiation, and moderate/weak signal strength in aging may be associated with neurodegenerative diseases. Comparisons of NDDs and cancer—including the observed mutation types, transcriptomic data, protein-protein interaction (PPI) networks, and protein conformational dynamics (Jang et al., 2023b) —can suggest the (relative) signaling strength. We expect signaling strength in benign pediatric tumors (Li and Langhans, 2021), like cutaneous neurofibromas, to resemble that in NDDs. Individuals with NF1 mutations might be at risk of developing certain tumors earlier in life (Landry et al., 2021), with unidentified latent embryonic mutations rendering them genetically predisposed.

The significance of resolving the connection between cancer and NDDs (Nussinov et al., 2022b; c), coupled with advances in the computational sciences, machine capabilities, and the accumulation of experimental data across size scales suggests diverse approaches to decipher the connection and causality between these two complex processes. The challenge is in selecting the questions, the right tools (artificial intelligence (AI) based such as machine learning (ML), deep learning (DL), or other, advanced, or classical methods) to target them, and the protocols (Zeng et al., 2022). Among these, molecular dynamics simulations, which are essentially a single molecule approach, can also be considered as a predictive tool of cell phenotype (Nussinov et al., 2023b; a). Tool selection is also dependent on how much data (and of what quality) are available. Below, we provide an overview of the computational modeling and pipelines aiming to elucidate the connections and causalities. Figure 2 overviews the toolboxes and resources we envision for potential applications and data processing.

Cancer and NDDs catalyze and are impacted by dysfunctions in biological processes and pathways, including chromatin remodelers, PI3K/mTOR, and MAPK pathways (Zheng et al., 2018; Nussinov et al., 2023c). Therefore, the identification of functional items—modules or pathways—is a critical component in the comparison of the two diseases. Upon comprehensive characterization of the implicated pathways, it would also be possible to propose pertinent combination therapies for effective treatment (Nussinov et al., 2024). In such a task, network models are useful tools as they methodically address complex disorders by looking at dysregulated modules rather than the effects of individual mutations or genes in the conventional reductionist paradigm (Williams and Auwerx, 2015). Molecular networks exhibit high modularity, which can be defined as a high probability linkage among subsets of nodes. Specific modules frequently consist of different genes or proteins engaged in the same biological processes (Choobdar et al., 2019; Cui et al., 2019). Network models can also explain hypothetical causal mechanisms where multiple perturbations, such as mutations, are linked to pathways and processes (Nogales et al., 2022; Amgalan et al., 2023). In the mechanistic explanation of disease networks, cellular signal transduction is effectively depicted with causal knowledge such as kinase-transcription factor (TF) and TF-target interactions (Liu et al., 2019; Dugourd et al., 2021). Modular comparisons of network models can illuminate functional details of the commonalities and disparities between NDD and cancer.

Computational pipelines start with an input dataset, which may include genomic alterations, transcriptomics, proteomics, metabolomics, and any prior knowledge combinations (Milan Picard et al., 2021; Demirel et al., 2022) (Figure 2). Numerical features can be obtained from raw feature data using feature extraction (or selection) techniques. They can be used to eliminate redundant data so that the chosen learning algorithm only includes the most pertinent information in order to reach a biological interpretation. This would allow for inferences, such as the removal of diseased samples from healthy ones and the extraction of biomarkers for diagnostic purposes.

As the success of a computational pipeline heavily depends on the data, the selection of the appropriate dataset(s) relevant to the biological question and the necessary analyses is a crucial first step. Cancer-related datasets are available in many documented and controlled databases, such as TCGA (Cerami et al., 2012; Liu et al., 2018), AACR Project GENIE (Consortium, 2017), MSK-IMPACT Clinical Sequencing Cohort (Zehir et al., 2017), and primary and metastatic tumors from the Hartwig Medical Foundation (Martínez-Jiménez et al., 2023), thanks to the perpetual focus on cancers heretofore. While not being as widespread as cancer datasets, NDD-related datasets are also becoming increasingly available, as interest in understanding both NDDs and the cancer-NDD association has been on the rise (see Table 1).

The online single- and multi-omics data resources focused on cancer have been cataloged in Das et al.'s comprehensive review (Das et al., 2020). Several reviews covering emerging omics technologies for prognosis, early cancer screening, and diagnosis were published, along with a roadmap for multi-omics data integration techniques based on statistical methods and artificial intelligence (Biswas and Chakrabarti, 2020; Arjmand et al., 2022; Cai et al., 2022; Ma et al., 2022; He et al., 2023). Multi-omics integration techniques were applied to cancer research as well as to phosphoproteomics data (Mantini et al., 2021). Utilizing a proteogenomic approach to tumor investigation, the Clinical Proteomic Tumor Analysis Consortium (CPTAC) of the National Cancer Institute generates rich multi-omics datasets that link genomic anomalies to cancer descriptors (Li et al., 2023). Data on the full exome, whole genome, transcriptome, proteome, and phosphoproteome are provided for 10 cancer types including glioblastoma multiforme (GBM), lung adenocarcinoma (LUAD) and breast cancer (BRCA).

Computational studies mostly harness genomic and transcriptomic datasets to remedy biological questions (Figure 3). The genetic etiology of the NDDs is widely investigated through mutations (obtained with whole genome/exome or targeted sequencing) and copy number variations (CNVs). Comparisons of transcriptomic datasets, either bulk or single-cell type RNA-seq data, can identify disease. Further functional and molecular insight can be captured through proteomics and metabolomics approaches as well as post-translational modifications (Nascimento and Martins-de-Souza, 2015; Mueller and Meador-Woodruff, 2020; Murtaza et al., 2020; Ristori et al., 2020; Kopylov et al., 2023). Additionally, prior knowledge deposited in curated databases can provide information, including on protein structures (Berman et al., 2000; Varadi et al., 2021), interactomes (Alanis-Lobato et al., 2016), regulatory networks (Ben Guebila et al., 2021), metabolic pathways (Kanehisa et al., 2019), cancer drivers (Martinez-Jimenez et al., 2020), disease-related genes (Arpi and Simpson, 2022), and more, for sample-derived knowledge.

Biological datasets are exceedingly complex and high dimensional as they cover thousands of genes, transcripts, and proteins. Feeding these datasets as-is into learning-based models will lead to incorporation of noisy or repetitive data and increase the computational costs. Feature extraction and selection methods prevent such problems as they enable extraction of the most relevant and simple features (Figure 4). Similarity matrices keep the pairwise similarities between the data points by utilizing measures such as Euclidean distance, cosine similarity, Jaccard similarity, and Pearson correlation coefficient (Guo et al., 2018). Dimensionality reduction methods–such as principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), singular value decomposition (SVD), or non-negative matrix factorization (NMF)– can be applied to reduce the dimension of the similarity matrix along with the statistical tests and feature rankings to represent the data with minimum information loss (Meng et al., 2016; Guo et al., 2018; Reel et al., 2021). The above-mentioned dimensionality reduction methods like PCA, tSNE, and UMAP can be used for data visualization in 2D or 3D by reducing the dimensional complexity while keeping the maximum amount of information (Kim et al., 2018; Do and Canzar, 2021). Networks, such as regulatory networks, coexpression networks, or protein-protein interaction networks, transform multiple input datasets into more comprehensible data by using global and local features of networks (Li and Gao, 2019; Muzio et al., 2020). A community of interacting genes/proteins in these networks is mainly associated with specific biological knowledge. Thus, the use of communities in downstream methods can be advantageous to transform complex data into interpretable knowledge. Data decomposition approaches computationally break complex data problems into several components (Malod-Dognin et al., 2019). Associated and consistent components or detected -hidden but significant-components constitute feature sets.

The differences in the experimental protocols, tissue heterogeneity, and more, are expected to cause inherent noise among the datasets (Rahnenfuhrer et al., 2023). It is not uncommon to find inaccurate, or spurious correlations between unrelated data points due to the high-throughput nature of omics data, which typically result from a larger number of compared features (Clarke et al., 2008; Pudjihartono et al., 2022). As high-throughput data contains uninformative features, overfitting can lead to low generalization performance, and purging such features is inevitable. However, eliminating features without incorporating domain knowledge can lose critical information (Clarke et al., 2008). The aim of the study, the nature of data, the possible sources of noise should be considered carefully before using dimensionality reduction and other statistical methodologies. Although determining a single concrete map is challenging, the points discussed above could help reducing information loss and wrong derivations could be minimized.

Learning-based algorithms try to learn from the data to be able to predict the correct or relatable outcomes (Figure 5). Supervised learning techniques such as regression analysis, random forest, and decision trees require labeled samples, such as diseased versus control, where the algorithm can create a logical connection between the labels and the data to label possible unknown samples. Unsupervised learning methods such as clustering algorithms and Gaussian mixture models, on the other hand, do not depend on such labels and try to come up with common patterns based on the data itself rather than the labels. Reinforcement learning algorithms aim to maximize long-term rewards by learning from the rewards and penalties based on their actions. Such algorithms can be useful for patient-specific treatment regimens where selections are based on the biological characteristics and treatment response of the patients (Eckardt et al., 2021). DL is based on artificial neural networks and utilizes a multilayered learning structure that may include supervised, unsupervised, and semi-supervised methods to extract non-linear and complex features from high-dimensional datasets. DL models can be used for various biological problems from disease assessment to subtyping and survival analysis to precision medicine (Tran et al., 2021).

Numerous computational techniques have been used to study cancer, taking advantage of massive and quickly expanding databases. In addition to medical records, patient populations, cancer types, and clinical presentations, there are large databases of sequences, mutations, transcriptomics, and experimental 3D structures. Over the last years, computations related to NDDs have also been ramped up (Qi et al., 2016; Jiang et al., 2022). However, to date, they are of a different nature. They include a computational perspective on autism (Rosenberg et al., 2015), including exploiting ML models to understand and diagnose its pathogenesis in the context of complex neurodevelopmental heterogeneity (Jacob et al., 2019). They also include computational neuroscience approaches to identify precise, objective, and quantifiable markers of autism spectrum disorders (ASD) in physiological, behavioral, and neural processing (Computational Models of Autism) and computational analysis of neurodevelopmental phenotypes to harmonize clinical features (Lewis-Smith et al., 2022). These NDDs-related computations are associated with features at a different size scale than the molecular-level methods discussed here. Only one literature record compared the gene expression profiles of ASD frontal cortex tissues and 22 cancer types obtained by differential expression meta-analysis with reported gene, pathway, and drug set-based overlaps (Forés-Martos et al., 2019). The results suggested that brain, kidney, thyroid, and pancreatic cancers are candidates for direct comorbid associations with ASD, whereas lung and prostate cancers are candidates for inverse comorbid associations with ASD.

The most common NDDs are ASD and attention-deficit/hyperactivity disorder (ADHD). Other common NDDs include cerebral palsy, communication disorders, intellectual disabilities, learning disorders, and neurodevelopmental motor disorders (FamilieSCN2A Foundation, 2023). Intellectual disorder (ID), ASD, ADHD, schizophrenia, and bipolar disorder (BD) display a neurodevelopmental continuum that can be explained by gene nucleotide substitutions and copy number variants, e.g., nonsense mutations or splice variants. For example, in ASD, there is a strong association between mutations in CDH8 (a gene encoding Cadherin 8, diseases associated with which include craniofacial-deafness-hand syndrome and ectodermal dysplasia) (Jiang et al., 2022). Similarly, 16p11.2 with duplication resulting in a low weight; a small head size (microcephaly), and with deletions playing a major role in developmental delay, especially in speech and language (Rosenfeld et al., 2010). Affected individuals also have an increased risk of behavioral problems, and SCN2A is commonly associated with early-onset epilepsy and linked to ASD and developmental delay (Morris-Rosendahl and Crocq, 2020; FamilieSCN2A Foundation, 2023).

25 distinct classes of brain cells originating from the primary motor cortex of the mammalians, mouse, marmoset and human, were identified, including 16 different neuronal classes, each composed of multiple subtypes (Network, 2021). The locations of cells related to the NDDs differ, e.g., cerebral palsy-related cells are in the part of the brain that controls movement, likely differing from that of ID. At the same time, comorbidity between certain NDD types suggests some adjoining genes partaking in a common chromosomal deletion or CNVs as in ASD 16p11.2 deletion. The mutations might be present on genes depending on the specific cell types and states, and these genes may exhibit varying levels of co-expression throughout embryonic brain development. Comorbidity between the NDDs is frequent. 22%–83% of children with ASD have symptoms that satisfy ADHD, and 30%–65% of children with ADHD have ASD symptoms (Morris-Rosendahl and Crocq, 2020). Studies focusing on mental disorders in children aged 7 through 12 have also observed that 55% of the children were diagnosed with at least one NDD. 40% did not have a diagnosed comorbid condition, and 26% had an anxiety disorder (Morris-Rosendahl and Crocq, 2020). ID, ASD, ADHD, and schizophrenia share specific genetic risks (Owen et al., 2011). Data suggests that in addition, intellectual disabilities share phenotypes with schizophrenia (Singh et al., 2017), further supporting the NDDs continuum (Owen and O'Donovan, 2017).

Below, we first describe computational studies in NDDs. While we do not dedicate a separate section to cancer-related computational studies, we encourage interested readers to explore recent reviews that focus on cancer and machine learning (Zhang et al., 2023; Ma et al., 2024; Unger and Kather, 2024) in addition to the ones mentioned earlier. We then proceed to the connection between NDDs and cancer. Finally, we provide some computational pipelines for such explorations.

We hypothesized that one major cause for the emergence of NDDs and cancer—two diseases (or conditions) with vastly different clinical features—is the mutation load in the same genes. To validate our hypothesis computationally, firstly we identified mutations that are common between NDDs and cancer, as well as the mutations that are preferentially observed in both diseases. Then, we concentrated on PTEN and PI3Kα, two proteins with NDD and cancer mutations that we are familiar with their structural and dynamic properties.

We investigated the domain distributions and the locations in the 3D structure of common mutations and the mutations occurring only in one of the diseases. We observed that cancer mutations accumulate in critical regions of these proteins that fully thwart PTEN’s tumor-suppressive activities while promoting PI3Kα′s oncogenicity (Jang et al., 2021; Jang et al., 2023a). PTEN’s tumor suppressive functions are also partially tallied in the presence of PTEN mutations in NDD samples. As to PI3Kα mutations that are present in NDD patients, we evaluated their oncogenic potential. In line with our thesis, we observed that they appear incapable of promoting tumorigenesis on their own.

This approach could be extended to all genes harboring common mutations and to mutations that are harbored by either NDDs or cancer. Molecular dynamics (MD) simulations of commonly mutated gene products can point to their potential consequences, although conducting the simulations requires a sufficiently long time (Jang et al., 2023a). Taketomi et al. employed MD simulations to reveal a novel mechanism relating a point mutation in the SPARCL1 gene to the molecular and cellular characteristics associated with ASD (Taketomi et al., 2022). To investigate the importance of microtubule-associated protein 2 (MAP2) phosphorylation in schizophrenia, the authors performed a phosphoproteomic analysis of MAP2 in the primary auditory cortex of schizophrenia and nonpsychiatric control (NPC) patients. Using network analysis, they discovered 18 distinct phosphopeptides and divided them into three modules, each with a unique link to the respective pathology. They also used MD simulations to go further into the most changed phosphorylation location, serine 1782 (pS1782), and discovered that phosphomimetic alteration at this point lowers microtubule interaction (Grubisha et al., 2021). We applied simulations to PTEN to figure out the mechanisms of cancer- and NDD-related mutations and differentiate between them (Jang et al., 2023a), with the aim of a priori identifying the mutations. Our premise was that strong oncoprotein mutational variants tend to visit the active, catalytic state more often than NDD variants do. Another metric for evaluating mutation strength is utilizing variant-associated prediction tools, in our case, MutPred2 (Pejaver et al., 2020). Such tools merge experimental data with predictions and classify the mutations as harmful or benign with a pathogenicity score varying from 0 to 1.

Much data has accumulated through single-cell and spatial omics studies of cancer (Lee et al., 2014; Lei et al., 2021; Zeng et al., 2022; Camps et al., 2023), which are lagging for NDDs. Single-cell sequencing technologies have enabled a groundbreaking resolution of omics where bulk-sequencing data fall short. Single-cell based methods can address areas ranging from revealing the tumor, and the tumor microenvironment (TME) heterogeneity (Pan et al., 2020; Tian et al., 2022), to identify subpopulations and cell states (Figure 3). They also offer a venue for a more comprehensive investigation of oncogenic mechanisms (Pan et al., 2020). Single cell spatial transcriptomics can inform cell-to-phenotype over time mapping by capturing the physical tissue structure, which is of special importance for tissues like the brain (Khodosevich and Sellgren, 2023; Piwecka et al., 2023; Vandereyken et al., 2023). These emerging technologies have unrivaled promise for signatures of tumor growth and progression (Pan et al., 2020; Tian et al., 2022). They allow higher-resolution analysis, which could dramatically illuminate molecular processes that fuel malignancies, aiding cancer surveillance systems and treatment approaches (Piwecka et al., 2023; Vandereyken et al., 2023).

These technologies could also be useful in NDDs, improving the understanding of human brain development. This can help in modeling NDDs and more precisely depicting key developmental disorders (Nowakowski et al., 2017; Moffitt et al., 2018; Fan et al., 2020; Kim et al., 2020; Eze et al., 2021; Khodosevich and Sellgren, 2023). For example, a single-cell based analysis by Skene et al. showed that common schizophrenia GWAS variants, and previously identified schizophrenia-associated genes, mapped to certain groups of brain cell types with a much higher frequency than others, hinting at the distinctive roles of different cells (Skene et al., 2018). Maynard et al. combined single-cell and spatial transcriptomics data of the six-layered human dorsolateral prefrontal cortex (DLPFC) and revealed layer-enriched expression signatures. Addition of schizophrenia and ASD-related genes to the analysis uncovered the presence of differential layer-enriched expression indicating clinical importance (Maynard et al., 2021). Replication time of single-cell RNA sequencing (scRNA-seq) data showed that gene clusters linked to cancer and ASD are restrained in late replication (Nassir et al., 2021). International efforts like The Human Cell Atlas Project, which aims to create a comprehensive molecular map of all human cells to aid studies of physiological states, could help in grasping the origin of cellular dysregulation (Regev et al., 2017).

Due to the heterogeneous nature of cancer and NDDs, it is challenging to draw parallels between two diseases. Exploiting all cancer types and NDD phenotypes implies bigger datasets to learn from; the heterogeneity increases the complexity. Narrowing the comparison to a single phenotype from both diseases may enable discoveries that would have been missed otherwise. This is especially the case for NDDs whose phenotypes overlap, resulting in different phenotypes observed by different researchers. While data availability is crucial in determining phenotypes, especially for NDDs, recent epidemiological cohorts provided convenient starting points. In a cohort study of 8,438 patients with autism, Chiang et al. observed an increased risk of genitourinary and ovarian cancers (Chiang et al., 2015). Another cohort study reported a higher risk of breast cancer in BD patients and their healthy siblings in comparison to the control (Chen et al., 2022), again in line with our thesis of the possible connection between cancer and NDDs. To uncover the underlying mechanisms, the phenotypes to be studied can be selected based on already reported connections, such as thyroid cancer and autism, or breast cancer (Figure 1A) and BD.

Computations can analyze the rapidly increasing volume of data, identify trends, and derive correlations at a scale that experiments are unable to do. By discovering causal relationships, they can also identify the most probable events, processes, states, and objects. In our case, such discoveries are extremely consequential as they can identify the likelihood of NDDs early in life even prior to emergence of the debilitating phenotype. No less important, they can project the likelihood of cancer, assisting with early pharmacology. Cancer onset is associated with abnormal cell proliferation (Brunner and Finley, 2023; de Visser and Joyce, 2023), while NDDs are mostly related to dysregulated differentiation (Khodosevich and Sellgren, 2023). In terms of cell cycle signaling (Demeter et al., 2022), they differ in signaling strength and duration (Yavuz et al., 2023a). Under physiologic conditions, mitogen-promoted strong signaling bursts over short duration are associated with cell proliferation. Weaker, extended mitogen-promoted signaling is associated with cell differentiation. When disease-related, due to activating mutations or overexpression, signaling is constitutive, thus always extended. When strong, the outcome is likely to be proliferation in cancer. The paramount factor is the population of proteins in their active states. Both strong, mitogen-promoted bursts, and mutation-elicited, lead to a larger population in the active state (Nussinov et al., 2022a). Higher expression level due to, e.g., gene duplication, or dysregulation, lead to the same outcome.

Cancer and NDDs are not the only diseases associated with cell cycle and signaling strength. Neurodegenerative diseases in old age, cardiovascular disease (Wiman and Zhivotovsky, 2017), vascular proliferative disorders (Sriram and Patterson, 2001), diseases with prolonged mitosis (Levine and Holland, 2018), possibly the outcome of low translation rates, and more, can all be related to the cell cycle.

Here, we discussed computational approaches, including AI, ML, and MD simulations, as well as others, such as networks construction, and their findings. Computational observations based on ‘large enough’, high quality data, can be powerful. We took up the linkage between NDDs and cancer, and emerging strategies for pursuing them. Our comprehensive searches of the literature uncovered an abundance of experimental and clinical studies. However, unfortunately, we failed to identify computational approaches, especially at the detailed molecular level addressing questions such as those that we raised above, leading us to describe the ones that we have adopted and possible extensions. We hope that these open the door to future studies of this emerging innovative and enthralling computational discipline.

Taken together, our broad review underscores the multi-scale aspect of this type of reasoning, rather than focus on a specific organ system or disease, with the examples that are used demonstrating what new understanding can be gained.