Strategies for Developing Mechanism of Action and Potency Assays for AI-Generated Advanced Therapy Medicinal Products (ATMPs)

Artificial intelligence (AI), ranging from protein language models and structure-aware diffusion networks to active-learning wet-lab loops, is reshaping discovery and development across Advanced Therapy Medicinal Products (ATMPs). This review synthesizes how AI is applied to (i) identify and optimize modality-specific leads, (ii) engineer delivery systems, (iii) predict safety and developability, and (iv) accelerate translation under evolving regulatory frameworks. Modality-focused vignettes highlight exemplar programs and emerging companies at the frontier.

What are Advanced Therapy Medicinal Products (ATMPs)?

ATMPs are medicinal products for human use based on genes, cells, or tissues that have undergone substantial manipulation to achieve a therapeutic, diagnostic, or preventive purpose. As classified by regulatory agencies such as the European Medicines Agency (EMA) and the U.S. Food and Drug Administration (FDA), ATMPs fall into three categories.

Gene Therapy Medicinal Products (GTMPs): GTMPs involve the insertion, alteration, or removal of genetic material within patient cells to treat diseases such as genetic disorders, cancer, or other conditions. Their active substance consists of recombinant nucleic acids designed to regulate, repair, replace, add, or delete a genetic sequence. The therapeutic effect is directly linked to the nucleic acid sequence itself or to the product derived from its expression. Delivery can be mediated by viral or non-viral vectors, including genetically modified viruses, lipid- or polymer-based nanoparticles, or engineered cells, engineered to target specific tissues while minimizing safety risks. In line with EMA guidelines (EMA/CAT/80183/2014), vectors are modified to eliminate genes associated with virulence, pathogenicity, immunotoxicity, or replication. GTMPs may have prophylactic, diagnostic, or therapeutic effects, with notable examples including CAR-T cell therapies and CRISPR-based genome editing platforms.

Somatic-Cell Therapy Medicinal Products (sCTMPs): sCTMPs consist of or contain cells or tissues that have been substantially manipulated to alter their biological characteristics, physiological functions, or structural properties relevant to the intended clinical. They may also include cells or tissues not intended to perform the same function in the donor and recipient. These products are designed to restore or modify biological functions and are applied in a wide range of diseases. For instance, mesenchymal stem cell (MSC)-based therapies are under investigation for inflammatory disorders, autoimmune diseases, and regenerative applications. Another example is Vertex’s cell therapy (zimislecel) for type 1 diabetes (T1D).

Tissue-Engineered Products (TEPs): TEPs (also referred to as Tissue-Engineered Medicines, TEMs) consist of or contain engineered cells or tissues intended to regenerate, repair, or replace human tissue. These products may be of human or animal origin, or a combination of both, and can include additional substances such as biomolecules, biomaterials, chemicals, scaffolds, or matrices to support tissue function. Examples include bioengineered corneal implants for ocular repair and artificial skin grafts for wound healing.

Each category poses unique challenges in standardization, quality control, and navigation of complex regulatory pathways.

Why AI Matters for ATMPs

ATMPs, spanning biologics, gene therapies, and cell therapies, occupy vast design spaces (10^n sequences/edits/circuits) and exhibit complex structure–function relationships while facing strict CMC and regulatory constraints. AI foundation models for biology and protein design are increasingly integrated into design-build-test-learn (DBTL) cycles and wet-lab workflows, as demonstrated by companies like Chai Discovery and Alloy Therapeutics, and academic groups such as the Baker Lab.

These models leverage sequence, structure, and functional data to generate and optimize candidates in silico before experimental validation, streamlining discovery. Reinforcement learning, generative modeling, and active-learning feedback loops now enable iterative refinement, accelerating therapeutic development. Companies such as Absci and BigHat have advanced AI-designed antibodies into preclinical and early clinical stages. Beyond optimization, generative AI has produced fully de novo antibodies and proteins, offering unprecedented diversity for drug pipelines.

Parallel advances are occurring in viral gene delivery, where AI systematically engineers adeno-associated virus (AAV) capsids for improved tropism, manufacturability, and immune evasion, reducing the search space far beyond traditional random screening.

AI Applications Across Key ATMP Modalities

Gene Therapy

Viral vectors: Vector design and engineering are being transformed by machine learning (ML) and computational modeling to improve safety, specificity, and efficacy of gene therapies. In capsid engineering, ML models analyze sequence–structure–function relationships to predict tropism, immune evasion, and transduction efficiency, enabling the design of safer and more targeted viral vectors such as AAV and lentivirus; Voyager Therapeutics, for example, applies AI to engineer AAV capsids with enhanced CNS delivery. Similarly, algorithms are used to optimize promoter and enhancer combinations for tissue-specific expression, as seen in Passage Bio’s AI-driven selection strategies for CNS therapies. Delivery optimization is further guided by physiologically based pharmacokinetic (PBPK) models, which predict ideal administration routes (e.g., intravenous, intraocular) and dosing strategies based on preclinical and patient-specific data. Safety considerations are also addressed through ML, with sequence-based models predicting genomic integration sites and assessing risks of insertional mutagenesis, while immunoinformatics tools identify potential immune responses against both the vector and the transgene. Together, these approaches streamline vector development and reduce risks in clinical translation.

Antisense Oligonucleotides (ASOs) & Small Interfering RNAs (siRNAs): Artificial intelligence (AI) and machine learning (ML) are driving advances in the design and development of antisense oligonucleotides (ASOs) and small interfering RNAs (siRNAs). In sequence design and optimization, AI algorithms generate ASO/siRNA sequences with high target mRNA binding affinity, reduced off-target interactions, and improved nuclease resistance; for instance, Ionis Pharmaceuticals employs AI platforms such as WingX to refine ASO design. Delivery system engineering is enhanced through ML models that predict the performance of chemical modifications and formulations, including GalNAc conjugates for liver targeting and lipid nanoparticles for broader tissue-specific delivery, as demonstrated by Alnylam Pharmaceuticals in optimizing GalNAc-conjugate chemistry. Safety is addressed by AI-driven toxicity prediction, where sequence-based and structural models evaluate off-target effects and immunostimulatory risks, such as Toll-like receptor activation; Pfizer, for example, applies AI to screen candidates for potential toxicities early in development. Additionally, mechanism of action (MoA) prediction uses computational approaches to assess the downstream consequences of gene knockdown across complex biological pathways, improving therapeutic design and reducing translational risks.

Gene Editing Therapies: CRISPR and Based Editing Technologies: Artificial intelligence (AI) is reshaping CRISPR-based genome editing, from guide RNA (gRNA) design to delivery optimization. In gRNA design, AI integrates sequence features, chromatin accessibility, and structural modeling to maximize efficiency and minimize off-target effects. Companies like Desktop Genetics and CRISPR Therapeutics use these tools, while ML models trained on CIRCLE-seq, GUIDE-seq, and DISCOVER-seq data enable genome-wide off-target prediction. Google’s DeepVariant adds sequencing-based error correction.

Beyond CRISPR-Cas9, base and prime editing are advancing as precision tools. AI supports design of cytosine/adenine base editors and prime editing guide RNAs (pegRNAs), improving efficiency and minimizing bystander edits. Beam Therapeutics and Prime Medicine lead in this space. Epigenome editing with dCas9 fusions is enhanced by AI, which predicts chromatin interactions and long-term expression outcomes.

For homology-directed repair (HDR), AI aids donor template design, including ssDNA, dsDNA, and AAV donors. It also drives next-generation CRISPR systems such as Cas13 for RNA targeting and CRISPR-associated transposases for large-fragment edits.

Delivery remains central. AI accelerates engineering of viral vectors (e.g., AAVs with reduced immunogenicity, expanded tropism) and lipid nanoparticles (LNPs) with greater stability and tissue specificity. Intellia Therapeutics applies AI to LNPs for in vivo therapies, while multi-omics integration aligns delivery vehicles with patient-specific profiles.

Together, the convergence of AI with CRISPR, base editing, prime editing, and epigenome modulation is paving the way for next-generation precision gene therapies with unprecedented accuracy and therapeutic potential.

Cell Therapy  

Machine learning (ML) is increasingly applied across the cell therapy pipeline, from target selection to product manufacturing and quality control. In target antigen selection, ML integrates data from RNA-seq and mass spectrometry to identify tumor-specific antigens with minimal off-tumor toxicity, as demonstrated by Carisma Therapeutics in selecting neoantigens for personalized TIL therapies. For cell product engineering, computational modeling optimizes CAR and TCR receptor designs for affinity, signaling, and persistence, with Century Therapeutics using AI to develop potent allogeneic CAR constructs. Gene editing is also enhanced by ML, which predicts CRISPR/Cas9 outcomes—including on-target efficiency and off-target risks—supporting safer modifications such as PD-1 knockouts; CRISPR Therapeutics, for example, leverages AI collaborations to refine guide RNA design. Manufacturing is streamlined by predictive analytics that monitor real-time bioreactor conditions (temperature, pH, metabolites) to forecast cell expansion, viability, and functionality, enabling closed-loop process control, as seen in Lineage Cell Therapeutics’ AI-driven systems. Finally, quality control is strengthened by AI-enabled assays for rapid, sensitive characterization of identity, purity, potency, and safety, with companies like Neochromosome pioneering comprehensive cell product QC.

Tissue Engineering

Artificial intelligence is increasingly being applied to the development of tissue-engineered products (TEPs), where it supports both design and functional evaluation of complex cellular–biomaterial systems. Machine learning models can integrate genomic, transcriptomic, and biomaterial data to predict how engineered cells interact with scaffolds, extracellular matrix components, and growth factors, guiding the rational design of constructs for tissue repair or regeneration. In scaffold engineering, AI-driven generative models help optimize pore size, stiffness, and biochemical modifications to enhance cell adhesion, differentiation, and vascularization. Computer vision and deep learning enable automated analysis of histology, live-cell imaging, and biomechanical testing, providing high-content characterization of engineered tissues while reducing manual bias. Additionally, AI assists in developing potency and mechanism-of-action assays by predicting which molecular markers or functional readouts—such as collagen deposition in artificial skin grafts or keratocyte activity in corneal implants—best correlate with clinical outcomes. Looking ahead, digital twins and AI-guided microphysiological systems promise to simulate tissue integration and long-term functionality in silico, accelerating translation of TEPs into safe, reproducible, and regulatory-compliant therapies.

From In-Silico Molecules to Mechanistic Assays: Hybrid AI Strategies for Developing MOA and Potency Assays for ATMPs

AI platforms are beginning to not only discover but also design ATMPs. These AI designed molecules emerge from in-silico fitness landscapes that are largely disconnected from classical biology. Consequently, the first technical challenge is to reverse-engineer the in-silico MOA into a set of in-vitro and ex-vivo assays that regulators will accept as “mechanistically relevant” and “quantitative”. Hence, following hybrid strategies are required:

• Reported Assays: experimental workflows already used for human-approved ATMPs.
• Speculative Assays: AI-driven assay-selection engines that continuously update the experimental plan in silico as new data are generated.

Reported MOA and Potency Assays for ATMPs

Cell-Based Assays: Cell-based assays represent the cornerstone of potency assessment, as they capture the complexity of the entire biological system and provide functional characterization directly aligned with the mechanism of action (MOA). These assays are tailored to the therapeutic modality. For cytotoxic therapies such as CAR-T cells, cytotoxicity assays are employed using luciferase-based bioluminescence platforms or impedance-based technologies like xCelligence to measure target cell killing in real time. For immunomodulators, cytokine release assays such as Luminex multiplexing are used to quantify the secretion of key cytokines, including IFN-γ, IL-2, and IL-6, upon target engagement. In the case of gene therapies, transduction efficiency assays using flow cytometry for reporter genes are combined with functional genomic assays such as qPCR/ddPCR for transgene expression and RNA-seq for transcriptomic profiling. The development strategy for these assays involves several critical considerations: selecting a biologically relevant cell line, establishing a robust co-culture system, optimizing the effector-to-target ratio, and identifying the most informative readout that correlates closely with clinical effect. This strategic framework ensures that potency assays provide meaningful and reproducible measures of therapeutic activity.

In Vitro Binding Studies (Biophysical Characterization): Surface Plasmon Resonance (SPR) and Bio-Layer Interferometry (BLI) are considered gold-standard techniques for quantifying binding kinetics, including association rate (ka), dissociation rate (kd), and equilibrium dissociation constant (KD), between an ATMP’s active moiety—such as a soluble CAR receptor or engineered protein—and its target antigen. These methods offer label-free, real-time measurements that provide detailed insights into both affinity and specificity, making them invaluable tools for characterizing molecular interactions.

ELISA-Based Assays: ELISA is a versatile technique that detects and quantifies specific biomolecules, such as cytokines and surface antigens, and is readily adaptable for potency assessment of advanced therapy medicinal products (ATMPs). Several methodological approaches have been reported for tailoring ELISA to different applications. Sandwich ELISA, which employs both capture and detection antibodies, provides high-sensitivity quantification of ATMP-induced biomarkers, such as IL-2 release from engineered T cells. Competitive ELISA is particularly useful for analyzing small-molecule ATMPs, including AI-designed peptides, by measuring their ability to compete with labeled analogs. More recently, electrochemiluminescence (ECL) ELISA has been introduced to improve detection sensitivity, enabling accurate measurement of low-abundance targets that may otherwise be difficult to quantify.

Speculative and AI-Augmented Strategies

Artificial intelligence is increasingly integral to assay selection and optimization for advanced therapy medicinal products (ATMPs). In target identification and pathway analysis, AI models can rapidly mine genomic, proteomic, and clinical datasets to predict primary drug targets, potential off-targets, and the most relevant physiological pathways. Machine learning helps prioritize which molecules, markers, or functions should be assayed; for instance, deep learning can screen for surface markers that define cell therapy product identity or predict outcomes of gene-editing events. AI also enables virtual assay simulation, where digital twins or in silico models are used to simulate product–target interactions, predict binding affinities, and optimize assay parameters before conducting experiments at the bench. As in vitro and clinical datasets expand, predictive potency modeling becomes possible, allowing AI to correlate specific assay readouts, such as cytokine release or cell lysis—with clinical potency. This improves the selection of surrogate markers for release testing and reduces dependence on costly and time-consuming functional assays. An AI-based step-wise assay development framework is shown in Table 1.

Table 1: AI-based step-wise assay development framework

AI-Mediated MOA De-Convolution: The AI-mediated MOA de-convolution approach involves applying multi-omics analyses—including transcriptomics, secretomics, and phospho-proteomics—on AI-generated ATMP-treated versus untreated target cells. To interpret these data, weighted gene co-expression network analysis (WGCNA) is used to infer signaling modules and identify potential mechanistic pathways. As a speculative extension, a graph neural network (GNN), trained on more than 10⁴ public MOA pathways, could be employed to transform the omics snapshot into a ranked list of mechanistic hypotheses, such as “increased granzyme-B exocytosis via LFA-1 conformational change.” Each generated hypothesis would then be annotated with a regulatory confidence score, derived from historical FDA reviews of similar mechanisms, thereby integrating computational inference with regulatory precedent.

AI in In Vitro and Cell-Based Assay Development: In the development of in vitro and cell-based assays, AI contributes significantly to both precision and efficiency. AI-assisted image analysis enhances the quantification of differentiation, cytotoxicity, and proliferation in high-content screening, while predictive analytics help determine which combinations of in vitro, ELISA, and cell-based outcomes most accurately represent clinical efficacy. This informs the design of matrix approaches and supports the identification of the most informative primary, secondary, and orthogonal assays. Importantly, regulatory precedents already guide assay development across cell and gene therapies: CAR-T cells are often assessed by IFN-γ release and CD107a degranulation, mesenchymal stem cells (MSCs) by tri-lineage differentiation and IDO activity, and NK cells by cytotoxicity assays such as Calcein-AM release. AI systems can build on these precedents to streamline assay strategy design. For example, reinforcement learning (RL) agents can be trained with a reward function balancing regulatory acceptance probability, cost, and time to select among canonical assay archetypes stored in curated knowledge bases. By leveraging transfer learning from prior assay qualification packages, including public FDA briefing documents, these systems can predict the minimal viable panel of assays that meets ICH Q2(R1) requirements.

Finally, AI also plays a pivotal role in quality and manufacturing optimization by applying machine learning algorithms to reduce batch variability. Through continuous monitoring and prediction of cell phenotypes, gene expression profiles, and other quality attributes, AI strengthens in-process controls and ensures greater assay reliability. The practical outcome of these AI-driven approaches is a ranked shortlist of assays tailored to regulatory precedent and efficiency—for example, prioritizing IFN-γ ELISA as a low-cost, high-precedent option, real-time cytotoxicity using IncuCyte live-cell imaging as a medium-cost assay that captures kinetics, and phospho-flow for STAT-3 activation as a high-specificity mechanistic assay with lower regulatory precedent. Together, these innovations demonstrate how AI is not only enhancing assay development but also aligning it more closely with regulatory expectations, clinical relevance, and manufacturing robustness.

AI-Guided Optimization of ELISA Format: AI-guided approaches are increasingly being applied to optimize ELISA design and performance. Protein language models such as ESM-2 can predict which single-point mutations in the scFv are likely to reduce off-target binding without compromising affinity. These predicted mutants can then be expressed in Expi293 cells and incorporated directly into the same ELISA plate as specificity controls, creating an in-plate specificity panel. This strategy minimizes the need for multiple standalone surface plasmon resonance (SPR) experiments, streamlining assay development while improving both specificity and efficiency.

AI-Augmented Potency Assay Design and 3D Micro-Physiological Systems: Speculative applications of AI in assay design highlight its potential to bridge in vitro and in vivo translation. Active learning algorithms can propose the minimal set of donor PBMC genotypes—such as HLA-A2⁺ or the CD16-V158F high-affinity variant—whose inclusion maximizes the in vitro–to–in vivo translation score, calculated from retrospective clinical databases. Complementing this, digital twin models of assays, built as agent-based simulations calibrated with live-cell imaging data, can predict how adjustments in effector-to-target (E:T) ratios or cytokine priming conditions, such as IL-15 pulsing, shift the potency curve. This enables early exploration of design space prior to any wet-lab experiments. In parallel, 3D micro-physiological systems are emerging as powerful platforms to evaluate ATMPs. Patient-derived iPSC organoids, such as B-cell lymphoma organoids cultured in microfluidic chips, can be exposed to therapeutic candidates while AI-driven image analysis performs segmentation and tracking to quantify tumor regression and T-cell infiltration simultaneously. These systems provide critical bridging data between classical 2D cytotoxicity assays and in vivo mouse efficacy models, enhancing translational predictability while reducing reliance on animal studies.

Statistical Qualification and Continuous Learning: Statistical methods are essential for ensuring robustness and reproducibility in assay development and application. Design of Experiments (DoE) can be applied to systematically evaluate sources of variability, such as operator differences, reagent lot changes, and day-to-day fluctuations, often using a three-factor full factorial design to capture interactions among variables. Looking ahead, more speculative approaches envision integrating advanced statistical models with AI-driven assay systems. For instance, a Bayesian hierarchical framework could pool data across multiple assay modalities—such as ELISA, SPR, and cell-based assays—to estimate the latent “true potency” of a product, complete with credible intervals. The resulting posterior distribution would then feed back into the AI design engine, enabling adaptive control of manufacturing parameters, such as vector copy number, to ensure that potency consistently remains within specification. This continuous learning loop would create a dynamic quality control system that evolves with accumulating data, strengthening both reliability and regulatory confidence.

Conclusion

AI-generated ATMPs shift assay development from the paradigm of “assay follows biology” to one in which “biology and assay co-evolve in silico.” By integrating established wet-lab modules (ELISA, SPR, cytotoxicity assays, organoids) with speculative AI-driven engines for MOA deconvolution and assay optimization, developers can shorten timelines, reduce animal use, and still meet the stringent evidentiary standards of modern regulators. The next milestone is to validate this hybrid workflow in prospective clinical programs—transforming AI-designed ATMPs into AI-qualified products.

 References

1. Guidelines relevant for advanced therapy medicinal products. European Medicines Agency. https://www.ema.europa.eu/en/human-regulatory-overview/advanced-therapy-medicinal-products-overview/guidelines-relevant-advanced-therapy-medicinal-products
2. Guidance for Industry Potency Tests for Cellular and Gene Therapy Products, FDA. https://www.fda.gov/downloads/BiologicsBloodVaccines/GuidanceComplianceRegulatoryInformation/Guidances/CellularandGeneTherapy/UCM243392.pdf
3. Advanced therapy medicinal products development – from guidelines to medicines in the market. Biotechnology Advances. 2025, 83:108612. https://doi.org/10.1016/j.biotechadv.2025.108612
4. Design-Build-Test-Learn: Impact of AI on the Synthetic Biology Process. Available from: https://www.ncbi.nlm.nih.gov/books/NBK614601/
5. Artificial intelligence-driven computational methods for antibody design and optimization. MAbs. 2025; 17:2528902. doi: 10.1080/19420862.2025.2528902.
6. Atomically accurate de novo design of antibodies with RFdiffusion. bioRxiv [Preprint]. 2025:2024.03.14.585103. doi: 10.1101/2024.03.14.585103.
7. Systematic multi-trait AAV capsid engineering for efficient gene delivery. Nat Commun. 2024; 15: 6602. https://doi.org/10.1038/s41467-024-50555-y.
8. PandaOmics: An AI-Driven Platform for Therapeutic Target and Biomarker. Journal of Chemical Information and Modeling. 2024;  Vol 64. https://doi.org/10.1021/acs.jcim.3c01619
9. Leveraging Artificial Intelligence to Expedite Antibody Design and Enhance Antibody-Antigen Interactions. Bioengineering (Basel). 2024; 11:185. doi: 10.3390/bioengineering11020185.
10. Revolutionizing oncology: the role of Artificial Intelligence (AI) as an antibody design, and optimization tools. Biomark Res. 2025; 13:52. doi: 10.1186/s40364-025-00764-4.
11. Advances in AAV capsid engineering: Integrating rational design, directed evolution and machine learning. Molecular Therapy. 2025; 33:1937-1945. https://doi.org/10.1016/j.ymthe.2025.03.056.
12. AAV Engineering for Improving Tropism to the Central Nervous System. Biology (Basel). 2023; 12:186. doi: 10.3390/biology12020186.
13. Artificial Intelligence-Based Approaches for AAV Vector Engineering. Advanced Science. 2025. https://doi.org/10.1002/advs.202411062
14. Revolution of AAV in Drug Discovery: From Delivery System to Clinical Application. J Med Virol. 2025; 97:e70447. doi: 10.1002/jmv.70447.
15. Structural basis for self-discrimination by neoantigen-specific TCRs. Nat Commun. 2024;15:2140. doi: 10.1038/s41467-024-46367-9.
16. Computational structural optimization enhances IL13Rα2 – B7-H3 tandem CAR T cells to overcome antigen-heterogeneity-mediated tumor escape. Molecular Therapy (published online 2025) DOI: 1016/j.ymthe.2025.07.044
17. Predicting T-cell quality during manufacturing through an artificial intelligence-based integrative multiomics analytical platform. Bioeng Transl Med. 2022; 7:e10282. doi: 10.1002/btm2.10282
18. Predicting T-cell quality during manufacturing through an artificial intelligence-based integrative multiomics analytical platform. Bioeng Transl Med. 2022; 7:e10282. doi: 10.1002/btm2.10282
19. Antisense Inhibition of Angiotensinogen With IONIS-AGT-L_Rx: Results of Phase 1 and Phase 2 Studies. JACC Basic Transl Sci. 2021;6:485-496. doi: 10.1016/j.jacbts.2021.04.004
20. Springer AD, Dowdy SF. GalNAc-siRNA Conjugates: Leading the Way for Delivery of RNAi Therapeutics. Nucleic Acid Ther. 2018; 28:109-118. doi: 10.1089/nat.2018.0736
21. Machine Learning for Toxicity Prediction Using Chemical Structures: Pillars for Success in the Real World. Chem Res Toxicol. 2025;38:759-807. doi: 10.1021/acs.chemrestox.5c00033.
22. Desktop Genetics. Per Med. 2016; 13:517-521. doi: 10.2217/pme-2016-0068.
23. Advancing genome editing with artificial intelligence: opportunities, challenges, and future direction. Front. Bioeng. Biotechnol. 2024;11:2023. https://doi.org/10.3389/fbioe.2023.1335901
24. Off-target predictions in CRISPR-Cas9 gene editing using deep learning. Bioinformatics, 2018; 34: i656–i663, https://doi.org/10.1093/bioinformatics/bty554
25. From bench to bedside: cutting-edge applications of base editing and prime editing in precision medicine.J Transl Med. 2024; 22:1133. doi: 1186/s12967-024-05957-3
26. Comparative analysis of assays to measure CAR T cell–mediated cytotoxicity. Nat Protoc. 2021; 16:1331–1342. doi: 1038/s41596-020-00467-0
27. Evaluation of CAR-T cell cytotoxicity: Real-time impedance-based analysis.Methods Cell Biol.  2022:167:81-98.doi: 10.1016/bs.mcb.2021.08.002.