Cellular Engineering

Choosing and modifying cells is a foundational problem in cellular agriculture: different cell types and genotypes differ enormously in how readily they proliferate, differentiate, suspension-adapt, and accumulate the protein and fat that make a product. The AI work in this area spans the full engineering arc — characterizing cell populations from single-cell data, building general-purpose models of cell state, predicting how cells respond to genetic or chemical perturbations, designing genetic parts, and automating the design–build–test loop itself. A recurring caveat applies: most of the single-cell methods below were developed on human and mouse data and are analogous to cell-ag rather than native to it, but a growing cross-species and explicitly cultivated-meat thread is beginning to close that gap.

Characterizing cells from single-cell data

Single-cell RNA sequencing is the workhorse for identifying and characterizing the cell types in a heterogeneous culture, and a cluster of deep-learning methods make that data more usable. Li et al. (2020, ref #5) introduced DESC, a deep embedded-clustering autoencoder that removes batch effects through iterative self-learning without requiring explicit batch labels. Three graph-neural-network methods cluster cells by learning from cell–cell relationship graphs: Wang et al. (2021, ref #13)‘s scGNN couples a left-truncated mixture-Gaussian expression model with iterative autoencoders for joint imputation and clustering; Ciortan & Defrance (2022, ref #8)‘s graph-sc learns cell embeddings from a cell–gene graph; and Shan et al. (2023, ref #12)‘s GLAE adaptively learns the cell-relation graph during training rather than fixing it in advance, making it robust to the noise that corrupts pre-computed graphs. Lin et al. (2020, ref #9)‘s DR-A casts dimensionality reduction as an adversarial variational autoencoder, reporting the best normalized-mutual-information clustering scores against PCA, scVI, and other baselines on several datasets. Further back, Chen et al. (2016, ref #4)‘s D-GEX showed a multi-task neural network could infer the expression of ~9,500 target genes from ~1,000 landmark genes, a cost-reduction strategy for profiling cell lines. These are general single-cell tools, but cell-type identification, batch correction, denoising, and cheap expression profiling are exactly what is needed to characterize and quality-control cultivated-cell populations across experiments. One image-based method here is explicitly cultivated-meat-native: Yang et al. (2025, ref #218) trained a deep network combining object detection and regression to read adipogenic differentiation directly from bright-field images of porcine suspension-cultured adipocytes — predicting per-cell lipid content (a BODIPY-fluorescence proxy) at R² ≈ 0.83 without the fluorescent stain at inference, and using the readout to screen fatty acids, flagging arachidonic acid as an adipogenesis promoter. It is a label-free quality-control and screening tool for cultivated-fat production in a directly cell-ag-relevant species.

Foundation models for single-cell biology

A more recent wave pretrains large transformer models on tens of millions of cells and adapts them to many downstream tasks. Masked-language-model architectures lead this group: Theodoris et al. (2023, ref #111)‘s Geneformer pretrains on ~30 million transcriptomes and transfers to network-biology tasks with limited data; Yang et al. (2022, ref #112)‘s scBERT uses a Performer backbone for scalable masked pretraining aimed at cell-type annotation and novel-type discovery; and Hao et al. (2024, ref #116)‘s scFoundation is a 100-million-parameter, read-depth-aware model spanning ~20,000 genes. A parallel set are generative/autoregressive: Shen et al. (2023, ref #115)‘s tGPT predicts the next gene over expression-ranked sequences; Cui et al. (2024, ref #117)‘s scGPT pretrains on >33 million cells and fine-tunes across annotation, integration, perturbation, and gene-regulatory-network tasks; and Pearce et al. (2026, ref #92)‘s TranscriptFormer is a generative atlas trained across twelve species spanning 1.5 billion years of evolution, using protein-sequence-derived gene embeddings to share a species-agnostic space. That cross-species capability is the most cell-ag-load-bearing sub-theme: Rosen et al. (2024, ref #118)‘s SATURN aligns datasets across species without relying on orthologs by grouping genes into protein-embedding-defined “macrogenes,” and Rosen et al. (2026, ref #119)‘s UCE produces zero-shot universal cell embeddings (a 650-million-parameter model over a 36-million-cell integrated atlas) that map new species into the representation space with no fine-tuning — directly relevant to characterizing non-model cultivated species that lack annotated atlases. Bridging to the LLM world, Rizvi et al. (2026, ref #120)‘s C2S-Scale trains true large language models (up to 27 billion parameters) on scRNA-seq profiles rendered as ordered “cell sentences,” enabling perturbation prediction and natural-language reasoning over cell state. A further large model in this masked-modeling family, Zeng et al. (2025, ref #235)‘s CellFM, is an 800-million-parameter single-cell foundation model pretrained on 100 million human cells with a masked-gene objective on a modified RetNet backbone, reporting gains over Geneformer, scBERT, and UCE on cell annotation, perturbation prediction, and gene-function tasks — though, like the rest of this group, its training data is human-only, so livestock applicability remains a transfer problem.

Predicting perturbation responses

Predicting what a genetic or chemical perturbation will do — without running every experiment — is central to rational cell engineering. Roohani et al. (2024, ref #121)‘s GEARS fuses a gene co-expression graph and a Gene Ontology graph to predict transcriptional outcomes of single- and multi-gene perturbations, including for genes never experimentally perturbed, reporting roughly 40% higher precision than prior methods on genetic-interaction subtypes. Adduri et al. (2025, ref #57)‘s State is a cell-state transformer trained on over 100 million perturbed cells that improves discrimination of perturbation effects by more than 30% and generalizes to cellular contexts unseen in training. Dong et al. (2026, ref #124)‘s Stack uses tabular attention over ~149 million cells to enable in-context “cell prompting” — predicting how a query population responds to a condition without fine-tuning — and produced Perturb-Sapiens, a whole-organism atlas of perturbed human cells. Littman et al. (2025, ref #123)‘s PRESAGE takes a deliberately simple, interpretable route, fusing gene embeddings from diverse knowledge sources, and shows that knowledge-source choice matters more than architectural complexity and that performance saturates quickly with training-set size — an argument for collecting sparse perturbation data across many biological systems rather than exhaustively profiling one. Magnusson et al. (2024, ref #122)‘s PreciCE pairs a computational method that predicts multi-gene activation/repression sets for cell-fate conversion with a CRISPR toolbox of dual-Cas systems to execute them — turning directed differentiation, the act of pushing stem or satellite cells toward muscle or fat while suppressing alternative lineages, into an engineering operation.

Designing genetic parts

Where perturbation models predict the effect of editing existing genes, a generative thread designs novel sequence. Zrimec et al. (2022, ref #10)‘s ExpressionGAN trains a generative adversarial network on paired genomic and transcriptomic data to design synthetic regulatory DNA with prespecified target expression levels; despite high divergence from natural sequence, 57% of its highly-expressed synthetic constructs surpassed strong natural controls when measured in vivo in yeast. Ji et al. (2021, ref #6)‘s DNABERT is, despite sitting in the deep-learning column, a BERT-style pretrained language model for genomic DNA: trained on the human genome and fine-tuned for promoter, splice-site, and transcription-factor-binding-site prediction, with nucleotide-level attention that aids motif and variant interpretation and transfers to other organisms. Together they represent the two halves of genetic design — generating regulatory parts that hit an expression target, and reading the regulatory grammar needed to engineer them.

Autonomous agents for cell engineering

The newest thread wraps the loop in LLM-based agents. Roohani et al. (2025, ref #125)‘s BioDiscoveryAgent designs CRISPR perturbation experiments using an LLM’s prior knowledge in place of a trained acquisition function, reporting a 21% average improvement (and 46% on the harder non-essential-gene task) over Bayesian-optimization baselines across six datasets. Hao et al. (2025, ref #97)‘s PerTurboAgent is a self-planning agent for sequential Perturb-seq design, using an action memory to let multi-round plans evolve adaptively. Tang et al. (2026, ref #93)‘s CellForge goes a level up, autonomously designing the neural-network architectures for perturbation-prediction tasks from raw multi-omics data and a task description. Two further agents build the supporting layer rather than engineering cells directly: Youngblut et al. (2025, ref #126)‘s scBaseCount is an agent that discovers and uniformly reprocesses public scRNA-seq data into a self-updating repository of more than 500 million cells across 27 organisms, and Yu et al. (2026, ref #90)‘s GenCellAgent is a training-free multi-agent framework that routes cellular images to the best segmentation tool and supports text-guided segmentation of structures existing models miss — an imaging and phenotyping capability for reading out engineered cells. An earlier autonomous-experimentation demonstration, Kanda et al. (2022, ref #16), used a robot with batch Bayesian optimization to optimize an iPSC-to-retinal-pigment-epithelium culture protocol, and shows the closed-loop ideal these agents pursue (it is discussed further under Media Optimization).

Cross-cutting and adjacent methods

A few entries sit at the edges of the column. Mathieu et al. (2025, ref #60) is a cultivated-meat-specific contribution: rather than a foundation model, it builds an integrative multi-omics interactome and a Target-Action-Metabolite framework, demonstrated on duck embryonic stem cells to generate media-intervention hypotheses while monitoring downstream interactome effects — tying cell engineering to media and quality/safety design. Xiao et al. (2026, ref #145)‘s MetaGEM is a dual-tower deep model (an ESM-2 protein tower and a UniMol-2 metabolite tower) that predicts enzyme–metabolite interactions to reconstruct genome-scale metabolic models bottom-up, reaching an AUROC of 0.9701 and reconstructing functional models for several microbes — work that bridges to Metabolic Modeling. Wu et al. (2026, ref #165)‘s CellFluxRL post-trains an image-based generative virtual-cell model with reinforcement learning against biologically meaningful reward functions (mode-of-action, nuclear and cytoplasmic structural validity), fixing physically implausible generations and offering its rewards as new benchmark metrics.

Three methods catalogued primarily under AI Tooling / Methodology also carry a Cellular Engineering matrix placement because they act directly on cell- or strain-engineering tasks: Pandi et al. (2022, ref #63)‘s versatile active-learning workflow for optimizing genetic and metabolic networks, Singh et al. (2023, ref #66)‘s automated scientist for designing and optimizing microbial production strains, and Li et al. (2024, ref #68)‘s LLM pipeline that mines metabolic-engineering literature into a knowledge base and feeds a hybrid deep-learning / genome-scale-model target predictor. Each is described more fully under AI Tooling.

Tools and data

Single-cell foundation models: the AI Agents & Foundation Models section in Software.md catalogues open implementations of nearly every model above — Geneformer, scGPT, scBERT, scFoundation, UCE, tGPT, Cell2Sentence (C2S-Scale), and TranscriptFormer.
Perturbation and agent tooling: GEARS, State + Cell-Eval, BioDiscoveryAgent, and CellForge.
Training and reference data: the Single-cell & perturbation corpora in Datasets/HumanReference.md — including Perturb-Sapiens (produced by ref #124) and Tahoe-100M — and the Cell Line & Single-Cell Reference Atlases in Databases.md (Cellosaurus, Human Cell Atlas, CZ CELLxGENE).

Open challenges for cell-ag

The methodological momentum here is real but the cell-ag-native evidence is thin. The great majority of the single-cell and foundation-model work is built and validated on human and mouse data; the cross-species models (SATURN, UCE, TranscriptFormer) are the most promising bridge to livestock and aquatic species, yet none is yet validated on a cultivated-meat-relevant cell line. Direct work on cultivated-meat-relevant species stays rare in the column — Yang et al.’s porcine adipocyte differentiation model (ref #218) is the clearest exception. The perturbation-prediction work depends on large perturbation atlases that do not exist for any farmed species — Perturb-Sapiens is the methodological template a future livestock-perturbation atlas would follow. The autonomous-agent demonstrations are early-stage and, in several cases, build the data and phenotyping layer rather than engineering cells. The clearest opportunities are concrete: a perturbation atlas for a muscle- or fat-relevant livestock cell line, cross-species foundation-model benchmarks that include farmed species, and the integration of perturbation prediction with the directed-differentiation tooling (ref #122) that would let a lab specify a target cell state and get back a buildable genetic program.