Artificial intelligence and systems biology analysis in stem cell research and therapeutics development

Abstract

Background: Stem cell research has rapidly advanced during the past decades, but the translation into approved clinical products is still lagging behind. Multiple barriers to effective clinical translation exist. We hypothesize that an ineffective use of the existing wealth of data from both product development and clinical trials is a crucial barrier that hampers effective clinical implementation of stem cell therapies. Methods and Results: Here, we summarize the contribution of systems biology (SysBio) and artificial intelligence (AI) in stem cell research and therapy development, to better understand and overcome these barriers to effective clinical translation. Advancements in cell product profiling technology, clinical trial design, and adjunct clinical monitoring, offer new opportunities for a more integrated understanding of both, product and patient performance. Synergy of SysBioAI analysis is boosting a more rapid, integrated, and informative analysis of large‑scale multi‑omics data sets of patient and clinical trial outcomes, thus enabling the “Iterative Circle of Refined Clinical Translation”. This SysBioAI‑supported concept can assist more effective development and clinical use of stem cell therapeutics through iterative adaptation cycles. This includes product‑ and patient‑centered clinical safety and efficacy/potency evaluation through paired identification of suitable biomarkers of clinical response.Conclusion: Integrated SysBioAI-use is a powerful tool to optimize the design and outcomes of clinical trials by identifying patient-specific responses, contributing to enhanced treatment safety and efficacy, and to spur new patient-centric and adaptable next-generation deep-medicine approaches.

Graphical abstract

Graphical Abstract.

Significance Statement.

The clinical translation of stem cell therapies has been hindered by persistent challenges, including product heterogeneity, incomplete mechanistic understanding, and limited predictive power of current trial designs. This review highlights how the integration of systems biology (SysBio) and artificial intelligence (AI) can transform these limitations into opportunities. By enabling the holistic analysis of multi-omics datasets, patient biomarkers, and clinical outcomes, SysBioAI offers an iterative framework for refining both therapeutic products and trial strategies. This patient-centered, data-driven approach not only promises to accelerate safe and effective deployment of stem cell therapies but also establishes a paradigm for next-generation precision and regenerative medicine.

Introduction

Stem cell research has advanced rapidly over the past decades, setting the basis for novel therapeutic approaches.^1–5 Translating these early advancements into mature therapies has faced significant challenges, unraveling numerous barriers to effective clinical translation that must be understood and overcome.^5–11 This review article highlights the potential of advanced bioinformatics, artificial intelligence (AI), and systems biology (SysBio) to assist in this important endeavor (Graphical Abstract).

This current Concise Review Article provides an overview on the concepts underlying the stem cell SysBioAI analysis done in our previous Scoping Review.¹² The Scoping Review reported a concise timeline of publications on stem cell SysBioAI topic registered at PubMed between 2000 and 2024 and numerous examples of the reported daily experience and practical use cases. In short, the prior Scoping Review found an 8 to 10-fold global increase in SysBioAI-related research output between 2000 and 2021, with a 10-fold increase in AI-related publications since 2010, and an increasing shift from preclinical basic research to clinically oriented translational stem cell research. As summarized in this prior Scoping Review,¹² SysBioAI-supported data analysis is now common at all stages, from molecular level to population level analysis, and from preclinical in vitro and in vivo research to clinical research (Figure 1A and B).

Figure 1.

Advanced bioinformatics, systems biology, and artificial intelligence in stem cell research and therapeutics development.

(A) Reductionist vs holistic approach: Integrating SysBio and AI/ML/DL for complex data analysis. Demonstration of the scale progression, from molecular studies to population-based research (cohorts), with enhancement of biomedical research through the integration of advanced bioinformatic data analysis, systems biology (SysBio), artificial intelligence (AI), machine learning (ML), artificial neural networks (ANNs), and deep learning (DL). (B) Use of advanced bioinformatics, AI and SysBio in preclinical and clinical stages of stem research and development. Datasets from all three phases in stem cell research (in vitro, in vivo, and pre-clinical studies) benefit from the integration of SysBio and AI analysis tools. These studies involve the use of AI or ML, employing their individual strengths in different settings. Typically, pre-clinical analysis includes omics data, cell morphology, tissue biopsies, laboratory parameters, but also analysis of clinical cohort outcomes and biomarkers with subsequent safety and efficacy analysis. In conclusion, the SysBio approach enables integrated data analysis, visualization, and interpretation of complex biological systems.

Based on the results and examples from this prior scoping review,¹² it becomes clear that SysBio, in particular in combination with AI technology, has already proven invaluable in boosting outcomes in many data-intense research fields, particularly in enhancing the analysis of large-scale multi-omics and 3D-spatial-temporal data.^12–17 Recent success in clinical development of chimeric antigen receptor (CAR)-based cell and gene therapies and advanced therapy medicinal products (CGTs/ATMPs) underscores the potential of SysBioAI analysis for more cost-effective CGT/ATMP development in the future.^18–27 Some recent practical examples include the use of AI in automated CAR-T cell manufacturing, safety and efficacy monitoring, and AI-driven predictive approaches for improving CAR-T cell signaling/prevent CAR-T cell exhaustion, to improve their “fitness” and clinical efficacy in patients.¹⁹,²⁰,²⁶ Overall, this approach may yield safer, more effective, and cost-efficient results.

Similarly to CAR-T-development, crucial factors that still hamper stem cell therapy development include: (1) High costs for manufacturing and eventual therapy, (2) Large product heterogeneity and partly insufficient definition, and (3) Incomplete understanding of the mechanism of action (MoA) and optimal clinical use.5^–10,^28–32 Another setback to clinical success may be the delayed adoption of advanced bioinformatics, SysBio, and AI tools, that may be needed to effectively harness the wealth of data generated during stem cell therapy development and adjunct clinical trial testing.^13–16,³³,³⁴ Although thousands of clinical trials with diverse stem cell products have been performed in the past,⁴,⁸,¹²,³⁵ advanced SysBioAI data analysis tools have only arrived rather late in the stem cell clinical trials arena compared to other fields.¹²

Opposed to traditional reductionist scientific methods that are based on “taking things apart,” SysBio is a holistic approach “to better understand the larger picture.” SysBio employs advanced computational and mathematical analysis to better model and understand complex biological systems in an integrative and holistic fashion.³⁶ This includes advanced data analysis and computational tools, for example artificial intelligence, machine learning, and deep learning (AI/ML/DL, respectively; Figure 1).^12–16,^36–40 Here, we provide a comprehensive review of the interesting interplay between SysBio and AI/ML/DL in stem cell biology and therapeutics development, exploring both advancements and challenges in this novel field to overcome potential bottlenecks and accelerate its effective implementation for the betterment of human health.

Potential of advanced bioinformatics, SysBio, and AI in stem cell research

Systems biology and artificial intelligence (SysBioAI) analysis approaches and tools have already proven its great value in other data-intense disciplines, such as immunology and infectious biology, that strive for a holistic understanding of disease process on different biological levels from molecular to population level analysis (Figure 1A, left panel).³¹,³³,^41–51 This includes: (1) Analysis on molecular level (eg, multi-omics and single-cell RNAseq),¹³,¹⁴,³⁴,⁴⁵,⁵²,⁵³ (2) Cellular, tissue, and organ analysis (eg, 3D-spatial and 4D-spatio-temporal analysis of cellular features in their 3D context, appreciating both space and time),^54–57 and (3) Studies and modeling of complex “living” systems [eg, organ-on-chip (OoC) models, animal models, and human studies/clinical trials from early to advanced stage].^58–60

Novel SysBio data analysis tools are increasingly useful for the holistic analysis of datasets in medicine and stem cell research.^13–16,^36–38,^61–66 Here, SysBio approaches go beyond traditional analysis by enabling the operator to better contextualize and understand the complexity of data using a “human-like artificial intelligence.” They combine the strength of conventional computational analysis/computing power with “self-learning” computational algorithms and neural networks that increasingly mimic complex non-linear learning processes of the human brain (Figure 1A, right panel).

These developments are fueled by the rapid progress in commonly available and affordable computer hardware and software, which can easily handle large datasets in conjunction with external servers.⁶⁷,⁶⁸ Developing novel bioinformatics software and data-processing tools has been another decisive step, integrating SysBio with AI/ML/DL tools in all phases of research (Figure 1B).¹²,⁴²,⁴³,⁴⁷,⁵⁰ Together, these innovations provide prospects for a rapid acceleration in medical breakthroughs and betterment of human health in the near future, including the stem cell field.

Stem cell research has emerged as one of the most dynamic fields in biomedical research, potentially transforming regenerative medicine, disease modeling, and drug discovery (Figures 1 and 2).²,⁵,⁶⁵,⁶⁹,⁷⁰ Their unique cell-intrinsic properties make stem cells a valuable tool for studying and understanding complex biological processes, such as embryonic development, tissue regeneration, and cellular homeostasis, and thus, also for developing concomitant innovative therapeutic interventions.^71–74 In this complex and ever-evolving research landscape, the convergence of two cutting-edge bioinformatic disciplines, SysBio and AI/ML/DL, is now paving the way for new transformative breakthroughs in stem cell science,^13–16,³⁹ including advancements in treating complex diseases like Parkinson’s, cardiovascular diseases, and diabetes.⁷³

Figure 2.

Opportunities and challenges in use of advanced bioinformatics stem cell research and therapeutics development.

(A) Opportunities in use of advanced bioinformatics in stem cell research and therapy. Popular stem cell types studied and employed in stem cell research and therapeutic settings grouped according to self-renewal capacity, lineage commitment and differentiation capacity (left panel). Multiple advances in SC research through integration of SysBio and AI analysis revolutionize regenerative medicine and cell therapy (right panel): This approach bears the potential to offer new treatment approaches, personalizing therapies based on SC with more effective and tailored treatments to meet individual patient needs. Promising tools such as SysBio, AI/ML drive the continuous development of advanced and personalized therapies. (1) Totipotent SCs: Originating from fertilization, play a unique and essential role in early embryonic development, limited to the initial phase. They can differentiate into any cell type, establishing the foundations for formation of all tissues and organs, (2) Pluripotent SCs: Located in the inner cell mass of the blastocyst are formed days after fertilization during embryogenesis, then, specialize into the three germ layers, eventually giving rise to a variety of tissues and organs in the developing organism. Induced pluripotent stem cells (iPSCs) are created through the reprogramming of adult cells, acquiring characteristics similar to embryonic pluripotent cells, and (3) Multipotent SCs: Are crucial for tissue regeneration and maintenance in multicellular organisms. Mesenchymal Stromal/Stem Cells (MSCs), originate from vascularized tissue sites like bone marrow and adipose tissue, and demonstrate potential differentiation into bone, fat, and cartilage cells, along with immunomodulatory properties. Hematopoietic Stem/Progenitor Cells (HSPCs) are found in the bone marrow, and are crucial for hematopoiesis, playing a vital role in transplants to treat blood diseases. Neural Stem Cells (NSCs) are found in the nervous system and have an essential role in neurogenesis, with promising therapeutic implications for neurological disorders. (B) Multiomics challenge with integration of complex data. Different types of “omics” technologies (left panel), including genomics, transcriptomics, proteomics, metabolomics and other types of high-throughput assessments of different biological components. A major challenge is the integration of multiple data types and sources across different omics layers in SysBio and AI based analysis. Omics studies provide crucial insights on several aspects of stem cell biology, including molecular signatures, regulatory networks, and functional properties, such as the differentiation potential of stem cells (right panel). (C) Technical challenges in advanced data generation and analysis. An adequate statistical planning of the type and number of features (j) and number of observations (k) in the dataset matrix, paired with optimized sample material and processing methods result in robust data. This can be analyzed and visualized with popular bioinformatics tools, available independently on various websites (webtools), or through computer programming in languages such as R and Python. Some examples include the hierarchical clustering, k-means analysis, principal component analysis, differential expression analysis, gene alignment and random forest.

A defining feature of stem cells is their ability to self-renew and differentiate into various specialized cell and tissue types depending on their state of potency (Figure 2A, left panel), thus, distinguishing between totipotent (eg, the fertilized egg), pluripotent (eg, ESCs and iPSCs), and more lineage-committed multipotent adult and perinatal stem cells (eg, HSCs, MSCs, and DSCs).⁵,^71–77 Totipotent stem cells can differentiate into all cell types within the organism, while pluripotent stem cells (PSCs), for example embryonic and induced pluripotent stem cells (ESCs and iPSCs, respectively), can differentiate into a wide range of more specialized cell types. Multipotent stem cells, also known as adult stem and progenitor cells (ASCs/ASPCs), have a more restricted differentiation capacity according to their developmental origin,⁵,^71–77 including hematopoietic stem and progenitor cells (HSCs/HSPCs), diverse adult and perinatal-tissue derived mesenchymal stem/stromal cells (eg, MSCs and DSCs),⁵,¹⁰,^76–78 and other types of more lineage committed tissue-specific stem cells, for example skin stem cells, entailing subsets of basal and epidermal stem cells.³,⁷⁴

The complexity and dynamic nature of stem cells and the multitude of factors influencing their behavior pose a significant challenge and an opportunity for scientific exploration and clinical applications.⁶⁵ Stem cell niche-dependent control of self-renewal and asymmetric division, lineage commitment, division arrest, and differentiation potential is governed by the tight control of pluripotency and lineage-specific genes/transcription factors in conjunction with chromatin organization and epigenetic regulation of accessibility and transcription (Figure 2A, bottom left).⁷⁴,⁷⁵

Traditional reductionist approaches often fall short in providing comprehensive insights into the intricate regulatory networks that govern stem cell fate and function.³³ Here, SysBio with its strong holistic perspective can play a vital role in understanding biological systems as a whole, encompassing diverse molecular interactions between genes, proteins, and cellular pathways (Figure 2A, right).³¹,³³,^41–51 This approach allows scientists to unravel the properties of complex biological systems and provides a foundation for predicting cellular behavior with unprecedented accuracy.

SysBio- and AI-driven computational models can analyze/integrate data from diverse sources by offering the analytical power and pipelines required to handle the massive datasets needed to decipher intricate patterns. This can predict cell behavior and identify novel therapeutic targets.⁶¹,^79–81 The synergy of SysBio and AI enhances our comprehension of the intricate nature of stem cells and propels the field toward personalized medicine and precision therapies (Figure 2A, right).⁸²

In turn, clinical trials benefit by becoming more targeted and patient-centric.⁸³ With AI-driven predictive models, researchers can better identify patient-specific responses to ensure more effective and safer treatments.¹⁰,⁷⁹,⁸⁴ By better understanding the genetic/epigenetic factors that govern stem cell responses, we are entering an era where regenerative medicine may be tailor-made for each patient.

Advanced bioinformatics: data generation and analysis with SysBio and AI

Advanced bioinformatics plays a pivotal role in better understanding complex biological systems although multiple challenges exist (Figure 2B and C).⁸⁵ The rise of high-throughput omics technologies, in particular the transition from bulk to single-cell RNA sequencing (scRNAseq) and other multi-omics approaches,⁸⁶ has transformed data generation to yield extensive datasets that require sophisticated analytical approaches across different layers of molecular analysis, such as genomics, transcriptomics, proteomics, metabolomics, and other methods (Figure 2B).⁸⁷ Proper integration of these omics data with the patients’ clinical background can reveal new insights into the complex interplay between typical molecular signatures, characteristic modulation of decisive regulatory networks, and respective functional outcomes.¹⁴,⁶² Different SysBio applications are characterized by three major approaches: (1) Top-down, (2) Bottom-up, and (3) Middle-out intermediate approaches, as described previously.^88–92

Top-down approaches

Begin with high-throughput technologies to measure certain classes of macromolecules (eg, RNA with transcriptomics), which are then analyzed with robust algorithms.¹³ Crucial steps in this analysis include (Figure 2C): (1) statistical planning of sample collection and analysis, ideally before the initiation of the experimental comparison/clinical trial, to adjust for the sample size accordingly (eg, power analysis of a number of samples required to obtain a significant result; Figure 2C, left),^93–95 (2) biological sample processing, requiring suitable starting material and fixation for later processing (eg, purity and integrity; Figure 2C, center), and (3) data analysis, visualization, and interpretation (eg, employing analytical tools; Figure 2C, right). Here, SysBio can be helpful in the integration of results from various analytical software tools and AI/ML/DL techniques to facilitate the classification and interpretation of results from complex data visualization approaches.

The last point, data analysis/visualization/interpretation, is often considered the most challenging step, since it employs multiple analytical tools that require an advanced understanding of bioinformatics software (Figure 2C, right),¹³ including: (1) hierarchical clustering (HCL) based on sample similarity, (2) K-means analysis based on dividing variables into a user-selected number of groups, (3) principle component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) of high-dimensional data to reduce the number of dimensions for improved visualization and comprehension, (4) differential expression and peak comparison analysis to determine if a set of genes is statistically different between two conditions, (5) gene alignment and gene ontology (GO) terms enrichment analysis to align identified genetic signatures to certain biological process stored in data bases, and eventually, (6) random forest (RF) and support vector machine (SVM) analysis as novel machine leaning approaches to conduct ranking and regression analysis, as outlined in the following sections.

Bottom-up and middle-out approaches

In contrast to going into more detail by breaking the whole system apart, bottom-up approaches form detailed models from subunits of data to simulate whole systems for modeling of physiological conditions (eg, genome-scale or flow simulations).^89–92 This approach has been used for tissue-specific simulations, for example systems pharmacology, to model blood flow/tissue responses with simulation of drug distribution, safety, and efficacy.⁹¹,⁹²,^96–98 It is not always clear whether these models accurately reflect clinical conditions. Thus, comparative studies of both top-down and bottom-up approaches need to be carried out to compare their ability to predict the clinical safety and potency of certain drugs.⁸⁹,⁹¹ To reduce the uncertainty in drug models, “intermediate or middle-out” approaches can be used, where top-down and bottom-up data analysis is combined.⁹² All these approaches have great potential to reduce the need for animal models and the length of clinical trials.

Strong need for statistical planning before starting your study

Before delving into the meticulous details of big data analytics, it is essential to evaluate key aspects that will greatly impact the final results, including the size of the dataset and an appropriate number of groups (variables and rows; Figure 2C).^93–95,⁹⁹ This decision significantly affects the subsequent selection of statistical tests, thereby shaping the methodological framework of the computational analysis.⁹³

For inferential statistics or AI/ML/DL approaches, it is important to remember that confidence intervals (CI) and/or P-value results are directly influenced by the sample size and data distribution of each variable. Here, various analysis methods exist to evaluate the suitability of a dataset for conducting robust statistical tests, including multiple linear regression, logistic regression, and the Wilcoxon-Signed-Rank test.^100–102 Tools such as the G-Power and its contained testing methods serve as valuable tools for ensuring the dataset’s adequacy and reliability for meaningful statistical analyses, thereby contributing to the overall rigor and validity of the research endeavor.⁵⁰,¹⁰³

Summarizing, the dataset’s structure, including the number of columns and rows, plays a vital role in the success of integrative big data and multi-omics analysis.⁴⁶ The multitude of columns—representing variables or features in the dataset—influences the choice of statistical methods (eg, regression analysis),⁵⁰,¹⁰⁴ which is particularly pertinent when exploring the relationships between different variables across different omics layers.⁴⁶ Simultaneously, the number of rows reflecting individual observations or samples becomes essential in determining the statistical power and precision of the study. The meticulous alignment of the dataset composition with suitable statistical tests during the planning phase ensures a robust foundation for accurate analyses and meaningful interpretations, paving the way for a comprehensive understanding of stem cell characteristics through integrative multi-omics analysis.¹⁰⁵

Intricacies of biological sample preparation and data integration strategies

Sample collection and processing demand careful consideration of the specific requirements tailored to the different readout methods (Figure 2C, center).¹⁰⁶ Whether employing genomics, transcriptomics, proteomics, or metabolomics, each distinct method comes with its own set of intricate sample processing requisites.⁸⁷,^107–109 Sample processing kits for a combined analyte extraction/readout with different omics approaches in one workflow are available and can reduce sampling requirements.⁸⁷,¹⁰⁸ The use of reference samples should be considered in multicenter studies.¹⁰⁹ When pursuing single omics approaches, the intricate requirements of each platform must be considered early on in experimental and clinical studies to prevent sample collection- and processing-related artifacts/degradation that may falsify the correct representation of the sample intrinsic information. The precautions below may increase the chance of an optimal and reproducible readout of the original sample encoded information.

DNA is generally considered a fairly stable molecule at ambient conditions, thus, also allowing for DNA extraction from biospecimens resting for a substantial time under ambient conditions—preferentially assisted by sequence alignment to already processed complete datasets/genomes found in specialized online databases (eg, EMBL-EBI, BLAST, and NCBI nucleotide search and sequence analysis databases).¹¹⁰ Nonetheless, any genomic analysis methods—such as different generations of DNA sequencing techniques¹¹¹—will aim to extract nonfragmented high-quality and high-purity DNA considering the anticipated biological source material of interest. These principle steps are paramount to eventually yield an accurate representation of the desired genome upon sequence alignment in well-established online databases, thus, requiring the necessary attention of investigators to optimal isolation and purification techniques to prevent DNA degradation/fragmentation/contamination.¹¹²

In transcriptomics and gene expression analysis, stringent protocols are required to maintain the RNA integrity and transcript stability to prevent potential bias, for example RNAse inhibitors and cold storage.¹¹³ On optimal extraction yield and molecular integrity, for example a lower cutoff of a RNA integrity number (RIN) value >7 is often used in bulk RNAseq approaches,¹¹⁴ of the different types of RNA species, such as messenger RNA (mRNA), micro RNA (miRNA), and long noncoding RNA (lncRNAs), to just name a few, is paramount—due to the high sensitivity of RNA to degradation by different types of ribonucleases (RNAses, eg, endonucleases and exonucleases) and different intrinsic stability of different RNA species depending to their susceptibility to cell intrinsic/extrinsic DNA degradation systems/mechanisms.¹¹⁵

Proteomic and metabolomic methods often involve some form of cold-storage of the samples/biospecimens with variable temperature requirements depending on the sample type (eg, storage of blood, plasma, serum, urine, or tissue samples at 4 °C fridges to –20 °C freezers for short term storage if no other means are available, or –80 to –150 degrees freezers for prolonged storage of samples with minimal degradation, or –196 degrees liquid or vapor-phase liquid nitrogen of cells to maintain their viability upon thawing).³⁰,¹¹⁶ The initial cold storage of biospecimens is often followed either by direct analysis of sample aliquots with suitable assays (eg, Affymetrix, ELISA, OLINK, or HPLC/GC platforms),^117–119 or by specialized analyte extraction procedures to capture the full spectrum of analytes in a sample.¹²⁰,¹²¹

Typical analytes include increasingly larger panels of cytokines, chemokines, growth factors, or small-sized metabolites indicative of biological processes, such as regenerative, immunomodulatory, or stem cell niche-related functions.¹²²,¹²³ For some sensitive analytes, analysis of fresh samples may be mandated, especially in the setting of routine analysis, as specified in the respective ISO guidelines.¹²⁴

Apart from the detailed transcriptomic analysis of immune cells and other cellular subsets with scRNAseq profiling, the high-parametric analysis of pre-defined clusters of differentiation (CD) marker panels with both flow and mass cytometry (FACS and CyTOF, respectively), still forms the strong analytic foundation of cellular analysis in most research labs today.¹¹⁴,^125–133 These profiling platforms allow for a solid readout of predefined surface and intracellular CD markers on different types of stem cells, immune cells, and therapeutic cell products according to their cell type intrinsic developmental differentiation pathways.¹²⁵ If the necessary requirements for biosample sourcing and subsequent analysis are met (eg, correct setting and speed of analysis, with adequate ambient temperature control and preservatives during shipping, and validated readout equipment with suitable standardized antibody panels, such as DuraClone panels), these analysis platforms may actually allow for a reproducible, standardized, and accredited/certified readout in specialized facilities/labs.¹¹⁴,^125–129,¹³³

The aforementioned details of proper statistical planning, sample collection, and processing are tightly linked to the broader challenges faced in bioinformatics, particularly concerning correct data integration and representation strategies.^134–137 The diverse nature of omics data requires innovative approaches to integrate information from the different analytic platforms cohesively, thus, requiring suitable strategies to harmonize the data from varied methodologies, ensuring a comprehensive and accurate representation of biological systems.⁸⁷ A correct match and representation of training and target datasets is essential for AI-based analysis.^134–137 Thus, developing and adopting advanced computational techniques and standardized protocols for data generation and integration is imperative in overcoming the complexities inherent in the diverse array of omics datasets, ultimately enhancing the reliability and utility of biological insights derived from integrated analyses.¹⁰⁹,¹³⁸

Bioinformatics tools for visualization and interpretation of SC data

Both stem cell research and its adjunct clinical trials generate vast datasets on different technology platforms, requiring sophisticated analytical tools for effective interpretation with integrated use of SysBio and AI (Figure 3A).¹²,¹⁶ Thus, nowadays, the effective use of advanced bioinformatics tools and software plays a crucial role in visualizing these highly complex datasets, thereby offering researchers the means to explore and better understand the molecular signatures, differentiation potentials, and regulatory networks that govern stem cell behavior and to communicate their results.¹⁶

Figure 3.

Increasing integrated use of AI and SysBio approaches.

(A) Intersection of AI and SysBio analysis in stem cell research and therapeutics development. The blue and green colors represent the AI/DL and AI/ML approaches, respectively. The blue and green lines represent connections that AI approaches can make to SysBio. The central square represents the connection between the approaches. (B) Typical machine learning algorithms and artificial neural networks analysis employed in SysBio and AI analysis. More detailed elaboration and practical examples can be found Supplementary Tables S1-S5.

Some freely available powerful web tools for analyzing omics data include: (1) Morpheus bioinformatics for the simulation and integration of cell-based models with ordinary differential equations and reaction-diffusion systems,¹³⁹ (2) Network analyst to provide a systems-level data understanding with the integration of gene expression data with protein-protein interaction (PPI) networks,¹⁴⁰ and (3) The Galaxy Platform for accessible, reproducible and collaborative biomedical analysis,¹⁴¹ amongst others. These tools enable scientists to gain meaningful insights from the wealth of data through graphical representations, statistical analysis, and interactive platforms.

Furthermore, SysBio researchers often use diverse power tools to integrate two or more datasets from the commonly used RNA sequencing approach.⁴⁶,¹⁰⁴,¹⁴²,¹⁴³ There are typically two different approaches to follow: (1) Consensus analysis,⁸⁸,¹⁴² or (2) ComBat analysis for statistical correction of batch effects,¹⁴⁴ for example by employing the Network analysis web tool.⁴⁶ These web tools are easily accessible and widely used in SysBio analysis. Another option is to employ programming languages like R,¹⁴⁵ Python,¹⁴⁶ and Shell/Linux—GNU.¹⁴⁷ For both R and Python, open-source software packages are available at websites such as Bioconductor,¹⁴⁸ that can facilitate complex statistical analysis. In general, the motto—learning by doing—applies.

Within the realm of programming languages, sophisticated analyses with ML methods employs statistical algorithms that are based on mathematical optimization and prediction models (Figure 3B and Table 1). These learn from existing datasets to generalize and predict outcomes for previously unseen data. In turn, data mining focuses on discovering unknown properties in the data (knowledge discovery) through employing ML methods, but with different goals. Thus, ML algorithms are described as “self-learning-algorithms” that perform tasks without explicit instructions.¹⁷²

Table 1.

Use of machine learning algorithm and artificial neural networks.

Author and references	Algorithm	Type	Description	Biological application
Liu et al.149	Linear regression	Supervised	Predicts a continuous output (numerical value) based on the linear relationship between the input variables (features) and the target output.	Gene expression analysis, dose-response relationships
Ban et al.151	Logistic regression	Supervised	Models binary or categorical outcomes by estimating probabilities using a logistic function, often applied to classification tasks.	Disease diagnosis and risk prediction, gene mutation and drug response prediction
Wu et al.153	Decision trees	Supervised	Constructs a tree-like model of decisions by splitting the dataset into subsets based on feature values, useful for both classification and regression.	Disease diagnosis and classification, gene function prediction
Imai et al.155	Random forest	Supervised	An ensemble learning method combining multiple decision trees to improve prediction accuracy and reduce overfitting.	Biomarker discovery, genome-wide association studies (GWAS), ecology and evolutionary biology
Lien et al.157	Support vector machines (SVM)	Supervised	Finds the optimal hyperplane that maximizes the margin between different classes for classification or regression tasks.	Protein structure and function prediction, transmembrane-protein topology prediction
Mota et al.159	K-nearest neighbors (KNN)	Supervised	Classifies a data point by analyzing the classes of its k-nearest neighbors in the feature space, often used for classification tasks.	Protein function prediction, medical image analysis
Wood et al.161	Naive Bayes	Supervised	A probabilistic classifier that assumes independence between features, efficient for large datasets and suitable for classification problems.	Disease classification (eg, classifying cancer subtypes), microbial classification, text mining in bioinformatics, gene classification
Buggenthin et al.163	Artificial neural networks (ANNs)	Supervised	Mimics the structure of biological neurons to learn complex patterns in data, used for tasks like image recognition, speech recognition, and more.	Drug discovery, personalized medicine, medical data mining, metabolic network modeling
Katarzyna et al.164	Dense neural networks	Supervised	A fully connected neural network where each node in one layer is connected to every node in the subsequent layer, used for a wide range of tasks.	Gene expression analysis, predicting drug-target interactions in drug discovery using molecular data.
Khouj et al.166	K-means clustering	Unsupervised	Groups data into K clusters by minimizing the distance between points and their assigned cluster centers based on feature similarity.	Grouping similar gene expression profiles, image segmentation, clustering cells in single-cell RNA sequencing
Zimmermann et al.168	Principal component analysis (PCA)	Unsupervised	Reduces dimensionality by transforming correlated features into a smaller set of uncorrelated components, often used for data compression.	Visualizing gene expression data, reducing high-dimensional omics data
Taherkhani et al.169	AdaBoost (adaptive boosting)	Ensemble	Combines multiple weak classifiers into a strong one by focusing on misclassified examples and adjusting weights to improve model performance.	Improving cancer diagnosis from imaging, integrating diverse biological datasets
Rapakoulia et al.170	Gradient boosting machines (GBM)	Ensemble	Sequentially builds models by correcting the errors of previous models, commonly used for regression and classification with high accuracy.	Predicting disease risk factors and discovering censored survival outcomes, identifying key genetic drivers of diseases

Popular ML methods include (Figure 3B), for example (1) Regression Analysis (eg, multiple linear regression),¹⁰⁰ (2) Principal Component Analysis (PCA; to simplify the complexity of high-dimensional data while retaining trends and patterns),¹⁷³ (3) Random Forests (RF; a combination of tree predictors, where each tree depends on values of a random vector sampled independently and with the same distribution for all trees in the forest),¹⁷⁴ (4) Support Vector Machine (SVM learning—a powerful analysis tool for classification and subtyping of data),¹⁷⁵,¹⁷⁶ and (5) Bootstrap (eg, tests or resampling methods that work through assigning measures of accuracy to sample estimates, including measures such as bias, variance, confidence intervals).¹⁷⁷

Regression analysis provides insights into the relationships between variables and enables forecasting of trends by modeling how changes in one or more independent variables impact a dependent variable. Linear regression is used when the dependent variable is continuous and the relationship with independent variables is assumed to be linear, allowing for the prediction of future values based on this linear relationship. Binomial logistic regression is applied when the dependent variable is categorical with two possible outcomes. This method estimates the probability of a particular outcome by modeling the relationship between the independent variables and the odds of the outcome, using a logistic function to ensure predictions fall between 0 and 1.

Random Forest and Gradient Boosting Machines (GBM) are advanced ensemble methods that build on decision trees to enhance predictive accuracy. Random Forest constructs multiple decision trees using random subsets of the data and features, combining their predictions to improve robustness and reduce overfitting. Gradient Boosting Machines create a series of decision trees sequentially, where each new tree focuses on correcting the errors of the previous ones, thereby refining the model’s performance through iterative improvements. Both methods leverage decision trees but differ in their approach to aggregating predictions and addressing model errors.

Support Vector Machines (SVMs) are classification and regression techniques that identify optimal hyperplanes that best separate different classes by maximizing the margin between data points of different classes, thus, ensuring a clear separation. For non-linear relationships, SVMs use kernel functions to transform the data into higher dimensions where a linear separation is possible. This approach is particularly effective in high-dimensional spaces and for complex classification problems.

K-means clustering partitions the data into a predefined number of clusters by iteratively assigning data points to the nearest cluster center and updating the centers to minimize the within-cluster variance. This method is effective for finding natural groupings in data but requires specifying the number of clusters in advance. Clustering encompasses various techniques for grouping data into clusters without predefined labels, including hierarchical clustering. These methods aim to uncover inherent structures in the data and can adapt to different types of cluster shapes and densities.

The ML algorithms enable the identification of complex patterns and non-linear relationships within datasets, thus, providing valuable insights for more in-depth and predictive analyses.¹⁷⁸ Artificial neural networks and deep learning approaches have recently surpassed many previous ML approaches in performance.

Integrated SysBio-AI analysis in SC research and therapeutics development

Integrated AI and SysBio analysis has become a field of increasing interest in both stem cell research and concomitant clinical trial analysis (Figure 3).³⁶ Importantly, the integrative SysBio approaches offer the opportunity for connecting the molecular components to physiological functions and organismal phenotypes within both a single biological scale (eg, on type of tissue) but also among different scales (eg, cells, tissues, or organ systems), through quantitative reasoning, computational models and high-throughput experimental technologies such as scRNAseq analysis (Figure 4).³⁶ This allows SysBio researchers to better decode the flow of information from genes, proteins, and subcellular components, intracellular/extracellular pathways, up to the regulatory control of cell, tissue, organ, and organismal level functions, which are all highly relevant in stem cell research and respective clinical trial efforts.

Figure 4.

SysBio workflow applied to data from stem cell research and therapy development.

(A) Typical SysBio workflow for scRNAseq analysis of stem cell samples. Bioinformatic workflow in SysBio assisted scRNAseq analysis of stem cell samples: scRNAseq can reveal the gene expression profile in different types of stem cells and through SysBio it is possible to identify different sets of genes being differentially expressed in each type of stem cell. Furthermore, other SysBio methods such as functional enrichment can be applied to better understand biological processes. (B) SysBio approaches can be applied in clinical stem cell therapy studies. The application of integrative SysBio methods to data generated in clinical stem cell studies can generate a better understanding for different complex diseases and enable more substantiated guidance in clinical trials.

Although the stem cell field has been one of the pioneers in multi-omics analysis, this has often been limited to detailed lab-research-centered preclinical, experimental, and modeling approaches in the past to better understand intrinsic stem cell behavior (eg, potency and fate decisions).¹⁴,³⁴,⁵³,⁶¹,¹⁷⁹ Long et al. recently gave an elaborate update on scRNAseq method advancements in MSC research, demonstrating the rapid evolution of methodology in this field.³⁴ In exemplary fashion, Miura et al. employed scRNAseq to identify an LRRC75A-expressing subpopulation of MSCs involved in the secretion of vascular endothelial growth factor (VEGF) under ischemic conditions in their quest to explore phenotypic and functional heterogeneity in MSC products in relation to their angiogenic potency.⁵³ Furthermore, scRNAseq of “out-of-thaw MSCs” can identify differences related to both tissue source and inter-donor variations,¹⁷⁹ thus, widening the scope of earlier studies on the potential impact of cryopreservation and freeze-thawing for clinical MSC product performance.⁹,³⁰,¹¹⁶,^180–189

By integrating different omics datasets with SysBio and advanced AI/ML/DL, it is possible to obtain a systemic view of the biological networks and stem cell (niche) interactions with either the local or the systemic environment, that exert impact on the potency, lineage differentiation, and the resulting progeny of stem cells.¹³,¹⁴,³⁹,⁶¹,⁶²,⁸⁷ Tools such as Clusterprofiler (An R package for comparing biological themes among gene clusters),¹⁹⁰ EnrichR,¹⁹¹ and GSEA¹⁹² (Gene set enrichment analysis for interpreting genome-wide expression profiles) identify biological functions and cellular processes enriched in a set of genes associated with stem cell biology (Figure 4A).¹⁹³

There are various categories of stem cells with varying degrees of plasticity, but despite their distinct phenotypes and functions, all these varieties of stem cells share common fundamental characteristics in their biology, such as the activation and repression of biological processes linked to a set of stem cell related genes.⁶⁵,⁷⁵,¹⁹³,¹⁹⁴ Currently, scRNAseq stands out for its ability to provide more refined details about the characteristic expression profile in various types of stem cells.¹⁹⁴,¹⁹⁵

Recently, Barata et al. identified core functional modules of stem cells and potential novel “stemness genes” by integrating known stemness gene signatures.¹⁹⁴ The authors compared 21 distinct stemness signatures for humans and mice, identifying a significant overlap between both species, thus, allowing for the identification of integrated stemness signatures (ISS) comprised of genes frequently occurring among stemness signatures.

In a rat model of Parkinson’s disease (PD),¹⁹⁵ Tiklová et al. employed scRNAseq in combination with histological analysis to characterize cellular diversity in intracerebral cellular grafts of transplanted human ESCs and fetal tissue. They demonstrated that neurons and astrocytes are major components of both graft types and identified a perivascular-like cell type in the stem cell-derived grafts.

Advanced tools like Harmony, LIGER, and Seurat (R toolkits for alignment of high-throughput scRNAseq datasets) offer powerful statistical approaches,^196–198 for example including batch-effect correction methods, moderated estimation of fold change and dispersion for RNAseq data with DESeq2, thus, simplifying the analysis and facilitating the visualization of complex data. These tools enable a more quantitative analysis focused on the individual strength of certain expression signatures rather than the mere presence of differential expression.¹⁹⁸ Here, batch-correction normalization and removal of batch effects is decisive for appropriate quantitative representation.

Furthermore, this approach allows for functional enrichment analysis, identifying distinct molecular signatures for each stem cell population and potential correlations between populations based on the expression profile. The synergy among various disciplines in interpreting results from complex biological models can lead to more sophisticated future conclusions regarding stem cell-related outcomes,¹⁹⁴ thereby establishing a solid foundation for the identification of specific markers and potential therapeutic targets and paving the way for promising avenues in future clinical trials and advancements in biomedical research.¹⁹⁹

The ability to map critical signaling pathways allows for a better biological understanding of stem cell-related processes in patients and respective therapeutic interventions, and thus, a more refined customization of clinical protocols.²⁰⁰ This not only bears the potential to enhance the treatment efficacy but it can also contribute to patient safety by minimizing the risks of adverse events. Moreover, a comprehensive understanding of the molecular characteristics of stem cell biology through SysBio analysis in the preclinical setting may facilitate the selection of patients for clinical trials, thus, maximizing the chances of success and expediting the translation of stem cell-based therapies into clinical practice (Figure 4B).⁸⁴

SysBio analysis in stem cell research and therapeutics development

Recent developments aim to employ SysBio for a better understanding of more complex physiological processes related to stem cell biology, for example organismal aging, regeneration, and therapeutic responses in a systemic context.⁶²,⁶⁶,⁶⁹ McNamara et al. recognized the value of SysBio for studying complex biological systems and hypothesis generation in the context of regenerative medicine, seeking to replace or repair tissues with compromised function to improve their functionality,⁶² while Raj et al. reviewed the value of SysBio for drug development.⁶⁹ Both studies exemplify the adaptability of SysBio to fields related to stem cells, regenerative medicine, and pharmacology.

Researchers in data-intense biomedical fields (eg, immunology and infectious biology) are the early adopters of SysBio analysis, a common tool for analyzing complex data in these fields.³¹,⁴⁴,⁴⁶,⁴⁸,⁴⁹ This early adoption of SysBio analysis was driven by the need to better integrate human clinical trial results and patient data in proper relation to their adjunct mechanistic analysis in the laboratory to best ensure the practical relevance of their experimental research.⁷⁰ Stem cell research and adjunct clinical trials rapidly follow suit to better integrate more holistic SysBio analysis in their development pipelines to yield significant results.¹⁴

In the past decades, stem cell product-based clinical trials have matured from small concept studies on few patients (eg, phase I studies with typically no more than 9 patients in the 3 × 3 design for early-stage dose-escalation trials), to an increasing use of advanced phase IIb and III studies, to prove the safety and efficacy of a rather diverse panel of distinct stem cell products in multiple clinical indications for regional and international marketing approval.⁹,¹⁸,^201–208 Such advanced studies provide the ideal training ground, complexity, and volume of data to explore the strengths of both SysBio and AI for advanced integrated data analysis (Figure 4B).

Examples of clinical stem cell trials with adjunct mechanistic studies include the approval of the AT-MSC-based cell product Cx601 for the treatment of complex perianal fistulas in Crohn’s disease,²⁰¹,²⁰⁹ systemic intravascular MSC therapy for the treatment of HSCT-related complications, for example acute Graft-versus-Host Disease (GvHD) and HSC engraftment failure,⁵,⁹,^202–205 systemic MSC therapy for sepsis and pulmonary complications, for example acute lung injury (ALI), acute respiratory distress syndrome (ARDS) and coronavirus-induced disease 2019 (COVID-19),⁵,²⁹,³¹,²¹⁰,²¹¹ but also local intramuscular administration of MSC-like cells (PLX-PAD) for treatment of muscle injury following hip arthroplasty and critical limb ischemia (CLI).^206–208

These advanced studies are characterized by more complex clinical trial design, larger patient cohorts—often including hundreds of patients, and an abundance of product-, patient-, and biomarker-related data. These are collected during the trials to better understand the trial outcome and gain more value from these precious studies. This approach aims to overcome the “black box” effect of completing trials with a too narrow or singular focus on the primary endpoint, but to gain more knowledge from a diverse panel of secondary readouts and adjunct biomarker panels.¹²⁵,¹²⁸,¹²⁹,²¹²

Thus, the need to better integrate preclinical (mechanistic) data with complex patient data and clinical trial outcomes has arrived in stem cell research and clinical trials. Here, SysBio approaches (in particular with the use of AI/ML/DL) offer a much deeper integration and understanding when integrating and analyzing the different obtained datasets (eg, cell product properties vs patient outcomes) to better understand the reasons for clinical trial success or failure, and to be able to further develop the established technology pipelines to successful therapeutics that eventually meet the high requirements of marketing approval.¹⁸,^213–215

From SysBio to multimodal AI-/ML-/DL-based analysis in stem cell research

To overcome limitations in high-throughput data analysis, SysBio analysis can benefit from the integration of AI/ML/DL approaches and, recently, multimodal learning as one of its most advanced disciplines (Figure 5A).^39–43,⁴⁵,⁴⁷,²¹⁶

Figure 5.

Enhancing interpretation of biomedical results through use of SysBio-AI synergy.

(A) Integrated use of AI in health care for improved diagnostics and treatment. The illustration depicts a health care scenario where human creativity, problem solving skills, and adjunct decision-making processes of health care professions are increasingly supported by the integrated use of SysBio and AI/ML/DL analysis tools, to yield improved outcomes for patients. (B) Supervised and unsupervised learning approaches in ML/DL. Simplified overview of the types of supervised and unsupervised AI/ML and AI/DL. Supervised and unsupervised artificial learning approaches utilize distinct statistical methods, including regression, clustering, and dimensionality reduction, to enhance the interpretation of large datasets. (C) Practical application of AI and ChatGPT. The increasing impact of AI use in biomedicine and health care is exemplified by the ChatGPT chatbot example, which can be employed to optimize statistical methods, result interpretation, and fostering constructive discussions to enhance treatments.

A particular challenge for AI considering the analysis of data is to use different modalities or sources of data at the same time in an integrated fashion, thus, requiring sophisticated computational tools and advanced processing power for multimodal data fusion and analysis (eg, see developments at the Fraunhofer-MEOS-Center).²¹⁶ Classically, this would apply to simultaneous/integrated analysis of different sensory data sources, such as text, picture, or audio, to represent/understand the necessary sensory environment to fulfill a dedicated task (eg, robotic-assisted surgery).²¹⁷

Multimodal learning can be of great value for integrated analysis of data from different sources in SysBio, for example in biomedicine, often different types of omics data, microscopy, phenotypic profiling, patient information. Although still in an early stage, in the future, multimodal AI may yield more contextualized analysis with a deeper understanding of the data in their larger biological context instead of evaluating parameters individually, for example integrated multiparametric analysis of (stem) cell phenotype, viability, metabolic activity, and transcriptomic profile.³⁹,¹¹⁴

This may be useful for: (1) Multiparametric imaging of cells/tissues in vitro, but also for automated imaging analysis of complex tissues, organs, and radiology data, (2) Combined analysis of gene expression, genome annotation, and protein binding, and (3) Identification of specific markers, transcription factors, and transcriptional regulatory networks etc.³⁹,⁴⁰,⁸¹ This may improve the speed, efficacy, and accuracy of diagnosis and treatment (Figure 5A).⁸¹ Nonetheless, most early ML/DL advances were/are still achieved with data originating from one source.

Kufel et al. recently gave an overview of distinct AI/ML/DL approaches with examples of their medical use (as summarized below and in Figure 5B).⁸¹ Kupfel et al. provide brief explanations of: (1) Commonly used ML models and methods (eg, k-Nearest Neighbor Algorithms, Linear Regression Algorithms, Logistic Regression Algorithms, Naïve Bayes Classifier Algorithms, Support Vector Machines, AdaBoost, and XGBoost) and (2) DL and neural networks that resemble the human brain [eg, Artificial Neural Networks (ANNs) with single layer ML, Deep Neural Networks (DNNs) with multiple hidden layers between input and output layer, and Convolutional Neural Networks (CNNs) consisting of input, spline, auxiliary, and output layers].⁸¹

Many of the ML/AI algorithms are designed for data regression or classification using available datasets, where an ML algorithm, supported by a mathematical model, generates predictions or specific decisions. The main types of ML are supervised and unsupervised learning (Figure 5B). Supervised ML assumes that the model has been trained on a similar dataset as the problem, consisting of input and output data, while unsupervised ML differs in its use of unannotated data, which were not previously labelled by humans nor algorithms. In supervised ML, once the model learns the relationship between the input and output, it can classify new unknown datasets and make predictions or decisions using classification (eg, binary two-class classification yes/no response) or regression (eg, prediction of continuous values based on input data to forecast future costs). In unsupervised ML, the model learns from input data without expected values. The algorithm focuses on grouping data based on their characteristics. The goal is to teach the machine to detect patterns and group the data without a single correct answer, using either clustering (eg, grouping the data based on similarities and differences) or association (eg, using specific relationships between data). In turn, DL and neural networks employ a much more complex layered architecture.⁸¹

The amount of data needed for ML training/testing depends on several factors, such as the type of ML problem (eg, supervised models need more data than unsupervised ones), model complexity (eg, the more complexity, the more data needed), data quality and accuracy (eg, amount of noise).²¹⁸ To get a first estimate, the “rule-of-thumb” may be used for smaller datasets, where you need at least 10 times as many data points as there are features in your dataset, to ensure a minimum quality input and avoid sample bias and underfitting.²¹⁹ Sample size and number of features are two different attributes that vary in different data types, for example images and omics data have different feature dimensions, but the number of samples depends on available data.

For AI-based analysis of larger datasets, more advanced statistical methods to adequately estimate the necessary sample size must be employed to guarantee optimal performance (see section strong need for statistical planning before starting your study),⁹⁹ including checks of dataset size, number of groups (features/variables and rows), data quality and robustness of analysis,⁵⁰,^93–95,¹⁰⁴ using, for example multiple linear regression, logistic regression, Wilcoxon-Signed-Rank test with tools such as G-Power software.⁵⁰,^100–103 For ML, it is advised to first train with 70%-80% of the sample data and then test with 20%-30% of the volume, and in particular, a correct match and representation of both training and target datasets is essential.^134–137

Several strategies can be employed to reduce the amount of data needed for an ML model to optimize computational calculation/processing requirements, for example using PCA with recursive feature elimination (RFE) to identify and remove redundant features from a dataset.²¹⁸ Dimensionality reduction techniques, such as singular value decomposition (SVD) and t-distributed stochastic neighbor embedding (t-SNE), can lower the number of dimensions in a dataset while preserving important information. Synthetic data generation techniques like generative adversarial networks (GANs) can generate more training examples from existing datasets.

Examples of AI/ML/DL-based analysis in the stem cell and RegMed field

Next, we also wish to briefly introduce practical example(s) for using AI in stem cell research, therapeutics development, and regenerative medicine (RegMed). Others have already given more in-depth overviews of practical applications.⁸¹,⁸²,²²⁰,²²¹ As outlined in our recent Scoping Review,¹² the stem cell SysBioAI field has been expanding at near-exponential rates in recent years. A comprehensive summary of the past and current developments is provided in this adjunct paper.

Similar to using the SysBio approach, one of the most typical scenarios where AI/ML/DL can be of substantial use is the analysis of large datasets from stem cell experiments and clinical trials that quantify numerous molecular and environmental variables related to clinical safety and efficacy. Due to the large and intricate network of potential molecular interactions, such analysis presents considerable challenges for any analyst, clinician, scientist, or regular entity alike. This includes, among others, the causal analysis of stem cell tumorigenicity or cancerous progression in vitro (eg, as a safety measure before clinical use) and in vivo (eg, to detect host intrinsic or therapy-induced cancerous transformation), for example by using stem cell phenotypic and genotypic datasets, and analysis of characteristic or subtype-specific cancer markers.⁷,²⁸,^222–228

More specialized questions include, for example hemocompatibility and hematotoxicity analysis of CAR-T and IV stem cell therapeutics.⁷,⁸,¹⁰,²⁸,²²²,²²⁷,²²⁸ Other topics of increasing interest are neuropsychology and neuroimmunology in the interplay of general body health, immunological status, and thus, regenerative capacity with typical manifestations of poor brain health or mental illness.⁴⁴,²²⁹,²³⁰ In the past, decisions were often based on individual biomarkers, while AI-based analysis also enables better identification of more complex molecular (autoantibody) network signatures (eg, multiple autoantibody-GPCR network interaction signatures instead of one).^41–43

In stem cell-based tissue engineering, recent innovations in advanced scaffold design, 3D-printing, and 3D-spatiotemporal cellular analysis have provided a new boost, but the stem cell-based replacement of damaged tissues and organs is still substantially held back by the enormous engineering, logistic, and analytical challenges required to grow functional replacements.^231–233 Tissue and organ replacement are still constrained by donor shortages, but potential alternatives such as organ generation and replacement with tissue-specific stem/progenitor cells and iPSCs are increasingly explored.²³⁴ The iPSCs offer personalized treatment without ethical or rejection concerns,²³⁴ but considerable practical challenges considering the reprogramming and differentiation efficiency still exist, alongside outcome variability. Here, AI/ML/DL, plays a crucial role in the quality control assessment of iPSCs, predicting induction and differentiation, and characterizing cellular functions (Figure 5B).²³⁵

Liu et al. recently published an example on using AI/ML/DL in the meta-analysis of published MSC research on cartilage repair.¹⁴⁹ The authors aimed to better use the existing wealth of data to predict crucial impactors on MSC efficacy in cartilage repair to overcome inconsistencies in their therapeutic use. They first identified 36 articles and respective datasets in PUBMED to compile a small data base, including data from 15 clinical trials and 29 animal models. They employed artificial neural network (ANN) analysis (Figure 5B), to compare the relative strength of potential impactors, such as cartilage defect area, defect depth, type of damage, body weight, tissue source of therapeutic MSCs, or implanted cell numbers, to identify the factors that are most important for therapeutic outcome. Eventually, the authors identified defect area, defect depth, and implanted MSC number are the three most critical impactors on cartilage repair in their model, with a dose of 17-25 million MSC/defect providing optimal cartilage repair. Interestingly, the authors conclude that this approach/model can also be adapted for use in other clinical indications/treatments. The authors also emphasize that handling incomplete or missing data was a crucial aspect in the development of this model, thus, emphasizing the need for sufficiently sized datasets and, in this case, actual databases in the development of AI models.

AI-based advanced chatbots, such as the Generative Pre-Trained Transformer chatbot (ChatGPT), can prove useful in stem cell research and SysBio data visualization and interpretation (Figure 5C and Tables S1-S4).²³⁶ ChatGPT is of use in many ways, for example: (1) to improve data cleanup and coding readability/documentation, (2) to interpret data more effectively, generate visual summaries, and improve your writing, and (3) to prompt engineering/designs, stay on top of new developments and increase novelty in your stem cell research field.¹⁷ ChatGPT is a software engineering solution built on Transformer Architecture a DL neural network known for its ability to model complex dependencies in data sequences, such as text. It employs pre-trained transformers—GPTs, which are the foundation models (Tables S1-S4). ChatGPT can “understand” and generate text based on broader and more intricate contexts. It is important to note that, like any other ML/AI model, ChatGPT may occasionally produce minor errors and requires a subject matter expert to interact with it effectively and rectify any mistakes. As more and more AI/ML models stand out for their ability to process increasingly large volumes of generated data and efficiently identify complex patterns, their capacity to learn from past experience and dynamically adjust to new information provides the crucial key for more effective data handling and interpretation in the future. The combination of human creativity and intellect with the strength and efficacy of AI-based analysis can provide a strong dynamic partnership that combines the strengths of both AI/ML and human intellect to advance knowledge and further accelerate and drive innovation in the fields of stem cell research and therapeutics development.

Overcoming barriers to stem cell therapy development through the use of advanced SysBioAI bioinformatics and the framework for its implementation

Bodies wear out, tissue thins and tears, organs stop functioning, and cells lose their biological way; as a result we become ill and disabled.²³⁷ This has been our reality, until we entered the era of regenerative medicine (RegMed) and its advanced therapies, determined to restore diminished physiology and quality of life (QoL).²³⁷ Stem cell therapies are a crucial component of the CGTs/ATMPs currently under development to address this important unmet medical need.¹⁸ Integrating advanced bioinformatics, SysBio, and AI, in analyzing complex biological data and advanced clinical trials is spurring promising next-generation patient-centric deep-medicine approaches.²²⁰ To develop these novel stem cell-based advanced therapies,¹⁸ rigorous clinical trials must be conducted to establish their safety and efficacy profile and to obtain regulatory approval for their clinical use, with systematic side studies to verify the underlying mechanism of action of the proposed therapeutic effect.⁵,⁷,⁹,¹⁸,²⁸

The growing medical need for regenerative therapies and adjunct growth potential of the regenerative medicine market (Figure 6A)²³⁸ is hampered by the low approval rates for new therapies and ATMPs/CGTs (Figure 6B). The likelihood of approval (LoA) rates of advanced clinical trials in relation to prior phase 1 trials are estimated to be less than 10% and 5%, respectively.^239–241 The integration of advanced bioinformatic tools, such as the use of advanced SysBio and AI, bears the potential to positively counteract the high failure rates (typically around 90% failure rate from early stage to regulatory approval) and long timelines (typically 12-15 years, estimated reduction 3-4 years through AI-enabled developments) that are common in the conventional development of drugs, stem cell therapeutics, and for the development of the complex ATMPs/CGTs in particular.²²⁰,^239–241

Figure 6.

Medical need and bottleneck for new therapy development in RegMed.

(A) Regenerative medicine market and adjunct medical need for RegMed therapies. The global RegMed market size was valued at around USD 30 billion in 2023 and is expected to grow at a compound annual growth rate (CAGR) of around 16.79% from 2024 to 2030. RegMed entails a collection of techniques and technologies, including stem cell technologies, that addresses the high medical need to restore lost organ and tissue function to restore patient health, mobility, physical and mental well-being, and overall quality of life (QoL). The descriptive data were extracted from the website below, Grand-View-Research-Report: https://www.grandviewresearch.com/industry-analysis/regenerative-medicine-market#. (B) Bottleneck for new therapies: the low likelihood of approval from phase 1 studies. Individual likelihood of approval (LoA) from phase 1 (%) for the development of conventional therapeutics is depicted according to different classes of clinical indications, including the average LoA of <10% reported by Mullard et al. in 2016 (the average LoA of <10% is depicted here as reported by Mullard, while only a representative selection of indications is shown, which may have a particular need for regenerative therapeutics). An estimated lower average LoA of <5% for the development of more complex biological advanced therapy medicinal products (ATMPs) and cell and gene therapies (CGTs) is shown, based on the lower approval rates reported for ATMPs/CGTs. The highest LoAs for therapeutics are reported in infectious disease and hematology studies with approximately 18%-25% (data extracted from Mullard et al. 2016). Similarly, higher success rates are also typically achieved for ATMPs/CGTs in the hematology and oncology field, for example: for the development of various CAR-T therapeutics.

Thus, to pave the way for faster and more efficient access to novel therapies, we here illustrate our “Iterative Circle of Refined Clinical Translation” (Figure 7A-C), including a framework for its implementation supported by SysBioAI analysis tools.¹² This iterative concept identifies at least three major needs to achieve better success rates in ATMPs/CGTs development: (1) Better use of open data and advanced data analysis tools (eg, integrated use of SysBio and AI), (2) Improved clinical trial design with option for iterative adjustments (eg, inclusion of exploratory biomarkers and informative secondary endpoints in addition to the primary endpoint), and (3) Iterative advancements/adjustment in biomedical engineering of product properties, clinical delivery and patient stratification, in case the study fails to meet its primary endpoints, but shows safety and promising efficacy in secondary endpoints.

Figure 7.

The iterative circle of refined clinical translation.

(A) Insufficient performance with conventional approaches mandates adjustments to yield faster and more efficient access to novel therapies, including: (1) better use of open data and advanced bioinformatics, (2) improved trials design with iterative adjustments, and (3) advancements in biomedical engineering to improve product properties and clinical delivery. (B) Measures to support better use of open data and advanced bioinformatics, such as SysBio and AI approaches. (C) Measures for improved clinical trial design, including stronger focus on secondary endpoints and exploratory biomarkers to support iterative adjustments in clinical trials design and treatment concepts for qualified restratification of patients according to therapy responder (green) and non-responder (red) status and improved survival.

Typical limitations to the implementation of such iterative learning approaches in clinical translation include: (1) Limitations in funding for complex studies, requiring suitable funding tools, such as larger EU-funded grants; (2) Lack of experience, training, planning and execution of advanced SysBioAI analysis in clinical trials; and (3) Resulting from points 1 and 2—General shortcomings in availability and sufficient quality of the primary data (as outlined in section advanced bioinformatics: data generation and analysis with SysBio and AI), data incompleteness and bias, need appropriate training datasets and batch correction, but also challenges resulting from model interpretation. However, the limitations in point 3 are mainly technical problems that can be overcome by improvements in points 1 and 2. Accordingly, the integrated use of open data and advanced data analysis tools, such as SysBioAI, is a key element often still lacking in many clinical trials and developmental approaches (Point 1 in Figure 7A). Its increased use could promote a better understanding and harnessing of the wealth of data (Figure 7B), which may be crucial to eventually overcoming existing barriers to effective clinical translation.^13–16

Many cell therapy trials still yield contradictory efficacy results, at least in part due to the immune safety and efficacy challenges in their preclinical assessment and progression to mature cell therapies, such as CAR-T and MSCs.⁸,¹⁰,²²⁷,²⁴²,²⁴³ Intriguingly, CAR-T cell therapy trials typically have a well-defined goal of targeting and eradicating tumor cells in the oncological setting, or the autoreactive B-cells/plasma cells in the autoimmune setting,¹⁸,²⁰,²¹,^23–26 which facilitates the development of adequate potency assays and its analysis.²⁰,²⁶ In contrast, MSCs are utilized in a multitude of degenerative and inflammatory disorders, where the exact targets of therapy response and their preclinical/clinical potency assessment are often yet to be defined.²⁰⁴,^243–248 Thus, in the case of MSCs, the medical and regulatory need to find an “assay matrix,” “surrogate markers,” or “patient biomarkers” that are indicative of therapy response and its MoA in patients is paramount.²⁴³,^248–253 Novel profiling technologies combined with SysBioAI-analysis may allow rapid and adequate patient monitoring (Section advanced bioinformatics: data generation and analysis with SysBio and AI).

Examples include in vitro assays for down-modulation of characteristic pro-inflammatory mediators or donor cell subsets in mixed lymphocyte reactions (MLRs) and other similar assays, or modulation of cellular properties, such as apoptosis, that correlate with functional response in patients.²⁴⁸,²⁵⁴ Such in vitro markers should correlate with adjunct in vivo modulation of similar mediators and cell subsets in patients,²⁴⁸,²⁵⁴ for example skewing of proinflammatory activated T-cells and M1 myeloid cells into anti-inflammatory Tregs and M2 macrophages, with adjunct downmodulation of characteristic soluble mediators, such as CRP and IL-6. Given their clinical relevance, such biomarker and potency assays should be done by certified laboratories, to warrant the timeliness, robustness, and validity of such measurements.¹¹⁴,^125–129 Importantly, SysBioAI supported analysis may be of advantage for the frequently required rapid and cost-effective decision-making processes in such a translational setting.²²⁷,²⁴²

The MSCs have a multitude of properties and effector pathways that modulate responder immune cell populations in the host to collectively mitigate pathology.^249–253 In addition, MSCs are often tested in advanced or refractory conditions (eg, steroid-resistant GvHD),²⁰⁴ where the host of pathobiology is versatile.²⁵⁵ Accordingly, one of the solutions that needs to be identified is the preemptive stratification of the patients who are likely to respond to MSC therapy.³⁰,¹⁸⁰,²⁰⁴,^255–257 This may include similar approaches as done in therapy for Her2-positive breast cancer patients,^258–260 which are selected for (immune)-therapy targeting Her2, while Her2 negative patients are treated otherwise. Thus, prior definition and subsequent assessment/screening of suitable patient inclusion/exclusion criteria with SysBioAI approaches could be a valuable asset to yield improved clinical trial outcomes.¹² The assessment of predefined patient parameters with subsequent analysis employing SysBioAI tools could provide the necessary accuracy and speed need for rapid clinical decision making, as already done in the assessment of histopathology, radiology, or in the kidney transplant setting to monitor potential rejection of allografts with an online multimodal diagnostic algorithm the “Banff Automation System.”¹²,^261–271

Considering MSC therapy, a practical example for inclusion/exclusion criteria may entail patient stratification according to their immune-aging status, with cutoffs for patient immune-age-related biomarkers.²⁷² MSC therapy for aGvHD seems to be more effective in children then adults,²⁷³ while severe COVID-19 patients of advanced age appear to be refractory to MSC treatment (patients below the age of 60 years appeared to respond better than patients above 60 years of age).³¹ This disparate outcome may be related to the low quantity or entire lack or of suitable “host responder cells” that need to be present in the patient to respond to the application of MSC therapeutics, for example by being reprogrammed into beneficial anti-inflammatory repair cells.³⁰,¹¹⁶,²⁴⁸,²⁵⁴

In line with this argument, host biomarkers of potential therapy response to MSC applications could include the identification of certain biomarkers or circulating immune cell subsets in the patient indicative of “immunoaging” before and after therapy (eg, soluble aging markers such as GDF5 or circulating TEMRA T cells), or the induction of distinct immune cell subsets, that are indicative of therapy response (eg, Tregs and M2 myeloid cells macrophages).¹²⁸,¹²⁹ Either way, any SysBioAI-supported monitoring and evaluation of patient suitability for MSC treatment may require highly automated manufacturing, with suitable automated online-assessment of the required parameters during the manufacturing and release process, in conjunction with the assessment of in vivo parameters before and after therapy, to evaluate treatment safety and efficacy. This will require suitable technology development, standardization and regulation, to allow its comparable use in multi-centric/national clinical trials, to eventually demonstrate MSC product safety and efficacy.¹¹⁴,^125–129

Importantly, negative trial results are less likely to be published than positive outcomes, which confounds meta-analysis based only on published outcomes.^274–276 This leads to a substantial loss of information considering non-published failed studies, and thus, loss in translational efficacy (Figure 7). The consequences are fewer learning outcomes from failed studies, waste of precious resources, and overall delayed access to novel therapies, with the financial loss carried by society as a whole. Thus, better use of advanced bioinformatics tools is important for an improved integrated understanding of large volumes of data, as is assessment, publishing, and access to information per se. This points out the role of “open data” policies and publishing policies for publicly and privately funded clinical studies alike to make required data more accessible. Eventually, one of the most crucial points remains to acquire suitable information through improved iterative clinical trial design (Figure 7C).

Conclusions, limitations, and outlook

Improvements in analytical capacity are crucial to overcoming existing barriers to effective clinical translation in stem cell therapy development and deployment as mature, approved clinical products. However, integrating SysBioAI-based approaches in analyzing biological data and clinical trials goes well beyond expanding analytical capabilities. The increasing synergy of SysBioAI actually supports a much more integrated understanding of complex multilayered information on multiple parameters from different data sources, typical for stem cell research, therapeutics development, and adjunct clinical trials. This synergy also enables a more patient-centric view even in the large, randomized trials necessary for therapy approval.

Deep-profiling of cell therapies’ phenotypic and functional properties with adjunct assessment of multiple secondary endpoints and biomarkers in patients is crucial to provide the necessary database for integrated SysBioAI-assisted analysis. In turn, this effort assists to better understand the clinical safety profile, pharmacokinetics, and pharmacodynamics of these novel therapeutics. This mandates careful planning with an integrated readout and analysis concept. Typical limitations include shortcomings in availability and quality of primary data, data incompleteness and bias, with a need for batch correction and appropriate training datasets, but also challenges resulting from model interpretation. If data assessment is planned well in advance to avoid such limitations, iterative adjustments in clinical trial design and treatments concepts, based on the identification of favorable product and patient parameters with restratification of patients according to therapy responder and non-responder status (or more stringent parameters, such as survival), may be a suitable way to support evidence-based iterative improvements to therapy performance, and thus, successful outcomes and product development in the long-run.

Here, we provide the necessary background, explanations, and structure for non-experts in advanced bioinformatics and SysBioAI-based analysis to familiarize themselves with these important new developments and to plan the next steps to implement such technology in their manufacturing and clinical trial workflow.¹² Future discussion of interest may include more detailed discussion on related topics, such as: (1) Emerging technologies (including organoids and 3D-bioprinting single cell and spatial transcriptomics for capturing cellular heterogeneity, and integration of wearable devices or real-time monitoring technologies with AI for personalized therapy); (2) Advanced Considerations on Data Challenges (including handling of missing, incomplete, or noisy data, which are common in omics and clinical datasets, standardization issues in data collection from multicenter trials, ethical and logistical concerns regarding data sharing and open data policies, computational bottlenecks in multimodel AI analysis, risk of algorithmic bias and overfitting in small, diverse patient cohorts, and challenges in the interpretation of high-dimensional data from poorly annotated datasets); and (3) Considerations on Different Stem Cell Types (including how SysBioAI-analysis can be tailored for specific stem cell types, such as MSCs, iPSCs, and ESCs, which each have their unique challenges).

Author contributions

Guido Moll was invited to contribute the review article to the journal, first conceived, designed, and supervised the writing of the review article and the final version of the manuscript, and is the corresponding author. All listed authors listed have made a substantial and intellectual contribution to the work, contributed to the manuscript writing, revisions, read and approved the submitted version.

Supplementary material

Supplementary material is available at Stem Cells Translational Medicine online.

Funding

G.M.’s contributions were made possible by funding from the German Federal Ministry for Education and Research (BMBF) and German Research Foundation (DFG; projects Nephroprotection #394046635, subproject A03, as part of CRC 1365, and EXPAND-PD; CA2816/1-1 and IMMME) and through the BIH Center for Regenerative Therapies (BCRT) and Berlin-Brandenburg School for Regenerative Therapies (BSRT, GSC203), respectively, and in part by the European Union’s Horizon 2020 Research and Innovation Program under grant agreements No 733006 (PACE) and 779293 (HIPGEN) and 754995 (EU-TRAIN) and 101095635 (PROTO). J.N.U. and T.S. received funding from Charité and M.H.B. (G.M./I.H.). We acknowledge financial support from the Open Access Publication Fund of Charité Universitätsmedizin Berlin and the DFG. O.C.M.’s, D.L.M.F.’s, and I.S.F.’s contributions were made possible by The São Paulo Research Foundation (FAPESP 2018/18886-9, 2020/01688-0, and 2020/07069-0 to O.C.M. and 2020/16246-2 and 2023/133356-0 to D.L.M.F and 2023/07806-2 to I.S.F.) and the National Council for Scientific and Technological Development (CNPq) Brazil (Grant: 309482/2022-4 to OCM). J.N.U. was supported by the Coordination of Superior Level Staff Improvement under Academic Excellence Program (CAPES/PROEX; Ref. No. 88887.917898/2023-00) and the German Academic Exchange Service (DAAD; Ref. No. 91898528). O.R. was supported by grants from the Swedish Cancer Society and was a recipient of a Distinguished Professor Award from Karolinska Institutet.

Conflicts of interest

The authors declare that the research was conducted without any commercial or financial relationships that could potentially create a conflict of interest.