Artificial intelligence and the transformation of cell and gene therapy development

Machine learning shifts viral vector engineering from evolutionary modification of existing designs to generative design. Models predict a vector's tropism and immunogenicity from its capsid protein sequence, and optimize promoter selection, codon usage and regulatory elements to maximize transgene expression while limiting off-target effects; generative models can propose novel capsid variants with improved tissue specificity and lower immunogenicity without blind high-throughput screening.

The sequence of a chimeric antigen receptor determines its binding affinity to the target, its signal transduction strength, and ultimately the engineered cell's persistence and killing efficacy — which makes it tractable to sequence-based prediction. Machine learning models trained on large datasets of receptor sequences paired with in vitro functional outcomes can predict optimal amino acid sequences for the single-chain variable fragment, hinge and co-stimulatory domains, including the balance between high binding affinity and low tonic signalling that governs CAR-T safety. Deep learning frameworks using language-model embeddings and contrastive learning can design peptide binders directly from sequence.

Nucleotide delivery formulation is a domain where AI outperforms factorial design-of-experiments, because the relationships among lipid composition, particle morphology, encapsulation efficiency, biodistribution and toxicity are nonlinear. Generative chemistry frameworks design novel ionizable lipids with optimized pKa and biodegradability profiles, Bayesian optimization and active learning reduce the experiments needed to find high-performing formulations, and sequence-level features such as mRNA length, GC content and secondary structure can be modelled jointly with formulation variables to predict stability, endosomal escape and tissue targeting.

A Design-Build-Test-Learn loop integrates computational modelling directly into the construct engineering workflow, with in silico mutation or sequence generation at DNA, mRNA and protein levels followed by prediction of regulatory elements and CRISPR off-target effects for DNA, GC content and minimum free energy for mRNA, and stability and binding affinity for protein. Experimental results — transcriptomics, protein presence, toxicity scores — feed back through active learning so each design cycle is more accurate than the last, compressing timelines that conventionally run to years.

Patient digital twins integrate omics data (genomics, transcriptomics, proteomics, metabolomics), imaging and clinical records into a unified predictive profile that can simulate both efficacy and safety of a cell or gene therapy. For CAR-T this means predicting in vivo persistence, proliferation and tumour-killing kinetics alongside the risk of cytokine release syndrome and immune effector cell-associated neurotoxicity syndrome; the CART-GPT framework interprets neurotoxicity and therapeutic outcomes from T-cell information. For gene therapy, digital twins simulate vector biodistribution, transgene kinetics and immune response from patient-specific anatomical and molecular data, supporting prediction of therapeutic windows and dose optimization.

Predictive risk scoring before infusion changes clinical management rather than just prognosis: machine learning models analyzing baseline patient omics such as inflammatory markers and cell repertoire features can assign a toxicity risk score before conditioning therapy, allowing clinicians to tailor conditioning regimens, choose bridging therapies or pre-emptively administer mitigating drugs. AI-enriched patient selection also raises the statistical power of the small trials characteristic of cell and gene therapy, supports adaptive designs, and enables virtual or external control arms built from historical data.

A manufacturing digital twin is a real-time virtual replica of a physical process such as a bioreactor run, fed continuously by biosensors, at-line analyzers and process control systems. It permits thousands of what-if simulations before a physical run to identify optimal parameters for yield and critical quality attributes, predicts quality attributes such as cell viability and product potency before the process completes so a failing batch can be rescued rather than lost, and can autonomously adjust equipment such as nutrient flow rates when it detects deviation — with reinforcement learning applied to that optimization.

Machine learning applied to historical batch data, including initial cell quality, donor characteristics and media components, can predict final cell yield and optimal harvest timing in CAR-T manufacturing, while equipment sensor data supports predictive maintenance that forestalls unplanned downtime. For autologous therapies the supply chain is inseparable from patient safety, and digital twins tracking the vein-to-vein journey monitor location, temperature and integrity to predict and prevent deviations, maintaining chain of identity.

A structural obstacle in biomanufacturing is that data arrive from multiple software systems with non-harmonized structures, which blocks trending and multivariate or meta-analysis. AI can harmonize data presentation across those disparate sources while preserving data integrity as GMP requires, enabling automated report generation and continuous evaluation of analytical method performance rather than the static snapshot that traditional single-timepoint validation provides. Training such models is itself constrained in cell and gene therapy by limited prior knowledge, absence of platform products, small batch numbers and few marketed products.

AI extends post-market surveillance for cell and gene therapies, where adverse events can be severe and may appear years after infusion. Natural language processing automatically processes adverse event reports, patient narratives and unstructured electronic health record and registry data, while machine learning detects signals across EHRs, registries, manufacturing data and patient-reported outcomes more sensitively than traditional statistical methods for patterns confined to patient subgroups or correlated with a manufacturing deviation. Linking manufacturing digital twin data to long-term outcomes can establish whether a bioreactor deviation years earlier correlates statistically with a late-onset adverse event, closing the quality loop well beyond product release.

Regulators are building specific frameworks for AI in this space rather than relying on general software guidance: the FDA's Framework for Regulatory Advanced Manufacturing Evaluation (FRAME) initiative assesses AI-powered manufacturing technologies, its 2024 cross-center strategy covers AI across the product lifecycle with CBER issuing preliminary guidelines emphasizing risk management for AI in biologics production, and separate FDA guidance addresses standards for submitting AI-derived data so predictions are trustworthy, traceable and understandable to reviewers. The EMA continues to adapt the distinct EU pathway for Advanced Therapy Medicinal Products, and WHO documents on regulatory convergence of cell and gene therapy products provide a framework for global harmonization.

Data scarcity is the fundamental bottleneck for AI in cell and gene therapy, not algorithmic capability: patient populations are small, data sit siloed across institutions, and collection, annotation and storage lack standardization. Gene therapy is worse affected, because rare target diseases and long follow-up requirements compound scarcity while heterogeneity in vector platforms, patient populations and clinical endpoints obstructs aggregation and generalizability. Real-world data is widely used to compensate, but differing national personal-data protection regimes limit how much adequately qualified data can be collected.

Regulatory acceptance of AI in this field is constrained less by absent rules than by absent shared definitions. Most national regulatory authorities adopt a horizontal AI definition under which the highest level of risk management applies, which may not match a sponsor's own risk assessment, and there is no harmonized terminology distinguishing explainability, understandability and interpretability for medical products — leaving requirements ambiguous and risk management difficult to standardize. The black-box character of deep learning models compounds this, as does the risk that non-diverse training data produces therapies less effective or riskier for particular demographic groups, so traceability, transparency and post-market performance monitoring carry particular weight.

Realizing AI-enabled cell and gene therapy depends on infrastructure and people as much as models: intelligent manufacturing requires substantial investment in new digital infrastructure and a workforce combining biomanufacturing expertise with data science skills, including upskilling within national regulatory authorities. Compliance with HIPAA in the United States and GDPR in Europe requires careful handling of protected health information, de-identification and data-sharing practice, while regulators increasingly expect transparency about data provenance, preprocessing, bias mitigation and version control for AI systems used in development or manufacturing.