Cell and Gene Therapy AI
Cell and gene therapy turns sequence design, cell engineering, delivery, manufacturing, and clinical evidence into a single product problem. AI systems may support receptor design, CRISPR guide selection, AAV capsid engineering, LNP formulation triage, cell-state characterization, potency assays, release testing, and patient stratification. The decisive questions remain biological and regulatory: what was changed, where it went, what it did, how long it persisted, how it was manufactured, and whether the benefit-risk profile holds in patients.
Use this chapter to:
- Use AI in engineered cells, genome editing, vector design, potency assays, manufacturing analytics, and translational evidence for living medicines.
- Product identity, edit profile, delivery, potency, persistence, safety, manufacturability, and long-term follow-up define the evidence standard.
What is this field trying to solve? Use AI in engineered cells, genome editing, vector design, potency assays, manufacturing analytics, and translational evidence for living medicines.
What is the core idea? Product identity, edit profile, delivery, potency, persistence, safety, manufacturability, and long-term follow-up define the evidence standard.
What is the current state of the field? Guide design, off-target nomination, capsid engineering, and manufacturing analytics are useful; product-level safety and efficacy require direct evidence.
What do we know, and what remains open? Known reference points include CRISPR guide models, CRISPOR, GUIDE-seq, CIRCLE-seq, CHANGE-seq, base and prime editing, AAV capsid libraries, CAR-T guidance, and FDA gene-therapy documents. What remains open is whether model-ranked edits, vectors, receptors, or process decisions improve product-level safety and efficacy.
Why does this matter? Cell and gene therapy links genomics, immunology, vector biology, manufacturing, biomarkers, clinical trials, and long-term safety.
Introduction
Cell and gene therapy products sit at the most complex edge of therapeutic translation. FDA’s cellular and gene therapy page describes CBER oversight of cellular therapy products, human gene therapy products, and related devices, with gene therapy defined around modifying or manipulating gene expression or altering biological properties of living cells for therapeutic use (FDA, 2026). That definition matters because the model may affect design, characterization, manufacturing, or evidence generation.
FDA’s January 2024 CAR-T guidance states that CAR T-cell products are human gene therapy products and gives recommendations for chemistry, manufacturing, and controls, pharmacology and toxicology, clinical study design, and comparability for autologous and allogeneic CAR-T products (FDA, 2024). FDA’s January 2024 genome-editing guidance focuses on product design, manufacturing, testing, nonclinical safety, and clinical trial design for human gene therapy products that incorporate genome editing (FDA, 2024). The April 2026 draft guidance on next-generation sequencing for genome-editing safety adds a specific regulatory signal: off-target editing and loss of genome integrity require direct nonclinical assessment, not only computational prediction (FDA, 2026).
The right evidence question is not “did AI design it?” The right question is what decision the model changed: target choice, receptor sequence, guide selection, vector tropism, cell-state release criterion, potency assay, patient eligibility, or manufacturing control. In cell and gene therapy, every design output has to pass through a living system.
Product Classes and Decision Points
Engineered immune-cell therapies. CAR-T, CAR-NK, TCR-modified T cells, tumor-infiltrating lymphocyte products, and edited immune-cell products create a chain of design and manufacturing questions. A receptor must bind the intended antigen, avoid unacceptable cross-reactivity, signal with the right strength, resist exhaustion, persist long enough to work, and avoid severe toxicity. The same receptor sequence may behave differently in autologous and allogeneic products, in CD4 and CD8 subsets, in naive and memory phenotypes, and after different expansion protocols.
Genome-editing therapies. CRISPR nuclease editing, base editing, and prime editing have different evidence problems. Nuclease editing creates double-strand breaks and therefore raises concerns about off-target cuts, translocations, large deletions, p53 selection, and chromosomal rearrangements. Base editors avoid double-strand breaks for single-base conversion but introduce editor-specific DNA and RNA off-target questions. Prime editing extends the edit menu but adds delivery and efficiency constraints because the editing machinery and guide architecture are larger and more complex.
Viral-vector gene therapies. AAV, lentiviral, adenoviral, and other vector programs need payload capacity, tropism, expression level, biodistribution, immunogenicity, durability, and manufacturing yield evidence. AAV design is especially active because capsid sequence determines packaging, tissue distribution, neutralizing antibody escape, and manufacturability.
Non-viral delivery. Lipid nanoparticles and other delivery systems are central for RNA and genome-editing payloads. A formulation score is meaningful only after it is tied to tissue distribution, endosomal escape, dose, repeat dosing, tolerability, and manufacturing reproducibility.
Manufacturing and quality control. Cell and gene therapy products are often process-defined. AI may help monitor process drift, predict batch failure, analyze single-cell product composition, interpret potency assays, and connect in-process parameters to release specifications. This is where model governance matters most because the model may become part of quality decision-making rather than research triage.
Current Models, Datasets, and Benchmarks
The field does not have one benchmark that proves readiness for cell or gene therapy translation. It has method-specific data layers. Each layer answers a narrower question.
| Area | Examples | What the evidence supports | What it does not support |
|---|---|---|---|
| CRISPR guide design | Doench guide-design models, CRISPOR, on-target and off-target scoring | Guide prioritization for specified nuclease, organism, locus, and assay context (Doench et al., 2016; Haeussler et al., 2016) | Clinical off-target safety by itself |
| Genome-wide off-target assays | GUIDE-seq, CIRCLE-seq, CHANGE-seq | Empirical off-target nomination and assay-specific risk discovery (Tsai et al., 2015; Tsai et al., 2017; Lazzarotto et al., 2020) | Exhaustive in vivo safety |
| Base and prime editing | Cytosine base editors, adenine base editors, prime editors | Mechanistic editing systems and edit-class constraints (Komor et al., 2016; Gaudelli et al., 2017; Anzalone et al., 2019) | Delivery, durability, and product-level benefit-risk |
| AAV capsid engineering | AAV fitness landscapes and machine-guided design | Viability, packaging, and defined capsid-property prediction in experimental libraries (Ogden et al., 2019; Bryant et al., 2021) | Human tissue tropism, immunogenicity, or clinical efficacy without in vivo evidence |
| Immune-cell state modeling | Single-cell RNA, protein markers, exhaustion signatures, clonal tracking | Product composition, persistence hypotheses, and comparability analysis | Replacement for potency, toxicity, or clinical response evidence |
| Manufacturing analytics | Process sensors, release-test data, potency assays, batch records | Process monitoring, drift detection, batch-risk triage | A substitute for validated release specifications |
Genome-editing product development also has a regulatory benchmark layer. FDA has now put explicit weight on next-generation sequencing methods for assessing genome-editing safety risks, including off-target editing and loss of genome integrity, in a draft guidance that supplements the 2024 genome-editing guidance (FDA, 2026). That does not make any single sequencing assay sufficient. It does make clear that computational off-target prediction is not the endpoint.
What is demonstrated?
Demonstrated capability includes guide RNA selection for bounded CRISPR tasks. Doench et al. trained guide-design models using large-scale activity measurements and showed that sequence features improve guide prioritization for SpCas9 in the tested context (Doench et al., 2016). CRISPOR combined guide selection and off-target annotation into an accessible tool that became widely used for CRISPR experiment design (Haeussler et al., 2016). These tools are useful because they reduce obviously weak guide choices. They do not establish therapeutic safety.
Demonstrated capability includes empirical off-target discovery methods that constrain model claims. GUIDE-seq profiles double-strand breaks in living cells (Tsai et al., 2015). CIRCLE-seq screens purified genomic DNA for nuclease cleavage sites with high sensitivity (Tsai et al., 2017). CHANGE-seq added a method that can expose genetic and epigenetic effects on CRISPR-Cas9 activity (Lazzarotto et al., 2020). The practical implication is direct: off-target prediction belongs upstream of empirical measurement.
Demonstrated capability includes machine-guided AAV capsid engineering in defined experimental windows. Ogden et al. mapped a first-order AAV2 capsid fitness landscape and used that information for machine-guided design (Ogden et al., 2019). Bryant et al. trained deep models on capsid data and designed AAV2 capsid variants that remained viable for packaging across a targeted region (Bryant et al., 2021). These are important engineering results. They show that measured library data can guide capsid exploration. They do not prove tissue-specific human delivery.
Demonstrated capability includes manufacturing and QC analytics when the model is tied to a defined process variable and a validated assay. In-process sensor data, vector-genome titers, transduction efficiency, viability, identity markers, and potency assays are measurable. Models that flag drift, rank batch-risk factors, or identify atypical cell-state mixtures are useful quality tools when they are locked, validated, monitored, and governed.
What is theoretical?
Theoretical capability includes integrated receptor design for CAR-T and TCR products. A mature system would jointly evaluate antigen density, tumor heterogeneity, normal-tissue expression, scFv or TCR binding, signaling strength, tonic signaling, exhaustion, cytokine release, persistence, and manufacturability. Current workflows handle pieces of that chain. The product-level outcome still depends on cell biology, dose, host immune context, tumor microenvironment, and clinical management.
Theoretical capability includes joint editing and delivery design. A program might select a guide, editor, promoter, cargo architecture, delivery vehicle, dose, and patient subgroup together. This is plausible in bounded settings, especially when active learning adds new measured data. It remains theoretical as a general therapeutic engine because editing efficiency, off-target risk, tissue delivery, immune response, and durability are not solved by the same model.
Theoretical capability includes potency assays that forecast clinical effect. Potency is product-specific. A model may help select features that correlate with killing, cytokine production, engraftment, editing, expression, or secretion. But a potency claim requires a biological mechanism, a release assay, comparability evidence, and clinical relevance. The model is not the potency assay unless the assay itself has been validated for that use.
What is beyond current capability?
Beyond current capabilities includes fully in-silico approval-grade prediction of safety and efficacy for a cell or gene therapy product. Living products change during manufacturing, delivery, engraftment, expansion, persistence, and host response. No current model resolves those layers well enough to replace nonclinical studies, clinical trials, or long-term follow-up.
Beyond current capabilities includes assuming that a successful molecular design result transfers directly to a cellular therapy product. A receptor, guide, capsid, or LNP can look strong in a design benchmark and still fail because of immunogenicity, toxicity, insufficient delivery, poor persistence, manufacturing drift, or loss of potency.
Beyond current capabilities includes treating CRISPR off-target prediction as an adequate safety assessment. FDA’s 2026 draft guidance explicitly points toward NGS-based safety assessment for human genome-editing products, including off-target editing and loss of genome integrity (FDA, 2026). A model can nominate risk. It cannot certify absence of risk.
What would make this more promising?
Cell and gene therapy AI becomes more promising when evidence moves the output from triage to a named product decision. A claim should name the product class, edit or payload, delivery route, target tissue or cell type, cell source, manufacturing process, intended indication, and stage of development. A CAR-T receptor-design claim, an ex vivo hematopoietic stem-cell editing claim, an in vivo AAV liver-delivery claim, and an LNP extrahepatic-delivery claim are not interchangeable.
Assay alignment defines the claim. A guide score should be tied to editing efficiency and off-target data. A capsid score should be tied to packaging, expression, biodistribution, and immunogenicity. A potency model should be tied to the product’s mechanism of action. A manufacturing model should be tied to locked process variables and release decisions.
Denominator reporting calibrates confidence. A credible claim reports how many candidates were proposed, filtered, manufactured, tested, failed, and advanced. The numerator-only version is common: one designed capsid worked, one receptor looked promising, one guide edited well. The denominator tells whether the system improved discovery or only selected one successful example from a large hidden set.
Comparability evidence defines the scope. Cell and gene therapy products are often sensitive to process changes. If a model changes a process, assay, feature definition, or release decision, comparability evidence is needed. This is especially important for autologous products with patient-to-patient variability and for allogeneic products where donor source and expansion protocol can shift the final cell state.
Follow-up evidence sets the evidentiary bar. FDA’s long-term follow-up guidance reflects the possibility of delayed adverse events after human gene therapy products, especially when products create permanent or long-acting changes (FDA, 2020). Computational involvement does not shorten that biological timeline.
Failure Modes
Off-target blind spots. Sequence models may miss off-target sites that arise from chromatin, variant genomes, editor chemistry, repair pathways, or assay sensitivity. Prediction should feed empirical assays, not replace them.
On-target genotoxicity. A guide that edits the intended locus may still create large deletions, rearrangements, translocations, or clonal selection. On-target editing is not automatically safe editing.
Delivery mismatch. AAV and LNP design often fails at the delivery layer. A capsid or nanoparticle that works in one species, tissue, disease state, or dose range may not transfer to the intended patient population.
Immune-context mismatch. Pre-existing neutralizing antibodies, innate immune activation, cytokine release, T-cell exhaustion, antigen escape, and tumor microenvironment effects can dominate the engineered design.
Manufacturing drift. Cell phenotype, vector quality, transduction efficiency, expansion duration, cryopreservation, lot-to-lot reagents, and operator or site differences can change product behavior. A model trained on historical manufacturing data may fail after a process change.
Potency proxy failure. Viability, marker expression, cytotoxicity, editing rate, or transgene expression may not track clinical effect. A potency assay should measure a function connected to mechanism of action, not merely a convenient feature.
Cross-program overgeneralization. Evidence from CAR-T does not automatically apply to TCR-T, edited NK cells, HSC editing, in vivo AAV, or LNP editing. Modality-specific evidence is mandatory.
What should researchers, biotech teams, funders, and program leaders do with this?
Use this checklist before treating a cell or gene therapy AI claim as decision-grade.
- Product class is named: CAR-T, TCR-T, CAR-NK, edited HSC, in vivo AAV, LNP, tissue-engineered product, or another defined class.
- Intended decision is named: design, guide selection, off-target triage, delivery choice, process monitoring, release testing, potency, patient selection, or clinical endpoint analysis.
- Input data are described by source, assay, organism, tissue, cell type, disease context, and batch structure.
- Training and test splits are biologically meaningful: held-out locus, held-out donor, held-out antigen, held-out capsid family, held-out manufacturing run, or held-out clinical site.
- Baselines are present: established guide tools, empirical screens, non-ML process controls, standard potency assays, and conventional statistical models.
- Wet-lab validation is specified: editing efficiency, off-target assays, chromosomal integrity, cell phenotype, potency, biodistribution, immunogenicity, or release testing.
- Denominators are reported: candidates designed, manufactured, tested, failed, and advanced.
- Assay thresholds are locked before confirmation testing.
- Safety risks are mapped to modality: integration-related mutagenesis, off-target editing, immune toxicity, cytokine release, vector shedding, germline exposure risk, replication-competent virus, or product contamination.
- Manufacturing impact is assessed through comparability, release specifications, batch records, and monitoring plans.
- Regulatory language is precise: investigational, cleared, approved, qualified, validated, or research-use only.
- Long-term follow-up needs are acknowledged for permanent or long-acting changes.
FAQ
Is an AI-designed CAR-T receptor ready for clinical use if it binds the target?
No. Binding is only one requirement. A receptor program still needs specificity, normal-tissue cross-reactivity assessment, signaling behavior, exhaustion and persistence data, cytokine-release risk assessment, manufacturability evidence, potency testing, and clinical evidence.
Are CRISPR off-target predictors enough for therapeutic guide selection?
No. They are useful triage tools. Therapeutic editing requires empirical off-target assessment, on-target integrity assessment, cell-type context, donor or patient variation analysis, and a risk plan aligned with the intended product.
Do base editors and prime editors remove the safety problem?
No. They reduce some risks associated with double-strand breaks and introduce different questions. Base editors need editor-specific DNA and RNA off-target assessment. Prime editors need efficiency, byproduct, delivery, and payload-size assessment. Both need product-level evidence.
Is AAV capsid machine learning clinically proven?
No. Machine-guided AAV capsid work has shown strong experimental engineering value in defined libraries. Clinical delivery requires biodistribution, dose, immunogenicity, durability, tissue specificity, manufacturing, and human safety evidence.
Where does AI add the most practical value today?
The strongest near-term value is triage and measurement: guide ranking, capsid library prioritization, cell-state characterization, batch-risk detection, assay interpretation, and candidate selection for wet-lab testing. Product claims require measured biological evidence.