Artificial intelligence-driven decoding of trajectories of cancer cell plasticity: A paradigm shift in oncological assessment and treatment

Abstract

Cellular plasticity characterized by changes in the fate and identity of cells, substantially fuels the process of tumorigenesis by enhancing growth, metastases, and resistance to therapy under selective pressures. PubMed, Scopus, and Web of Science databases were comprehensively searched for the full papers published during 2013-2026 using the keywords such as artificial intelligence (AI), cellular plasticity, genomic, (epi)genomic, transcriptomic factors, oncological assessment, and therapies. The present mini-review discusses the applications of AI-driven deep learning and machine learning models integrated with advanced next-generation sequencing and single-cell RNA analysis in elucidating the changes in the trajectories of plasticity in tumor ecosystem. It facilitates the deciphering of the molecular landscape associated with normal, metaplastic, and pre-malignant tissues in animal models and human clinical specimens. This would help in establishing new predictive signatures for risk stratification of tumors and predicting the response to drugs. AI plays an important role in drug development, boosts three-dimensional structure of protein based drug discovery, analyzes the pharmacokinetic properties of drugs, and widens the scope of repurposed drugs. Recent trends foresee its application in providing tailored diagnosis, risk assessments, accelerating drug discovery, and individualizing treatment regimens.

Key Words: Artificial intelligence; Cellular plasticity; Diagnostics and therapeutics; (Epi)genome and transcriptomic landscape; Tumor heterogeneity

Core Tip: Dynamic switching between cellular states generates tumoral heterogeneity as explained by stochastic and hierarchy models, and is driven by genetic and epigenetic alterations in a tumor ecosystem. Recent advancements in the applications of artificial intelligence-driven deep learning and machine learning models integrated with next-generation sequencing and single-cell RNA analysis lead to the elucidation of the changes in the trajectories of cellular plasticity. artificial intelligence-powered approaches decipher the (epi)genomic and transcriptomic landscape associated with normal, metaplastic, and pre-malignant tissues and thus help in establishing the new predictive signatures for tailored diagnosis, risk assessments, accelerating drug discovery, and individualizing treatment regimens.

INTRODUCTION

Tumor recurrences, its progression to clinically detectable metastatic disease, and drug resistances are the prime cause for increased global cancer burden over the recent decades. Cellular plasticity is characterized by change in the fate and identity of the cells. It is essential for embryonic differentiation, restoring homeostasis following tissue damage, inflammation and senescence. It also substantially fuels the process of tumorigenesis by enhancing the growth, metastases, and resistance to therapy under selective pressures. Phenomenon including dedifferentiation, transdifferentiation, and the epithelial-to-mesenchymal transition (EMT)/mesenchymal-to-epithelial transition (MET) reprogram the cells to attain plasticity in response to intrinsic and extrinsic factors. Dedifferentiation refers to the phenomenon where differentiated cell turns back into undifferentiated cell within same lineage. Transdifferentiation is the process where differentiated cell turns into another differentiated cell lineage. The EMT is characterized by the loss of properties of epithelial (E) cells like apicobasal polarity, cell-cell junctions and cells acquire mesenchymal phenotype. Mesenchymal (M) cells when acquire epithelial phenotype undergo MET. The EMT/MET represents a phenotypic spectrum which is characterized by different degrees of epithelial and mesenchymal features. Among the various cellular plasticity mechanisms, the EMT/MET and specifically, partial-EMT state, facilitates collective migration of tumor cells. These mechanisms confer carcinoma cells with a higher degree of metastatic competence as compared to complete the EMT programs[1-3].

The dynamic switching between cellular state generates tumoral heterogeneity, and is driven by genetic and epigenetic alterations. Stochastic and hierarchy models of tumorigenesis help us to understand the concept of tumoral heterogeneity. It is majorly responsible for therapy failure, disease relapse and short survival of patients. Unique driver mutations in any tumor cell (cell-of-origin) contributes to tumoral heterogeneity by producing genetically different subclones (stochastic model). However, oncogenic hit turns normal adult stem and progenitor cells into cancer stem cells (CSCs) and cancer progenitor cells respectively (hierarchy model). It thereby changes the cancer landscape (Figure 1). It allows the cancer cells to clonally evolve when subjected to selective pressures, invoke anti-apoptotic pathways, reprogram angiogenic and immune recognition[4].

Figure 1 Schematic representation of hierarchy model and stochastic model of tumorigenesis. This figure was prepared on power point and Canva. Two foundational models of tumorigenesis explain the origin and growth of tumors as well as maintenance of cellular diversity. Hierarchy model proposes the ability of only specialized subpopulation of cells to drive the tumor development. The stochastic model argues that any cell has an equal and random chance to promote tumor development. CSC: Cancer stem cell.

Tumor microenvironment (TME) provides a favorable metabolic environment to support the cellular plasticity and tumoral heterogeneity. Volatile TME comprises of: (1) Diverse extracellular matrix (ECM) components; (2) Stroma containing networks of cytokines and growth factors; (3) Infiltrating endothelial, hematopoietic and perivascular cells or their progenitors; (4) Immune cells; and (5) Cancer-associated fibroblasts (CAFs)[5]. Metabolic changes such as hypoxia and nutrient fluctuations further aggravate the complexity during crosstalk between tumor cells and TME in a tumor ecosystem. The complex TME increases the tumor robustness by allowing the tumor cells to undergo genetic and epigenetic/transcriptional fluctuations. These alterations including activation of oncogenes, loss-of-function mutations of tumor suppressors and aberrant expressions of transcription factors in chromatin landscapes result in dynamic changes in gene expressions. Such structural changes in the (epi)genome may have different effects in different cells of origin[6]. Oncogenic stimulus may drive the tumor cells to change their phenotype, acquire cellular plasticity and expand the tumor-initiating capacities in a context-dependent manner. Regardless of the precise mechanism, cellular plasticity consistently provides a renewable source of tumor cells with high malignant potential and imposes challenge to therapeutic approaches[7,8].

Advancements in next-generation sequencing (NGS) and single-cell RNA (scRNA) analysis, combined with artificial intelligence (AI) technologies may help in elucidating the changes in the molecular landscape. Thereby driving the tumor cells to acquire cellular plasticity in the EMT/MET dependent and independent manner in a tumor ecosystem. Further careful examination of the chromatin states associated with normal, metaplastic, and pre-malignant tissues in animal models and human clinical specimens would help in designing new predictive signatures. These predictive signatures may guide risk stratification of tumors with accuracy, and predict response to therapies. The present mini-review discusses the applications of AI driven tools in elucidating the changes in the trajectories of cellular plasticity and tumoral heterogeneity during their crosstalk with TME. Integrating advanced sequencing methods with AI driven deep learning (DL), machine learning (ML), and existing drug discovery techniques could help in faster and accurate assessment of oncological outcome. Besides, it may also help in the development of effective new drug compounds or molecules, thereby revolutionizing the personalized cancer therapeutic approaches[9].

MECHANISTIC INSIGHTS OF THE DYNAMICS OF CELLULAR PLASTICITY AND TUMORAL HETEROGENEITY

Cancer cell plasticity and tumor heterogeneity in a tumor ecosystem confers tumors not only the notorious ability to evade immune and therapy responses, but also underpin their aggressive metastatic potential. Digging into the molecular insights of the cancer cell dynamism would help in understanding the multifaceted nature of cancer cells. This would be instrumental in delving into clinical applications, therapeutics and disease management.

Other than embryonic development and tissue repair processes, the EMT is implicated in pathophysiological conditions like organ fibrosis and tumor development. TME cues such as hypoxia, interactions with the ECM, and release of growth factors and cytokines facilitate the activation of key signaling cascades. These signaling cascades including [hedgehog (Hh), Notch, Wnt, receptor tyrosine kinases, and transforming growth factor beta (TGFβ)] and specific transcription factors, and altered gene expression patterns ultimately drive the EMT (Figure 2). The EMT is characterized by loss of cell-cell adhesion, loss of apicobasal polarity, cytoskeletal reorganization, disassembly of basement membranes and a shift from E to M phenotype. E phenotype is marked by reduced levels of epithelial proteins including E-cadherin, claudins, desmoplakin, occludin, specific cytokeratin intermediate filament proteins, epithelial cell adhesion molecule, mucin1, zonula occludens-1, syndecan-1 and laminin. M phenotype is associated with increased levels of vimentin, N-cadherin, fibronectin, α-smooth muscle actin, S100 calcium-binding protein A4 and matrix metalloproteinases (MMPs) (MMP-2 and MMP-9). Bidirectional conversions of tumor cells among E phenotype, M phenotype, and partial/hybrid E/M phenotypes via the EMT and MET contribute to tumor heterogeneity and cellular/epithelial plasticity[10]. It is the best known cause of tumor cell plasticity and it contributes to stemness, enhanced proliferative capacity and tumorigenic potential. Understanding the mechanisms of tumor plasticity may help in designing effective strategies to combat the lethal aspects of cancer that is metastasis and therapy resistance[6].

Figure 2 Mechanistic insights of the processes which regulate cancer cell plasticity in tumor microenvironment. This figure was prepared on power point and Canva. Environmental cues in a tumor ecosystem such as hypoxia and nutrients fluctuations or injury and oncogenic insults activate signaling pathways such as Notch, Wnt, mitogen-activated protein kinase, receptor tyrosine kinase, hedgehog and transforming growth factor beta. Dynamic tumor microenvironment results in changes in genetic, epigenetic, transcriptional and translational landscape enabling the cells to undergo epithelial-mesenchymal transition, mesenchymal-epithelial transition and partial/hybrid transition to promote cancer cell plasticity, invasion and metastatic potential and therapy resistance. AKT: Protein kinase B; CAFs: Cancer associated fibroblasts; ECM: Extracellular matrix; E/M: Partial/hybrid; EMT: Epithelial-mesenchymal transition; EMT-ATFs: Epithelial-to-mesenchymal transition-associated transcription factors; GLI1/2: Glioma-associated oncogene homolog 1 and glioma-associated oncogene homolog 2; JNK: Jun N-terminal kinase; MAPK: Mitogen-activated protein kinase; MEK1: Mitogen-activated protein kinase kinase 1; MET: Mesenchymal-epithelial transition; mTOR: Mammalian target of rapamycin, NICD: Notch intracellular domain; PI3K: Phosphoinositide 3-kinase; RTK: Receptor tyrosine kinase; Smad: Suppressor of mother against decapentaplegic; TGFβ: Transforming growth factor beta.

Activated key signaling cascades trigger the expressions of the EMT-activating transcription factors (the EMT-ATFs). These EMT-ATFs include: (1) Snail family of zinc-finger transcription factors (Snail, Slug and Smuc); (2) Two-handed zinc-finger factors of d-crystallin/E2 box factor family proteins [zinc-finger E-box binding (ZEB) 1 homeobox and suppressor of mother against decapentaplegic (Smad)-interacting protein 1/ZEB2]; (3) Basic helix-loop-helix factors (Twist1 and Twist2); and (4) E12/E47 and T-box transcription factor 3. The EMT-ATFs regulate the EMT program by binding to the promoter regions of epithelial genes involved in cell-cell adhesion, and repressing their expressions, while enhancing the expressions of mesenchymal proteins[11,12].

TGFβ signaling

TGFβ although induces apoptosis but during later stages of tumor development, Smad dependent and independent TGFβ receptor signaling pathways are activated. Type I receptor and type II receptor are serine-threonine kinase receptors and activate Smad2 and Smad3 upon binding to TGFβ ligand. Smad4 forms complex with activated Smad2/3. Translocation of it into the nucleus results in the regulation of the expression of target genes including the EMT-ATFs. Non-canonical TGFβ signaling/Smad independent pathway activated by TGFβ receptors is mediated by phosphatidylinositol 3 kinase (PI3K), protein kinase B, mitogen-activated protein kinase (MAPK) and small guanosine triphosphatases of the Rho family. It facilitates the EMT switch, cell migration and invasion of cancer cells. Process of the EMT mediated by TGF-β signaling confers stem cell properties and leads to the acquisition of migratory and invasive phenotype and resistance to anoikis[13]. Altered TGFβ signaling can manifest cancer via: (1) Expression and activation of TGFβ ligands; (2) Expression, post-translational modification, and mutations in TGFβ receptors and Smads; and (3) Perturbations in Smad co-regulators[14].

Notch signaling

Notch signaling between adjacent cells is initiated by binding of Notch receptors (Notch1-4) to ligands (Jagged1, 2 and delta like 1, 3, 4). This is followed by release of Notch intracellular domain upon intramembrane cleavage of Notch receptor by γ-secretase. It is then translocated to the nucleus and forms complex with: (1) Transcription factor, C-promoter binding factor 1/suppressor of hairless in Drosophila/Lin-12 and glp-1 in C. elegans; (2) Recombination signal binding protein for immunoglobulin kappa J region; (3) Mastermind like 1; and (4) Histone acetyltransferase, E1A-associated protein p300/cyclic-adenosine monophosphate response element-binding protein, to regulate the expressions of target genes. These target genes include cell cycle regulator p21, myc, hairy/enhancer of split and the hairy/enhancer of split related repressor (HERP, HRT and HEY) families. TGFβ is shown to increase the Notch activity through Smad3, upregulate Jagged1, HEY1 and Slug expression, and suppress E-cadherin. Expressions of the delta like 1, 3 and 4 ligands enhanced by Wnt signaling pathway promote tumorigenic phenotype. Notch regulates increased expression of lysyl oxidase by recruiting hypoxia inducible factor-1 alpha to lysyl oxidase promoter and stabilizes Snail, induces EMT and tumor development[11-13,15]. Notch signaling was examined to trigger Wnt signaling to promote CSC proliferation. Further activation of Jagged-1 and Notch-2 by Wnt generates a regulatory crosstalk between Notch and Wnt signaling cascades. However, Wnt was found to inhibit Notch signaling in the presence of hypoxia-inducible factor alpha in glioblastoma stem like cells, leading to decrease in stem like properties[16].

Wnt signaling

Canonical Wnt signaling is initiated by binding of Wnt1 and Wnt3a ligands to frizzled, a seven transmembrane domain receptor, and the lipoprotein receptor related protein complexes. Activated pathway inhibits the destruction complex which is composed of adenomatous polyposis coli, glycogen synthase kinase 3β, Axin and casein kinase-1α, and stabilizes β-catenin. Nuclear translocated β-catenin interacts with lymphoid enhancer factor/T cell factor complex and activates the transcription of target genes involved in cell proliferation, EMT and tumor development. Canonical Wnt signaling causes expansion of rapidly cycling CSCs and modulates immune surveillance and immune tolerance[17]. Non-canonical Wnt signaling activated by specific Wnt ligands via binding to frizzled receptors alongside co-receptors such as receptor tyrosine kinase-like orphan receptor 1/2 or receptor like tyrosine kinase. It regulates cell polarity and manages cytoskeleton rearrangement by activating Rho-kinase, c-Jun N-terminal kinase, and Ras-related C3 botulinum toxin substrate 1. Non-canonical Wnt signaling supports maintenance of slow cycling, quiescent or dormant CSCs and promotes EMT via crosstalk with TGFβ signaling. It coordinates the functions of CAFs, endothelial cells and immune cells in TME, and leads to the fine-tuning of stemness in melanoma, breast, ovary, pancreatic, colorectal, gastric and prostate, lung and uterine cancers, bulk-tumor expansion and invasion/metastasis[17]. Crosstalk between Axin and Smad3 activates TGFβ pathway. Their interaction enhances the Smad3 protein by TGFβRI, which results in separation of Smad3 from both Axin and TGFβRI. Studies show that inhibition of one pathway impacts the induction of other pathway[18]. Wnt signaling is promoted by TGFβ by inducing Wnt ligands and low-density lipoprotein receptor-related protein. TGFβ forms a co-ordination complex with lymphoid enhancer factor 1/T cell factor with the help of Smads. Nevertheless, TGFβ suppresses signal transducer and activator of transcription 3, and subsequently suppresses Wnt3a expression, resulting in prostate cancer regression[19].

Hh signaling

Hh signaling regulates the EMT, cell polarity, and maintains stem-cell characteristics. Binding of Hh ligands (Sonic, Desert, and Indian) to a twelve pass transmembrane protein, patched, derepresses smoothened. Its translocation to primary cilium, internalization and activation regulates the activation of zinc-finger transcription factors (glioma-associated oncogene homolog [GLI] -1, -2 and -3] and transcription of GLI target genes. Sonic Hh-GLI1 signals via complex signaling network (TGFβ, Ras, PI3K/protein kinase B, Wnt), growth factors, integrins, transmembrane 4 superfamily and S100A4 and thus promotes the EMT and tumor development[11-13,15]. Hh signaling can also activate Notch signaling through Jagged-1 upregulation. Their crosstalk can result in tumor progression and stemness retention. The interaction among the Hh, Wnt, and Notch is studied to promote proliferation of breast CSC[16]. However, downregulation of GLI1 through Hes1 may result in inhibition of Hh by Notch signaling[20].

Tumor necrosis factor alpha signaling

Tumor necrosis factor alpha (TNF-α), a tumor-promoting pleiotropic cytokine fuels tumor cell growth, invasion, and metastasis. Signaling is initiated by two distinct cell TNF receptors (TNFRs) 1 and 2. TNFR1-associated death domain protein recruits TNFR-associated factor and receptor-interacting protein. This is followed by the recruitment of the IκB kinase complex and its activation in a receptor-interacting protein-dependent manner. Phosphorylation of inhibitory protein IκB by IκB kinase complex promotes its rapid ubiquitination and proteasome mediated degradation and releases nuclear factor kappa B (NF-κB). Nuclear translocation of free NF-κB increases the transcription of EMT-ATFs including Snail, Slug, Twist, and zinc finger E-box binding homeobox 1 (ZEB1)/ZEB2, mesenchymal protein (vimentin) and MMPs, such as MMP-2 and MMP-9[11].

MAPK signaling

Binding of growth factors to the extracellular receptor domain activates receptor tyrosine kinases and their downstream signaling effectors, such as MAPK or PI3K and activate EMT and cell proliferation. MAPKs constitute a family of serine/threonine protein kinases and the most extensively studied ones belong to the subfamilies. These subfamilies include extracellular signal-regulated kinases 1 and 2, p38 isoforms (α, β, γ, and δ), c-Jun N-terminal kinases 1-3, and the extracellular signal-regulated kinase 5 pathway. Phosphorylation and activation of MAPK is observed to inhibit tumor cell death and promote resistance to anti-cancer therapies[21]. Extracellular signal-regulated kinase-MAPK pathway inhibition is examined to restore E-cadherin expression in cells with moderate levels of Ras signaling[22].

Epigenetic alterations including chromatin remodeling mediated by SWItch/Sucrose non-fermentable complex, and covalent histone modifications (methylation, demethylation, acetylation, and insulation) impact the chromatin landscape. Epigenetic modifications may result in bivalent chromatin state defined by active and repressive markers close to specific genes, thereby lead to improper cellular reprogramming or differentiation arrest. These characteristics help the cells to preserve cancer stemness properties, proliferative, migratory and invasive potential of tumor cells. Genetic abnormalities/mutations including chromosomal abnormalities (aneuploidy), structural rearrangements (translocations, amplifications, deletions, insertions), genomic instability, faulty DNA repair mechanism, and loss or gain-of-function mutations potentially change the behavior of cells. These alterations promote the flexibility/epithelial plasticity of cancer cells, improve stemness related features and increase the migratory, invasive and tumorigenic potential[23].

Subsequent section discusses the transformative potential offered by NGS and single cell technologies when coupled with AI in understanding the dynamics of cancer cell plasticity, oncological assessment and designing better treatment strategies.

AI-DRIVEN DECODING OF TRAJECTORIES OF CANCER CELL PLASTICITY

Advancements within the AI models, its ML and DL subtypes and speed of NGS methods make it possible to identify the molecular landscape of transient cellular states and the changes in TME that underlie plasticity even at single cell resolution (Figure 3).

Figure 3 Artificial intelligence and its applications in cancer diagnostics, prognostics and therapeutics. This figure was prepared on power point and Canva. Machine learning and Deep learning models are two important subsets of artificial intelligence that enable the computers to learn and interpret the experimental data on cancer cell plasticity dynamics to provide tailored diagnosis, risk assessment, accelerating drug discovery and individualizing treatment regimens. AI: Artificial intelligence.

Processing of big datasets using traditional processing methods to convert raw reads into clinically actionable form is complicated, time-consuming and irreproducible. Early traditional processing bioinformatics-based methods mostly used pathway enrichment analysis, transcriptomics, survival data, genomics to understand the tumor characteristics. However, these methods masked the activities of individual cells making it difficult to understand cell transitions[24,25]. Besides, there is a high risk of insufficient, largely sized data with noise and inconsistent values obstructing the interpretation of various molecular dynamics[25,26].

AI-based ML method is one of the modern advancements of industrial revolution which, by default, uses the self-learning techniques to advance its system using its past stored data[27]. It provides a platform to test the assumptions made based on the layouts of the new dataset using statistical approaches, making it an effective tool to analyze various types of complicated data collections. Thus can be employed for diagnostic and prognostic assessment[28]. ML uses supervised learning, unsupervised learning, reinforcement learning and DL computational methods in health care imaging, patient categorization and diagnostics; drug designing and therapeutics[29]. Supervised learning utilizes the saved data to predict the outcome of disease, its recurrence, progression and survival status along with therapeutic responses[30]. Under supervised method, random forests and support vector machines are highly preferred algorithms and can be used for detailed classification and regression of NGS data. However, unsupervised method is recommended to study the trends in unannotated NGS data[31]. DL methods with improved hardware, software and parallelization of algorithms overcome the challenges of scalability and dimensionality of single cell data. It layers the algorithms into artificial neural networks to analyze raw sequence data by aligning it with the reference genome. A neural network database uses three different layers: (1) An input layer that incorporates unprocessed data; (2) Hidden layer that modifies the data based on biases, coefficients and induction factors; and (3) Output layer that presents the final anticipation or characterization[32]. It removes the duplications, deletions and insertions, does the base realignment and recalibration and thereby, improves the quality of data[33].

AI-driven diagnosis

Geneformer, a transfer learning dependent DL tool, recognizes the gene regulatory framework from millions of single cell transcriptomes. By using self-supervised method, it identifies transcriptional activities and later uses it for assessment of various cellular states within small sample numbers[34]. Transcriptional inference based on gene expression and regulatory data, a bayesian model dependent AI tool, analyses and processes the information about the connection between the active and suppressed genes to identify the prominent regulatory interactions. The recovered data are then used to analyze the functions and alterations of transcription factors in regulatory frameworks[35]. TRENDY is an AI dependent gene regulatory network tool that uses transformer model to retrieve gene-regulatory interconnections from expressed genes[36].

ML analysis on pancreatic ductal adenocarcinoma-derived phase images was done to detect changes in tumor cell differentiation and intratumoral heterogeneity of distinct pancreatic ductal adenocarcinoma subtypes upon EMT induction and under treatment-imposed pressure in murine and patient model systems[37]. Quantitative experiments combined with AI, examined the rapid transitions in cell phenotypes across diverse ECM conditions and phenotype-dependent motility of invading tumor spheroids composed of breast cancer cells[38]. Study by Calderaro et al[39], analyzed the series of 405 patients with combined hepatocellular-cholangiocarcinomas (CCA), a rare variant of liver cancer, to reclassify the combined hepatocellular-CCA tumors into hepatocellular carcinoma or intrahepatic CCA based on the histopathology images using DL models[39]. Hepatocellular carcinoma and intrahepatic CCA represent two clinically distinct liver tumor types with completely different risk factors, genetic and molecular characteristics and treatment strategies[39]. Experimental investigation in association with ML identified cluster of differentiation 81 (CD81), a tetraspanin transmembrane protein enriched in extracellular vesicles as a marker/driver of breast cancer stemness and metastasis. Study also examined its clinical significance in association with the outcomes of triple negative breast cancer patients[40]. The study further demonstrated the interaction of CD81 with CD44 to promote tumor cell cluster formation and lung metastasis in triple negative breast cancer patients using protein structure modeling and interface prediction-guided mutagenesis[40]. Locally embedded anomaly pattern hunter is an unsupervised ML algorithm to identify functional phenotypes along a continuum. When combined with pointwise mutual information and network biology, it resulted in exploring the outcome-associated microdomains as distinct spatial configurations of heterogeneous functional phenotypes[41]. Strong association of image dataset of immunofluorescence-based 51 biomarkers with CSC maintenance, immunosuppression and recurrence phenotype was noted in colorectal carcinoma primary tumors (n = 213)[41].

AI agent for high-optimization and precision medicine focused on Wnt is an AI model and specifically designed to check the status of Wnt signaling dysregulation in colorectal cancer. It uses large language models, prompts-code engines and predefined analysis workflow to process mutational frequency, survival assessment and perform clinical and socio-economical characterization of the variables. This AI model was used to confirm the result from prior studies performed on high risk colorectal cancer patients. The model validated the outcome from previous studies by showing consistency with better survival chances in Wnt-dysregulated high risk colorectal cancer patients and enhanced ring finger 43 mutations in Hispanic/Latino patients. As an advantage over traditional method, this model demonstrated the association of ring finger protein 43, adenomatous polyposis coli and Axin alterations with the therapeutic outcome, tumor localization, survival and patient characteristics making it a promising model for Wnt dyregulation analysis[42]. Similarly, AI-agent for high-optimization and precision medicine-PI3K was created to check the PI3K signaling dysregulation in colorectal cancer patients. AI-agent for high-optimization and precision medicine-PI3K accurately demonstrated the known associations and identified novel outcomes including poorer survival in colon vs rectal tumors with PI3K alterations, enrichment of inositol polyphosphate-4-phosphatase type II mutations in Hispanic/Latino patients and better survival outcomes in FOLFIRI-treated patients with high tumor mutational burden[43].

AI-driven prognosis

An AI-driven spatial cell omics strategy incorporating histopathology, ML and multiplex microscopy was used to interpret the 43 cellular phenotypes of TME in non-small cell lung cancer patients[44]. Foundational AI and radiomics based models were demonstrated as prominent prediction tools in the identification of significant signatures of outcomes in non-small cell lung cancer patients[45]. An eXtreme Gradient Boosting, highly efficient ML library and algorithm, in combination with sequencing methods identified common epigenetic pattern of enhancer activation characterized by histone H3 lysine 27 acetylation to predict transcriptomic expression pattern in glioblastoma stem cells across 11 patients[46].

AI-driven single cell trajectory analysis

Autoencoder, a DL method, compresses and reconstructs the manifold representations of data with minimal information loss. DL methods including sparse autoencoder, stochastic cell-level variational inference, and variational autoencoder are used for single cell data analysis. These methods recognize the patterns in data and interpret it with high biological variance[47-49]. A two-stage variational autoencoder refines the latent representation of data and is used for image reconstruction suitable for biological inference. Traditional DL method, archetypal analysis, captures the complex non-linear relationships and describes the spectrum of cell states[50]. This method can be trained to identify particular cancer cell state or transition state that share the same archetype. Hence, labelling of cancer cells can be avoided and cells can be placed within a continuous landscape. This approach helps the analysis of genomic programs contributing to cancer cell state plasticity and tumor heterogeneity. Changes in the transcriptomic landscape during cellular plasticity are determined by pseudotime trajectory method. The raw data generated during scRNA sequencing as input are used by computational algorithms for mapping the cells on continuous trajectory for their relative positions and pseudotime[51-53]. scVAG, a DL tool is used to obtain accurate biological structure and interpret interaction between gene patterns. It uses variational autoencoder for dimensional reduction of single cell datasets and then with the help of graph attention autoencoder, it generates graph based interaction between the datasets for further clustering[54]. scCompressSA tool identifies gene-gene relationship using self-attention model and then incorporates the structural pattern of the same to improve the clustering ability[55]. A heterogenous graph learning model stKeep helps in detection of cellular plasticity within TME and gene-gene relationships using cellular and spatial transcriptomic information which were further verified in clinical cohorts[56]. Advancements in computational optimal transport like Waddington optimal transport quantifies the minimum cost mapping during induced pluripotent stem cell reprogramming[57]. ScEGOT, comprehensive framework for single-cell trajectory inference, is built on Gaussian mixture optimal transport. It is presented as a generative model with high interpretability and low computational cost. It accurately identified human primordial germ cell-like cell population and bifurcation time of segregation[58]. Identification of divergences in malignant cell lineages among different cancer cell subtypes could be an important clinical tool to accurately predict the tumor specific features like progression and treatment resistance[59]. A novel ML architecture analyzes the transcriptome profiles of 33 types of RNA from The Cancer Genome Atlas database and classifies the site-specific metastases in 16 cancer types that metastasize to 12 different locations[60]. Various single cell trajectory inference algorithms were employed to understand the pediatric glioblastoma cell-fate dynamics and tumor-immune interactions. Clinically relevant targets were identified as guanine-adenine-thymine-adenine transcription factor 2, protein tyrosine phosphatase receptor-type Z polypeptide 1, tumor protein translationally-controlled 1, mitochondrially encoded 16S rRNA-like 1/2, oligodendrocyte transcription factor 1/2, sex determining region-box transcription factor 11, phenylalanine-any amino acid-tyrosine-aspartic acid domain-containing ion transport regulator 6, seizure related 6 homolog like, platelet-derived growth factor receptor alpha, epidermal growth factor receptor, S100 calcium binding protein B, Wnt, TNF-α, and NF-κB[61]. Multi-omics identified the origin of lung adenocarcinoma in alveolar type 2 cells, while the lineage plasticity via sex determining region-box transcription factor 2/Wnt/Yes-associated protein 1 pathways was observed to be responsible for aggressive subtypes. AI driven TME modelling identified immune subsets (CXCL13⁺CD8⁺T cells, M1/M2 macrophages) and antigen-presenting CAFs as therapeutic targets by targeting lineage plasticity. Thereby, reprogramming the TME, modulating the microbiome and overcoming the immune checkpoint blockade/tyrosine kinase inhibitor resistance[62].

AI-driven development of anticancer therapeutics

Characterization of the molecular dynamics of cellular plasticity and its landscape may offer promising strategies for designing effective targeting therapies. Incorporation of AI in traditional methods can facilitate the development of allosteric drugs which can restore sensitivity for resistant markers, proposing more efficient chemotherapies specialized for various cancer plasticity states[9]. AI integrated with clustered regularly interspaced short palindromic repeats based technology plays an instrumental role in preclinical drug development by generating in silico-designed broad spectrum of molecules and analogs with the anticipation of resistance mechanisms. AlphaFold2, an AI system utilizes DL and boosts the structure-based drug discovery by accurately identifying the distances between amino acids and the angles of chemical bonds in proteins and predicting the 3D structures of proteins from amino acid sequences. Certain limitations associated with AlphaFold2 include incorrect prediction of protein structures with multiple domains, e.g., transmembrane receptors; inability to predict the change in the protein structures during interaction with targets; and inability to anticipate the effects of mutations on protein structures[63,64].

Advancements in vector machine models make it possible to predict the pharmacokinetic properties, blood-brain permeability and intestinal absorption of drugs. It helps in widening the use of conventional drugs as repurpose drugs by anticipating their probable mechanisms beyond their existing medical indications[65]. The efficacy of drug can be calculated in terms of drug score on the basis of cell type proportion; significances of reversing pattern of differential gene expression during drug treatment in every selected cell cluster; and ratio of significantly deregulated genes which can be reversed upon drug treatment in each selected cell cluster[66].

AI sets up the in silico/virtual clinical trials based on the modelization of cumulative clinical experience and biological data generated upon the use of previously developed molecules that belong to the same category of drugs in question. Different randomized clinical trial experimental designs were compared using the in silico simulations for accurate estimation of treatment effects based on dose-effect interactions for quantitative or qualitative outcomes. One thousand clinical trials were simulated for each design[67]. In silico simulation could strengthen the selection of experimental design and treatment effects for future clinical trials. Drug Response prediction Utilizing Genomic features Screening, a DL tool can anticipate the cellular reactions developed in human oncogenic cell lines using the expression pattern of genes to interpret resistance and identify synergic therapies[68]. PDGrapher, a graph neural network based tool for the combinatorial prediction of therapeutic perturbations, is designed to identify sets of genes or molecular/therapeutic targets (perturbagens) to reverse disease phenotypes. It holds the ability to forecast the combinations of therapeutic modalities that can counteract the reaction provided by the diseased states[69]. ScTherapy, a ML based model was used to prioritize multi-targeting treatment options using single cell transcriptomic profiles of individual patients with hematological or solid malignancies[70]. The study exhibited selective efficacy or synergy in 96% of the multi-targeting treatments whereas 83% demonstrated low toxicity to normal cells. Study on pan-cancer analysis across five cancer types validated the sharing of 25% of predicted treatments among the patients of same tumor type, whereas 19% of the treatments were observed to be patient-specific[70]. The area under the curve value of 0.71 observed upon applying DeLong test demonstrated the superior predictive performance of scTherapy over the other two models namely scDrug and BeyondCell[70]. PERsonalized single-cell expression-based planning for treatments in oncology predicts the personalized therapeutic response in patients using single cell tumor transcriptomic profile. This tool showed success in predicting responses to targeted therapies in cultured cells, patient-tumor-derived primary cells as well as in two clinical trials for multiple myeloma and breast cancer[71]. Besides, PERsonalized single-cell expression-based planning for treatments in oncology is examined to successfully capture the resistance development in lung cancer patients treated with tyrosine kinase inhibitors[71]. SYNDEEP, a DL tool utilizes physicochemical and genomics features and predicts synergistic drug combination responses with 92.21% prediction accuracy and 97.32% area under the curve metrics in tenfold cross validation[72]. AI-based pharmacogenomics models like DeepDRA, MOICVAE, superFELT and various other multi-omics dependent tools capture indicators of drug response and distinguish vulnerable phenotypes from resistant ones[73]. AI databases developed by utilizing NGS pharmacogenomics information and therapeutic responses in clinical setting can accurately anticipate drug responsiveness as well as resistance in various cancer types. This encourages the adoption of AI-based approaches in tailoring the optimal treatment for individual cancer patients[74,75].

AI offers the advantage of flexibility to increase the volume of data, significant reduction in the estimated time period and cost of therapeutic trials along with avoidance of duplicacy in anticancer drug development programs. Nevertheless, underfitting, overfitting and lack of clear interpretation of data due to complexity in the models are few limitations associated with it.

CHALLENGES AND FUTURE PERSPECTIVES

Integration of AI with NGS and sc-omics provides opportunities to overcome the challenges in understanding the multifaceted nature of cellular states. AI-powered platform presents a robust paradigm shift in oncological assessment, predicting drug response and designing personalized treatment strategies. It not only reduces the reading time frames but also the risk of misdiagnosis. It can evaluate vast and diverse data related to thousands of the past patient medical records in fraction of time to predict survival probabilities and risk of cancer recurrence/spread. One of the foremost advantages of AI powered approaches is to effectively design the drug response that matches to the patient’s unique tumor genetics. Thereby reduces the trial-and-error period of testing multiple drug treatments and side-affects associated with them.

Despite advantages, number of shortcomings including patient information privacy vulnerabilities, unsupervised learning applications, poorly regulated training period of system, data size, insufficient regulation guidelines, limited human-software interaction for clinical assistance, deployment of software solely based on the judgement of the developer, algorithmic bias, lack of transparency, lack of external validation, reproducibility and ethical concerns limits its complete clinical integration.

AI-powered models trained on static datasets may struggle to predict the dynamic biological shifts of cancer cells. Future AI-integrated approaches need upgrades which include better computational strategies for integrating diverse data types, better algorithms, multifactor prediction tools and its authentication in various ethnic groups, development of multimodal frameworks and dynamic trajectory models based on integrating transcriptomic, epigenomic, spatial transcriptomic, and imaging data. This will enable the researcher to decipher the intricate interplay between intrinsic and extrinsic factors and map the complex cellular pathways that affects the dynamics of cancer cell plasticity in real time. These transformative advancements will help the clinicians to make accurate disease assessment, predict survival outcome and design targeted, personalized cancer treatments.

CONCLUSION

Advancements in NGS, scRNA analysis, and AI-based technologies make it possible to explore the trajectories of cellular plasticity in EMT dependent and independent manner in a tumor ecosystem. It offers the best experimental design, reduces the reading time frame, risk of misdiagnosis and trial-and-error period for future clinical trials in drug development. Nevertheless, its wider inclusion is still awaited. Recent trends foresee its application in providing tailored diagnosis, risk assessment, accelerating drug discovery and individualizing treatment regimens.