Harnessing AI in Cell and Gene Therapy: Advances in CAR-T Therapies

Cell and gene therapies are at the forefront of biotech innovation, offering groundbreaking treatments like CAR-T cell immunotherapies, viral vector-based gene therapies, and CRISPR gene editing for previously intractable diseases. These advanced therapies are highly complex. CAR-T cell products are living cells engineered for each patient, gene therapies rely on delicate viral delivery systems, and CRISPR enables precise but potentially off-target genomic edits. Developing and delivering such therapies is expensive and data-intensive; for example, a single personalized CAR-T treatment can cost on the order of $400,0001.

To accelerate discovery and improve clinical outcomes, researchers are increasingly turning to artificial intelligence (AI) and machine learning (ML). AI can sift vast multi-modal datasets – from genomic and proteomic profiles to patient outcomes to finding patterns and optimize each stage of therapy development.

This article explores how AI is currently applied in the key domain of cell and gene therapy, i.e., CAR-T- highlighting state-of-the-art models and platforms. I also discuss illustrative metrics and integrated data sources (e.g. multi-omics analyses, clinical trial databases, real-world patient data) that have driven AI advances, and examine the challenges (regulatory, data quality, interoperability, organizational) that must be addressed. Finally, we outline what must happen for AI adoption to scale successfully in these therapy areas, and why the convergence of AI with advanced therapies is critical to the future of biotechnology.

AI use cases in CAR-T Cell Therapies

Chimeric antigen receptor T-cell (CAR-T) therapies have revolutionized treatment for heme malignancies like Multiple Myeloma, DLBCL etc., but these therapies face challenges in design, manufacturing, and patient management. AI is emerging as a powerful tool across the CAR-T value chain, from optimizing CAR designs in silico to improving manufacturing processes and personalized patient monitoring. Integrating AI can increase efficiency and lower costs in CAR-T development1, although it also introduces new considerations around data and validation.

Below we detail key AI use cases in CAR-T therapy:

1. Optimizing CAR Design and Function: Designing an effective CAR requires balancing multiple factors like antigen binding, T-cell activation strength, and safety (avoiding overactivation or “tonic” signaling). AI-driven platforms now assist researchers in this complex design space. For example, Wang et al. (2024) introduced CAR-Toner, a deep learning platform that predicts and optimizes CAR constructs to modulate tonic signaling2.
   • CAR-Toner computes a Positively Charged Patch (PCP) score for CAR protein sequences – a metric correlating with spontaneous (antigen-independent) CAR signaling – and helps identify designs that keep this tonic signal in an optimal range.2 This AI-guided approach allowed the team to define a “peak” range of tonic signaling that maximizes CAR-T cell fitness and longevity in vivo. By evaluating vast libraries of CAR variants in silico, AI tools like CAR-Toner accelerate the iterative improvement of CAR constructs, aiming to increase persistence of CAR-T cells while reducing exhaustion and toxicity.
   • Similarly, other researchers have applied ML to correlate CAR genetic features or marker profiles with therapeutic efficacy, which could inform the design of next-generation CAR-T cells with better tumor killing and safety profiles.1
2. Manufacturing Process Control: CAR-T manufacturing is a multi-step, patient-specific process that is difficult to scale. Early efforts to automate CAR-T production are now generating large datasets that AI can leverage to optimize bioprocessing.1
   • For instance, the EU-supported AIDPATH project is developing an AI-integrated, decentralized CAR-T manufacturing platform. One use case in AIDPATH is constructing a digital twin of the T-cell expansion bioreactor – a virtual model that uses mechanistic knowledge and real-time sensor data to simulate cell growth.1 By applying machine learning “soft sensors” to streaming bioreactor data (e.g. nutrient levels, metabolites, cell density), the system can forecast cell expansion over 1–2 days.
     • These predictions inform optimal harvest timing, i.e. when a culture likely reaches the target cell dose.
   • Another use case employs an ensemble of AI algorithms to monitor and control the expansion in real-time: data from multiple physical sensors are fused by soft-sensor algorithms (including statistical models and neural networks) to alert operators and adjust conditions proactively.1
   • Such AI-driven process control has the potential to improve product consistency (critical quality attributes) and reduce failures.
3. Patient Selection and Efficacy Prediction: Not all patients derive equal benefit from CAR-T, and severe side effects like cytokine release syndrome (CRS) can be life-threatening. AI is being applied to improve patient stratification and predict outcomes.
   • Researchers have trained ML models on clinical datasets to predict which patients are at high risk of adverse events like CRS or sepsis after CAR-T therapy.1
     • In one approach, models ingest patients’ pre-infusion biomarkers and other clinical parameters to forecast CRS severity, enabling clinicians to plan proactive interventions (some models for CRS risk stratification have achieved useful accuracy, as cited by Bedoya et al., 2020.3
     • On the efficacy side, AI can help identify the best candidates for CAR-T. Researchers have demonstrated an ML model for predicting which leukemia patients are most likely to achieve remission with CAR-T, based on features like tumor burden and immune cell profiles.4
   • While these approaches are still being validated, they point to a future where AI-driven decision support tools assist oncologists in matching CAR-T therapies to patients with the highest expected benefit.

Bottlenecks in adoption of AI in CAR-T development

Despite these advances, the integration of AI into CAR-T development is still nascent. The opportunities – smarter design, automated production, personalized treatment – are balanced by challenges such as data siloing, validation needs, and the complexity of biological systems. Nonetheless, early successes demonstrate that AI can enhance CAR-T therapy at every stage, from designing better CARs to ensuring patients get the right therapy at the right time. By converting data into actionable insight, AI may help CAR-T treatments become more effective, safe, and broadly accessible in the years ahead.

1. Robust Data Ecosystems: Since CAR-T and other advanced therapies are developed in limited patient populations, there is a critical need for solving the data limitation problem. This entails building consortia and data-sharing agreements where companies, academia, and healthcare institutions pool anonymized data on patients, vectors, cell products, and outcomes.
   • Initiatives like the upcoming cell & gene therapy data atlas or public-private partnerships (e.g. the NIH’s somatic cell genome editing program)5 can provide repositories of training data for AI models.
   • Additionally, implementing end-to-end digital data capture in every process step (from automated cell culture instruments to electronic lab notebooks in R&D to clinical trial electronic data capture) will ensure rich datasets are generated and stored in formats amenable to AI analysis.
   • Adopting common data standards and ontologies for cell and gene therapy (similar to how the genomics field has standard file formats) will improve interoperability. Ultimately, a “big data” foundation is needed; without it, scaling AI will hit a ceiling. Companies might also explore simulated data or AI-generated synthetic data to augment real datasets for model training, provided those synthetic datasets sufficiently reflect reality.
2. Validated Platforms and “Proof Points”: As few organizations have fully operational AI pipelines in advanced therapies today, each success becomes a proof point that can be scaled or transferred.
   • For example, if a company deploys an AI model in manufacturing that reduces batch failures by 30%, documenting and publishing that result (even as a case study) will build confidence industry-wide and likely prompt others to adopt similar approaches.
   • Open-source or commercial platforms that package AI methods for ease of use will also help with scaling.
   • We may see the emergence of standard AI software for, say, CAR-T data management and analysis – analogous to how flow cytometry or Next-Gen Sequencing got dedicated software.
   • Moreover, validation of AI tools through regulatory channels will greatly aid scaling: if the FDA were to approve a trial design that heavily used AI for patient selection, it sets a precedent that others can follow with more confidence. External validation in the form of regulatory green lights or peer-reviewed clinical outcomes will be key to convincing skeptics that AI is not just hype but rather a practical necessity in this field.
3. Regulatory Adaptation and Guidance: On the regulatory front, continued evolution of frameworks that accommodate AI is needed.
   • The FDA’s risk-based guidance is a start; as it solidifies, companies will have clearer goalposts for what is required to get AI-supported products approved.6 Successful scaling will likely involve close dialog with regulators – engaging them early with AI development plans, potentially using sandbox programs or adaptive licensing pathways for innovative approaches.
   • Regulators may need to adapt approval processes to be more iterative, recognizing that AI models can improve continuously. For instance, establishing protocols for life-cycle management of AI models (re-training with new data under oversight) will allow AI to get better over time rather than freezing a model at approval.
   • International regulatory harmonization on AI in therapeutics would also help companies scale solutions globally rather than reinventing the wheel country by country. In essence, a flexible yet rigorous regulatory environment – one that demands proof of safety and efficacy from AI but also understands its dynamic nature – is necessary. Regulatory sandboxes, qualification of AI as biomarkers or companion diagnostics, and guidance on AI transparency will all facilitate broader adoption.
4. Organizational Readiness and Culture: Scaling AI is as much about culture as tech. Organizations that treat data as a strategic asset and AI as a core capability (not an experimental side project) will leap ahead. This may involve leadership setting clear directives for “AI-first” approaches where applicable, allocating budget for digital transformation, and measuring R&D success in part by digital metrics.
   • Change management is crucial – staff should be brought on board with the vision that AI will augment, not replace, human expertise, and that upskilling will be provided. When people see AI making their jobs easier (e.g. automating tedious data analyses, providing insights they couldn’t get otherwise), they become champions for it. The end state is an organizational culture where decisions – whether in discovery, clinical development, or manufacturing – are data-driven, and AI is naturally incorporated alongside experimental results in decision-making processes. This will likely differentiate the biotech players who succeed in the next decade from those who lag.

Conclusion

In summary, for AI adoption to scale, data infrastructure, human infrastructure, and regulatory infrastructure all need to mature in tandem. Encouraging trends are already visible: multi-stakeholder consortia are tackling data standards, cross-functional expertise is increasingly valued in biotech hiring, and regulators are actively engaging with AI topics. If these trends continue, the industry will move from isolated AI pilot projects to broad, routine use of AI platforms in developing cell and gene therapies. The result will be faster R&D cycles, more predictable manufacturing, and therapies that are tailored with unprecedented precision to patient needs.

Sources:
[1] Bäckel, N., Hort, S., et.al., (2023). Elaborating the potential of Artificial Intelligence in automated CAR-T cell manufacturing. Frontiers in molecular medicine, 3, 1250508

[2] Qiu, S., Chen, J., et.al., (2024). CAR-Toner: an AI-driven approach for CAR tonic signaling prediction and optimization. Cell Research, 34(5), 386-388

[3] Bedoya, A. D., Futoma, J., Clement, M. E., Corey, K., Brajer, N., Lin, A., et al. (2020). Machine learning for early detection of sepsis: an internal and temporal validation study. Jamia Open 3 (2), 252–260. doi:10.1093/jamiaopen/ooaa006

[4] Liberini, V., Laudicella, R., Capozza, M., Huellner, M. W., Burger, I. A., Baldari, S., et al. (2021). The future of cancer diagnosis, treatment and surveillance: A systemic review on immunotherapy and immuno- pet radiotracers. Molecules 26 (26), 2201. doi:10.3390/molecules26082201

[5] https://commonfund.nih.gov/editing

[6] https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

About the author:

Renu Pandey, PhD is a Life sciences and Healthcare strategy consultant supporting pharmaceutical and biotech companies with corporate strategy, product commercialization, and market expansion for precision and advanced therapies e.g., Cell and Gene therapies, RNA and Radioligand therapies. She is deeply passionate about all things biotech and is particularly interested in the intersection of science and business. Renu earned her PhD in Biomedical Research from University of Delhi.