Matrix Taxonomy

The Papers.md matrix pairs an AI/ML method (row) with a cellular-agriculture research area (column). This page is the canonical definition of every row and column: what it covers, what falls out of scope, and (for categories that are easy to confuse) how to tell it apart from its neighbours. It serves two audiences: readers orienting themselves in the field, and the automated classification audit, which decides a paper’s placement from the paper’s own methods text measured against these definitions.

A guiding rule for every category: a placement rests on what a paper demonstrably did, read from its methods text rather than from what it could be applied to or what its title gestures at. What that evidence has to establish is set out once in the methods preamble below, and it is narrower than mere presence: a row records a paper’s contribution, not every technique its methods happen to mention.

Two axes, and how to tell which one you are reading

This file defines three vocabularies, and two of them are subject axes over different populations:

• Research areas are the matrix’s columns. They are a classification instrument: one axis of AI method x research area, covering only work that applies a method to a problem.
• Subject themes are the browse taxonomy behind /topics/. They tag any content type, so they reach databases, datasets and software that never enter the matrix.

The two are paired but never equated. Each theme names exactly one research area and each research area is named by exactly one theme, a correspondence stored as topics.area_key and asserted by db:check. The assertion runs through that key and never by comparing labels, because the populations genuinely differ: on every pair the theme is the larger, because it also tags datasets, software and databases that never enter the matrix, and the size of the gap varies widely from pair to pair. Their counts must never be added together or compared.

Because the axes are paired, their labels are deliberately shaped differently so that no label appears on both:

• a research area reads as a problem (Media Optimization, Sensory Prediction, Food Safety Prediction);
• a subject theme reads as an &-joined subject (Media & Growth Factors, Sensory & Flavor).

The one pair that once shared a label, Bioprocess & Scale-Up, is the one pair that produced a bug: a whole-file flatten let the theme’s two-line blurb silently overwrite the column’s scope definition. The theme is therefore named Bioprocess & Manufacturing while the column keeps Bioprocess & Scale-Up. Theme labels are independent of their frozen topic: slugs, so that naming carries no identifier change.

This convention governs the two subject axes only. It says nothing about the section names used in Software.md and Databases.md, which group entries for browsing and are free to read differently.

Research areas (columns)

Media Optimization

Designing and optimizing the culture medium: basal formulations, serum-free replacements, growth factors, and supplement concentrations: to improve cell growth, cost, or product quality. In scope: media-composition search, formulation design, and the experimental-design loops that drive it. Out of scope: operating or scaling the bioreactor (→ Bioprocess & Scale-Up) and the cells’ own biology (→ Cellular Engineering).

Also in scope: the other designed solutions a cell line is held in, chiefly cryoprotectant cocktails and vitrification solutions for cell banking. The optimized object is the same kind of thing, a multi-component fluid searched over composition under a measured cell response, and the search methods are the same; what differs is that the response is post-thaw viability rather than proliferation. That difference is not a reason to split the column.

What puts such a paper in scope is the optimized object rather than the sector it names. Cryoprotectant optimization has close twins outside this field, in reproductive medicine, cell therapy and biopharma cell banking, and this column claims these papers in spite of that rather than by pretending otherwise: banking a line is itself a step in a cultivated-meat process, so the work is in scope whoever the authors had in mind. Other columns weigh a paper’s stated application differently and each states its own test, so do not read this paragraph as describing theirs. Where that reading is contested, raise it for re-audit rather than moving the cell.

Cellular Engineering

Understanding and engineering the cells themselves: gene expression and regulation, cell-state and perturbation modeling, differentiation, cell-line characterization, and single-cell analysis used to interpret or modify cells. In scope: scRNA-seq analysis, perturbation-response prediction, gene regulatory modeling, cell-type annotation, and genetic/sequence design. Out of scope: what feeds the cells (→ Media Optimization) and the reactor they grow in (→ Bioprocess & Scale-Up).

Bioprocess & Scale-Up

Running and scaling the bioprocess: bioreactor design, process monitoring and control, fermentation strategy, and the transport phenomena (mixing, mass transfer, fluid dynamics) that govern scale-up. In scope: reactor and impeller design, CFD/mixing simulation, fed-batch and feeding-strategy control, real-time process monitoring, precision-fermentation process design (microbial-host production of media proteins and growth factors, an in-scope cellular-agriculture route), and the chemical/biological-engineering know-how cultivated-meat manufacturing depends on at volume, including methods demonstrated on non-biological reactor physics that transfer directly to bioreactor scale-up. Out of scope: the medium recipe itself (→ Media Optimization) and intracellular biology (→ Cellular Engineering). (Renamed from “Bioprocess control” to make the reactor/scale-up engineering scope explicit.)

Scaffolding

Designing and predicting edible, food-grade scaffolds and structural biomaterials (including the plant-derived structural matrices that cellular agriculture deliberately favors for a fully animal-free product: textured or extruded plant protein, biopolymer gels, decellularized plant scaffolds, mycelium) that give cultured tissue and hybrid products their three-dimensional form, mechanical properties, and structural texture. In scope: scaffold geometry and architecture; mechanical, rheological, and structural-texture prediction and optimization of the food-grade matrix; tissue moulds; and biomaterial selection. A paper that predicts or designs the structural matrix as a material sits here. It is additionally dual-classified with Sensory Prediction when both of the following hold: (a) the quantity it predicts or optimizes is an eating-quality attribute, meaning either a human sensory rating or a validated instrumental proxy for one (TPA hardness, chewiness, springiness, cohesiveness; shear force; instrumental color); and (b) the paper motivates that quantity as a property of the finished product as eaten, rather than as a process-control or material-characterization endpoint. Out of scope: the soluble medium (→ Media Optimization); flavor, taste, aroma, and quality assessment of intact finished food (→ Sensory Prediction); and non-edible biomedical scaffolds (e.g. orthopedic or bone tissue-engineering constructs) whose methodology may transfer but whose application is not cellular agriculture, and which belong in AI Tooling / Methodology.

Sensory Prediction

Predicting sensory, organoleptic, and quality attributes as perceived outcomes of the product: flavor, odor, taste, aroma, perceived texture and mouthfeel, color, and freshness/spoilage. In scope: odor/taste/flavor prediction from chemical structure or spectra; perceived-texture and eating-quality prediction; and meat-quality/freshness classification. A perceived-texture result is dual-classified with Scaffolding when the paper also designs or predicts the underlying structural matrix as a material, judged by the same two-part test given under Scaffolding. Out of scope: designing the structural biomaterial itself as a material (→ Scaffolding), and purely nutritional composition with no sensory target.

Metabolic Modeling

Reconstructing, simulating, and searching metabolism itself: genome-scale metabolic models (GEMs), flux balance analysis and related constraint-based methods, kinetic/ODE models of metabolic pathways, pathway and strain design, and the metabolic-network resources these depend on. In scope: GEM reconstruction, curation, and simulation; flux prediction; pathway enumeration and enzyme or pathway selection decided against a network model; strain design; and machine learning trained on metabolic-network structure or flux data. Out of scope: black-box optimization of an enzyme cascade’s composition, where the demonstrated result is a formulation and no network is modelled; choosing what goes into the medium, even when a metabolic model motivates the choice, where the demonstrated result is a formulation (→ Media Optimization); gene regulation and cell-state modeling that does not run through a metabolic network (→ Cellular Engineering); reactor operation and scale-up (→ Bioprocess & Scale-Up); and, importantly, metabolomics used as an analytical input to predict an eating-quality attribute, which is a sensory result measured by metabolite chemistry rather than a model of metabolism (→ Sensory Prediction). The distinguishing question is whether the paper models the metabolic network, or merely measures metabolites.

Food Safety Prediction

Computationally assessing the safety of the proteins and compounds an alternative-protein product introduces: allergenicity and immunogenicity prediction, IgE-epitope mapping, cross-reactivity assessment, and toxicity prediction for novel or heterologously-expressed proteins. In scope: trained classifiers and learned representations that predict allergenic, immunogenic, or toxic potential from sequence, structure, or physicochemical descriptors, including protein-language-model approaches. Out of scope: the allergen reference databases such a method is trained or screened against, which are catalogued in Databases.md rather than placed in the matrix; sensory or spoilage quality, which is a different kind of safety-adjacent question (→ Sensory Prediction); and a general-purpose predictor presented with no alternative-protein or cell-ag framing (→ AI Tooling / Methodology).

A boundary this column needs stated, because titles hide it. Not every published allergenicity predictor applies an AI/ML method, and a tool that applies none has no row to sit on and is therefore not matrix-eligible at all, however well it fits this column’s subject. Sequence-similarity screening against a reference set (a sliding-window identity threshold of the kind the FAO/WHO Codex Alimentarius prescribes) and fingerprint comparison by a similarity coefficient are retrieval procedures, not trained models. Judge this from the methods section rather than the name: several tools in this space are named like classifiers and are not. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

How to apply the framing test above, because the words are usually absent. Papers that belong in this column rarely say “alternative protein” or “cultivated” anywhere, so a keyword test for the sector rejects them all. What separates an in-scope paper from an identical method presented for another field is what its body text orients itself around. A paper that argues from the novel-food regulatory apparatus (FAO/WHO allergenicity guidance, Codex Alimentarius, the dossier requirements of a food-safety authority) is framed for this column whether or not it names the sector. A paper whose stated motivation is clinical or pharmaceutical, such as disease risk or the immunogenicity of a therapeutic protein, is not, however closely its method matches. Read this from the body text: the title and abstract of an allergenicity paper look the same either way. This is never on its own grounds for a removal. Where the out-of-scope clause above applies on its own terms, it names the destination; where it does not, because the paper is narrowly scoped to another field’s question rather than general-purpose, raise the placement rather than unseating it.

AI Tooling / Methodology

General-purpose AI methods, agents, tools, and frameworks that are applicable to cellular agriculture but are not yet tied to a specific applied cell-ag result. This is the home for a method described in general form (a biomedical agent, an agent framework, a knowledge-graph protocol) that could be applied to a cell-ag problem. In scope: general agents/tools/frameworks and methodology papers. Out of scope: a tool applied to a concrete cell-ag problem, which belongs in that problem’s column (e.g. an agent used for media design → Media Optimization).

AI/ML methods (rows)

A row records what a paper contributed, not what appears in its methods section. A paper earns a method row by building or applying that method as part of its reported result. It does not earn one by consuming the method as a service, by using it to configure the machinery that produced the result, or by being measured against it. The same technique can be a contribution in one paper and a component in another, so the question is never whether the method is present but whether the paper is making a claim about it.

Rows that restate it: Foundation Models, Benchmarks & Evaluation Frameworks, Genetic Algorithms.

Each states it in the terms its own boundary needs, because each has been misread at least once: building a model versus invoking one, shipping a benchmark versus being scored by it, and where a search’s output appears. Read the row’s own text for the test that applies. That list is the authoritative scope of this principle and is read mechanically, so a row named there without a clause in its definition, or a clause deleted while the row stays named, is caught rather than noticed.

This is not the same as the routing rules elsewhere in this section, which say a paper belongs in a different row (a CNN paper goes to CNN rather than Deep Learning; a data-driven model trained on simulation output is not Hybrid Mechanistic-ML). Those answer “which row”; this one answers “any row at all”. Do not merge them.

Bayesian Optimization

Sequential, surrogate-model-based optimization: a probabilistic model (usually a Gaussian process) plus an acquisition function chooses the next experiment to run, sample-efficiently. In cell-ag, the workhorse for media and process optimization under a tight experimental budget. Distinct from Bayesian inference (posterior or flux estimation that quantifies uncertainty without acquisition-driven experiment selection); that is not this row.

Deep Learning

Multi-layer neural networks not better described by a more specific row: deep MLPs, generic feed-forward and recurrent nets, and autoencoders/transformers used as plain predictors. A catch-all: a paper whose core architecture is a CNN, GNN, GAN/VAE, or a pretrained foundation model belongs in that row instead. Not shallow models (e.g. RBF networks, single-hidden-layer MLPs used only as a baseline) and not non-neural methods (chemometrics/PLS, network propagation) that are sometimes loosely called “deep learning”.

GNN

Graph neural networks: learned, message-passing models over graph-structured data (GCN, GAT, GIN, graph autoencoders). In cell-ag: cell–cell similarity graphs, molecular graphs, and gene-interaction networks. Distinct from classical graph algorithms, random-walk network propagation, and non-learned graph embeddings (e.g. metapath2vec, node2vec), which are not trained GNNs.

CNN

Convolutional neural networks for grid- or image-structured data. In cell-ag: microscopy and product imaging, hyperspectral/spectral data, and 3-D scaffold tomography. The placement requires genuine spatial convolution on image-like input, a dense regression surrogate merely labelled “CNN” is better placed under Deep Learning.

GAN / VAE

Deep generative models (generative adversarial networks and variational autoencoders) used for data synthesis, augmentation, or generative dimensionality reduction. In cell-ag: synthetic regulatory- sequence design and scRNA-seq embedding.

Hybrid Mechanistic-ML Models

Models that couple a first-principles description to a learned component, so that neither half stands alone: mass-balance or kinetic ODEs whose unknown rates are supplied by a network, physics-informed networks carrying a governing equation in the loss, and grey-box serial or parallel structures. The row is earned by the coupling, not by the network, which is why a paper whose learned part is a small shallow net belongs here rather than in Deep Learning, and why a paper with a genuinely deep network can hold both rows. Distinct from a purely data-driven model that merely uses a mechanistic simulation to generate its training data, and from an ensemble that averages several learned models and calls the combination hybrid: in both of those the first-principles structure is absent from the deployed model.

Genetic Algorithms

Population-based evolutionary search: genetic algorithms, evolutionary strategies, NSGA, differential evolution. In cell-ag: optimizing media, process parameters, and tissue/scaffold designs over large combinatorial spaces.

The test is where the search’s output appears. The row is earned when what the search returned is part of what the paper reports: a formulation, a physical design, a process setting, or the fitted model itself. It is not earned when the output only configures the machinery that produced the reported result, which covers a search over hyperparameters, over model architecture (except as the third shape below sets out), and over which preprocessing or feature-selection technique to apply. Both kinds are real optimisation and both are often described at length, so the length of the description decides nothing. What decides it is whether the answer the search returned is one the paper makes a claim about.

A feature subset can itself be a finding, since “these are the variables that matter” is a claim about the problem. Feature selection is therefore not excluded as a category; what is excluded is a search whose answer is which technique to run.

Two shapes, written out because a single sentence here has already been read both ways. A scoring-card method whose evolutionary search produces the score table that is the predictor earns the row, because the table is the result. A study whose search identifies the best combination of feature-selection techniques to reduce features before training does not, because its reported result is about the biology and the search’s answer is a step in the pipeline. A paper that does both earns the row on the first, so read the whole methods section rather than the first evolutionary-search heading in it.

A third shape, because architecture search sits on the line the first two do not. A search over a model’s architecture produces a fitted model, which the positive test admits, and configures the machinery behind a result, which the exclusion refuses. What separates them is not the search but the paper: where the method for producing that model is the paper’s stated contribution, and its results characterise the model the search returned, the row is earned. Where the tuned model is a means to a claim about something else, and the paper’s results are about that something else, it is not. The question to ask is what the paper would still be reporting if the search had been replaced by a default configuration. This clause widens the row, so where it bears on a placement already made, raise it for re-audit rather than silently unseating or adding one.

SVM

Support vector machines and support vector regression, kernel methods for classification and regression on modest, well-featurized datasets. Common in sensory and media/process modeling.

Ensemble Learning

Tree ensembles and model averaging: random forests, gradient boosting / XGBoost, AdaBoost, and stacked ensembles. The default strong baseline for tabular prediction across sensory, media, scaffold, and process problems. The row is earned when trees, boosting, or stacking over classical learners is the primary learner. Not an ensemble assembled over networks that already hold their own architecture row: model averaging, snapshot ensembling, or stacking over CNNs, GNNs, or deep MLPs does not additionally earn this row, because what the row records is the learner, not the combination rule placed on top of it. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

K-Nearest Neighbors

Instance-based prediction from nearest-neighbor similarity, for small-data classification and regression. Usually appears alongside other classical models in comparative studies.

Linear & Regularized Models

Classical linear predictors and their regularized variants, ordinary and penalized linear/logistic regression (LASSO, ridge, elastic net) and linear additive scoring models (e.g. propensity-score scoring cards). Valued for interpretability and small-data stability: the fitted coefficients name which features drive the prediction, and L1 penalties perform explicit feature selection. In cell-ag: morphology- and sequence-feature regression for cell-quality and taste prediction where the feature set is modest and interpretability matters. Distinct from the SVM row (no kernel or margin), Ensemble Learning (no trees, bagging, or boosting), and Chemometrics (not restricted to spectral latent-variable projection).

Chemometrics

Multivariate statistics for extracting quantitative information from chemical measurements, overwhelmingly spectra. The workhorses are Partial Least Squares (PLS) regression, PLS-DA (discriminant analysis), PCA, and OPLS. In cell-ag this is the backbone of spectroscopy-based Process Analytical Technology: NIR/Raman monitoring of glucose, lactate, and biomass in a bioreactor, and hyperspectral quality/freshness prediction. Distinct from the kernel (SVM), tree-ensemble, and neural-net rows: it is latent-variable linear projection, not learning by kernels, trees, or backpropagation.

Active Learning

Iterative design-of-experiments in which the model selects which experiments to run next to label, typically wrapped around a surrogate model, closing a predict→experiment→retrain loop. In cell-ag: media and strain optimization campaigns. Adjacent to Bayesian Optimization (a specific acquisition- driven instance); this row is the broader iterative-selection framing.

Reinforcement Learning

Learning policies from reward signals: classical RL, policy-gradient methods, and LLM post-training (GRPO, RLHF). In cell-ag: reward-driven post-training of models and virtual-cell control.

Foundation Models: Next-Token Prediction

Large, pretrained, transferable models trained with an autoregressive next-token objective (GPT-style), applied to biology, e.g. generative transcriptome and cell-atlas models. The defining test is that the model is a pretrained foundation model, not a task-specific network. Not a paper that merely invokes such a model, as an agent’s tool call or a queried service: that paper keeps its own row, and this one records building the model rather than using it. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

Foundation Models: Masked Language Modeling

Pretrained, transferable foundation models trained with a masked (BERT-style) objective, the single-cell foundation models such as scBERT, Geneformer, and scFoundation. Defined by masked self-supervised pretraining at scale plus downstream transfer. Not a paper that merely invokes such a model, as an agent’s tool call or a queried service: that paper keeps its own row. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

Foundation Models: LM + Biological Priors

Foundation models that combine a language-model backbone with explicit biological priors or structured knowledge (e.g. protein-language-model representations integrated with curated biology). The model must itself be such a foundation model, a task-specific network that merely consumes pretrained embeddings as input features does not belong here (it is Deep Learning / GNN), and neither does an agent that invokes one as a tool (it keeps its agent row). Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

Foundation Models: Cell-State & Perturbation Prediction

Pretrained, transferable foundation models whose target is predicting cell state and perturbation responses across contexts. The discriminator is pretraining-and-transfer: a single-task supervised predictor (for example a GNN trained end-to-end on one perturbation dataset) is not a foundation model, even though it predicts perturbations, it belongs in its architecture’s row. Not a paper that merely invokes such a model, as an agent’s tool call or a queried service. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

Foundation Models (other modalities)

Pretrained foundation models for modalities beyond single-cell transcriptomics not covered by the other foundation-model rows: multimodal omics, spectra, and small-molecule/natural-product models. Not a paper that merely invokes such a model, as an agent’s tool call or a queried service. Where an existing placement is affected, raise it for re-audit rather than silently unseating it.

Scientific Literature & Discovery Agents

LLM agents for literature search, synthesis, and (semi-)autonomous discovery: retrieval-augmented question answering over papers and idea→experiment→manuscript pipelines. Among the agent rows, this one is defined by its focus on knowledge synthesis and the discovery workflow.

General-Purpose Biomedical Agents

LLM agents built to be broadly applicable across biomedical tasks without task-specific tuning, and that say so. Distinguished from Domain-Specific Biomedical Agents by breadth of intended use.

Chemistry / Synthesis Agents

LLM agents specialized for chemistry: synthesis planning, molecular reasoning, and autonomous chemical experimentation.

Domain-Specific Biomedical Agents

LLM agents purpose-built for one specific biomedical task or domain: NGS/RNA-seq analysis, spatial biology, a single named assay or modality. Distinguished from General-Purpose Biomedical Agents by narrow, dedicated focus.

Robot Scientists & Lab Automation

Systems that couple AI to physical laboratory automation for autonomous experimentation, closing the loop from hypothesis to robotically executed experiment. The discriminator is physical/automated wet-lab execution, not an in-silico agent alone.

Benchmarks & Evaluation Frameworks

Papers whose primary contribution is a benchmark or evaluation framework: datasets paired with tasks, protocols, and metrics, or a method for scoring/verifying AI and agent performance. Not a model that is merely evaluated on a benchmark (→ its own method × area), a dataset shipped without an evaluation protocol (→ Datasets/), a leaderboard service (→ Databases.md), or a correspondence/commentary about an evaluation (→ Reviews & Perspectives).

A paper in this row takes the column of the research area its benchmark measures, read off the task set rather than the title: a benchmark of protein-fitness or gene-perturbation tasks is Cellular Engineering, one of gel-microstructure tasks is Scaffolding, and one whose tasks span many sciences with no applied cell-ag endpoint is AI Tooling / Methodology. What decides it is the quantity the benchmark grades, not how much of its subject matter sounds biological: a suite scoring a statistical estimate recovered from simulated data measures analytic skill, whichever field supplied the scenario. A multi-domain suite whose task set genuinely splits, with each side substantially represented and one side outside every research area, is dual-classified across both columns rather than forced into one.

This row is otherwise easy to confuse with Comparative Studies, and the boundary between them runs on reuse. A benchmark ships a reusable task suite, protocol and metrics for others to score against; a comparative study evaluates existing methods to answer its own question and ships nothing for reuse. A paper may do both, and then it takes both rows. The same test decides whether a paper earns a cell in each method family it ran: it does so only where the evaluation is the paper’s own method study, and not where it supplies a suite for others to run against models it did not build.

Comparative Studies

Papers whose primary contribution is a head-to-head evaluation of existing methods on the authors’ own data, run to answer a question the authors have rather than to supply a resource: which model family to adopt for a given prediction task, or which measurement modality feeds it best. In scope: sweeps across model families, and studies whose stated result is a recommendation about method choice. A paper here takes the column of the research area it evaluates methods for, on the same reading as Benchmarks & Evaluation Frameworks above: read off what the study actually measured. Out of scope: an applied paper that happens to try several models before reporting the one it uses, since the row records a paper’s contribution and not its model count.

Agent Infrastructure (Frameworks, KGs, Protocols)

The substrate that agents run on: agent frameworks, knowledge graphs, tool/communication protocols (e.g. MCP), and orchestration ecosystems. The discriminator is that the contribution is plumbing/infrastructure, not an agent that itself solves a downstream task.

Subject themes (topic tags)

A separate axis from the matrix: subject-matter themes that group any content type (papers, software, databases, datasets) for cross-content discovery. Unlike research areas, which are problems an AI method is applied to: themes are what the material is about, so they fit databases and datasets that never enter the matrix (per the Databases-vs-ResearchAreas distinction). Themes are a fixed backbone; finer tags live under them and are minted only when several items cluster. Every theme names exactly one matrix research area, and every research area is named by exactly one theme; see Two axes, and how to tell which one you are reading above for what that pairing does and does not license.

Media & Growth Factors

Culture-medium composition and optimization: serum-free/serum-reduced formulations, growth factors and cytokines, recombinant medium proteins, and design-of-experiments for medium screening. Links to the Media Optimization research area.

Cell Lines & Engineering

The cells themselves and their manipulation: cell-line establishment and characterization, genetic engineering (CRISPR, perturbation), differentiation and myogenesis, stemness/senescence, and single-cell atlases. Links to Cellular Engineering.

Bioprocess & Manufacturing

Growing cells at scale: bioreactor design and CFD, perfusion and fed-batch, soft sensors and process control, and scale-up engineering. Links to Bioprocess & Scale-Up.

Scaffolding & Biomaterials

The structural substrate for tissue: scaffolds, biomaterials and hydrogels, 3D bioprinting, and the texture/architecture of cultured tissue. Links to Scaffolding.

Sensory & Flavor

The eating quality of the product and the analytics behind it: flavor and taste prediction, aroma, and the mass-spectrometry / chemometrics / metabolomics infrastructure used to measure them. Links to Sensory Prediction.

Metabolism & Modeling

Metabolic knowledge and models: genome-scale metabolic models (GEMs), flux analysis, pathway and strain design, and metabolic-network resources. Pairs with the Metabolic Modeling research area, which is narrower: the column holds papers that model a metabolic network, while this theme also collects the model files, databases and tools that never enter the matrix.

Food Safety

The safety assessment of the proteins a cultured product introduces: allergenicity and immunogenicity prediction, IgE-epitope mapping, and the allergen reference databases that screening pipelines compare against. Pairs with the Food Safety Prediction research area; the allergenicity fine tag groups the in-silico screening tools, allergen databases, and their methods papers, most of which are catalogue entries rather than matrix placements.

AI Methods & Tooling

General-purpose AI capability and infrastructure applied across cell-ag: agents and foundation models, benchmarks and evaluation, frameworks, ontologies, and lookup databases/tooling. Links to AI Tooling / Methodology.

Crosswalk to GFI’s solutions taxonomy

The Good Food Institute organises its solutions database on two facets, Technology sector and Value chain segment. Readers arriving from GFI’s vocabulary need to know how it lines up with the matrix columns, so this table maps each CAAIL research area to its nearest GFI technology sector and says how good the fit is.

Adopting GFI’s facets as the matrix column axis was considered and rejected. The reason is visible in the table: GFI’s axis answers who should act on this gap, while a CAAIL column answers what did this paper demonstrate, and GFI has no facet at all for a general-purpose AI method. The largest CAAIL column is exactly that, so adopting their axis would leave a large share of the matrix with nowhere to sit. The crosswalk buys the legibility without the axis change.

CAAIL research area	Nearest GFI technology sector	Fit
Media Optimization	Cell culture media	direct
Cellular Engineering	Cell line development and Host strain development	splits across two
Bioprocess & Scale-Up	Bioprocess design	direct
Scaffolding	Scaffolding	direct
Sensory Prediction	End product formulation & manufacturing	partial: GFI’s is a production activity, CAAIL’s is a prediction target
Metabolic Modeling	Host strain development	partial: nearest neighbour only, GFI has no modelling facet
Food Safety Prediction	(none)	no GFI sector covers safety assessment
AI Tooling / Methodology	(none)	no GFI sector covers general-purpose AI method work

The gaps run in both directions, and both are real rather than an artefact of mapping:

• CAAIL areas with no GFI sector: Food Safety Prediction and AI Tooling / Methodology, with Sensory Prediction and Metabolic Modeling matching only loosely.
• GFI sectors with no CAAIL column: Crop development, Ingredient optimization, and Target molecule selection. These are largely plant-based and fermentation-ingredient concerns that the matrix’s cultivated-cell focus does not reach.

Two details worth stating because they are easy to get wrong. GFI lists Scaffolding on both facets, as a technology sector and as a value chain segment, where CAAIL has only the one axis. And almost every paper in the matrix would fall under GFI’s R&D value chain segment, so that facet does not discriminate between CAAIL columns and is not mapped per-row here.

GFI’s facet lists were read from gfi.org/solutions on 2026-08-20 and are reproduced verbatim. They are a third party’s taxonomy and can change without notice; re-read the page rather than trusting this snapshot.