Deep Learning

This page describes the Deep Learning row of the Papers.md matrix: multi-layer neural networks not better described by a more specific row, including deep MLPs, generic feed-forward and recurrent networks, and autoencoders or transformers used as plain predictors. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

This is the axis’s declared catch-all, and reading it correctly means reading the exclusions first. A paper whose core architecture is a CNN, a GNN, a GAN or VAE, or a pretrained foundation model belongs in that row instead. The taxonomy also excludes two things that are easy to admit by mistake: shallow models, such as radial-basis-function networks or single-hidden-layer MLPs used only as a baseline, and non-neural methods sometimes loosely called deep learning, such as Chemometrics and network propagation.

Because it is a residual category, this row is not a coherent research programme and should not be read as one. What it does give is the broadest coverage of any method row, appearing in more research areas than any other, which makes it a reasonable starting point when the question is “has anyone applied a neural network to this part of the process at all”. Many of its references are described in more depth on the page for whichever other row they also occupy.

Media Optimization

#15 (Yoshida et al. 2023, Journal of Bioscience and Bioengineering): a deep neural network as the surrogate under Bayesian optimization for 31-component E. coli media, including the honest account of a first optimization round predicting badly and being fixed by retraining. Described in Bayesian Optimization.
#18 (Cosenza 2022, UC Davis dissertation): the neural-network-guided genetic algorithm and multi-information-source Bayesian work behind the cell-ag media thread. Described in Bayesian Optimization.
#212 (Tu et al. 2022, Scientific Reports): artificial neural networks compared against Gaussian process regression for bull-semen cryopreservation extender design, with GPR winning. Described in Bayesian Optimization.

Cellular Engineering

#5 DESC (Li et al. 2020, Nature Communications): unsupervised deep embedding that clusters scRNA-seq while removing batch effects, without being given batch labels. The mechanism is iterative self-learning: clusters are refined against an auxiliary target distribution derived from the current soft assignment, so cell clustering and gene selection are solved jointly rather than in sequence. The authors state the condition under which it works, that technical differences across batches must be smaller than true biological variation, which is the kind of stated precondition worth carrying into any reuse. Soft cluster-assignment probabilities are biologically interpretable rather than hard labels. Code at eleozzr/desc.
#118 SATURN (Rosen et al. 2024, Nature Methods, Leskovec lab): the most cell-ag-load-bearing reference in this column. Cross-species single-cell analysis is normally limited to homologous genes, which throws away most of the data; SATURN instead couples protein-language-model embeddings of each gene’s protein product with RNA expression, and learns a shared space of macrogenes, weighted combinations of genes across species, so datasets integrate regardless of gene correspondence. Initial cell-type labels are used within each species but never matched across them, so the cross-species alignment is learned rather than supplied. This is the published route to reading a bovine or porcine dataset against a human atlas.
#57 State (Adduri et al. 2025, bioRxiv, Arc Institute): the perturbation transformer, described in Cell-State & Perturbation Prediction.
#122 PreciCE (Magnusson et al. 2024, bioRxiv, Qi and Leskovec labs): directed differentiation by data-driven multi-gene control, and one of the few references that pairs a model with new CRISPR tooling built for it. Bidirectional systems based on dCas12a, Cas13d and dCas9 activate and repress many genes at once, and because the combinatorial space of multi-gene regulation is vast, a machine-learning component chooses which combination to apply. The stated problem, that current cell-fate control is inefficient, imprecise and relies on trial and error, is exactly the problem of moving a livestock cell line reliably into myogenic or adipogenic fate. Code at snap-stanford/precice.
#123 PRESAGE (Littman et al. 2025, bioRxiv, Regev and Genentech): predicts perturbation-induced expression change by transforming gene embeddings from diverse knowledge sources through an attention-based model, and is deliberately simple, modular and interpretable rather than architecturally ambitious. Evaluated across five Perturb-seq datasets from four cell lines (Replogle et al. and Nadig et al., spanning K562, RPE1, HepG2 and Jurkat, including a genome-scale screen of 9,866 expressed genes). Its two findings are the useful part and both are about where effort pays: the choice of knowledge source matters more than architectural complexity, and performance saturates quickly with training-set size, which argues for collecting sparse perturbation data across many contexts rather than deep data in one.
#263 (Vaishnav et al. 2022, Nature, Regev lab): builds sequence-to-expression models over millions of randomly sampled yeast promoter sequences measured by a gigantic parallel reporter assay, and uses them to construct fitness landscapes mapping regulatory DNA to expression and fitness. The scale of the measurement, not the architecture, is what makes the result possible, and it is the empirical foundation under generative regulatory-sequence design such as #10 ExpressionGAN. Data at GSE163045.

Bioprocess & Scale-Up

This column is the row’s densest, and it traces one clear methodological arc: from a plain network as a black-box surrogate, through hybrids that put mechanistic structure back in, to models that quantify their own uncertainty.

#30 (Peng et al. 2013, Bioresource Technology): the black-box starting point, a “5-10-1” network over fermentation parameters with a genetic algorithm deriving a control trajectory. Described in Genetic Algorithms.
#29 (Del Rio-Chanona et al. 2019, AIChE Journal): a surrogate for microalgal photobioreactor design, trained on simulations from an integrated model coupling computational fluid dynamics with kinetics, then searched by hybrid stochastic optimization. The motivation is that the integrated physical model becomes computationally intractable and numerically unstable at large scale, so it can be simulated for a few design scenarios but not optimized over. Bioreactor scale-up is the same problem.
#253 (Park et al. 2023, Biotechnology and Bioengineering): multistep-ahead forecasting of mammalian culture profiles toward bioprocess digital twins, over 24 fed-batch CHO-K1 batches across four media conditions, cultured 14 days with daily viable cell density and viability and metabolites every two days. Forecasting a profile several steps ahead is a harder and more useful target than predicting the next measurement, because it is what a controller needs.
#254 (Lu et al. 2025, Industrial & Engineering Chemistry Research): recurrent networks for time-varying behaviour, on a genuinely industrial dataset of 243 manufacturing runs of an IgG process spanning 12,000 L production and 80, 400 and 2,000 L seed-train stages. The problem it names is the one that breaks static models: cell metabolism shifts as the culture progresses, and similar starting conditions can produce very different time-varying behaviour. Data at this scale across seed train and production is rare in the matrix.
#255 Hybrid-EKF (Narayanan et al. 2020, Biotechnology and Bioengineering): couples a hybrid model with an extended Kalman filter for real-time monitoring and control, developed first on simulated data and then applied to a real industrial dataset. The framing is the honest middle position this column keeps arriving at: process datasets are not large enough for purely data-driven methods, and the underlying biology is not understood well enough for purely mechanistic ones, so hybrids are a pragmatic answer rather than an elegant one.
#205 (Yang et al. 2024, Industrial & Engineering Chemistry Research): a physics-informed neural network for large-scale pilot CHO fed-batch, embedding Monod kinetics into the network so it respects the bioreactor’s differential equations rather than only fitting the data. The stated payoff is extrapolation between sparse data points, which is exactly what a pure network fails at when experimental data is limited.
#204 (Chiu & Du 2025, Canadian Journal of Chemical Engineering): a neural-ODE model predictive controller regulating glucose in fed-batch CHO, with the network as the MPC’s prediction model over six state variables (viable cell concentration, mAb, glucose, glutamine, lactate, ammonium). It targets three problems simultaneously, control accuracy, model-plant mismatch and computational load, which existing MPC formulations trade against each other.
#257 (Wang et al. 2023, ACS Omega): probabilistic Bayesian deep learning for online forecasting of 2-keto-L-gulonic acid production by a mixed culture of K. vulgare and B. megaterium in industrial air-lift bioreactors. The operational detail explains the need: temperature, pH, aeration and level are monitored continuously by the control system, but product and substrate concentration still require manual offline assay, so a forecast is what stands in between. Returning a distribution rather than a point estimate is what makes it usable for a decision to stop or continue a batch.

Scaffolding

#20 (Rafieyan et al. 2024, Biofabrication): predicts the quality of 3D-(bio)printed scaffolds, and the dataset is the contribution as much as the model. Records were assembled from the literature by Scopus search and filtered hard for completeness, excluding any paper that did not specify printing conditions (pressure, temperature, nozzle diameter), scaffold properties (materials and compositions) or cell line and density, and any that reported them too vaguely to extract. That exclusion criterion is the reason the dataset is usable and a fair warning about how much published bioprinting work cannot be pooled. Also in Ensemble Learning. Code at saeedrafieyan/MLATE.
#215 (Shin et al. 2025, Advanced Science): optimizes five printing parameters (bioink viscosity, nozzle size, printing time, printing pressure, cell concentration) for droplet bioprinting of organoids at a target size, replacing manual pipetting that is labour-intensive and varies batch to batch. Consistency rather than peak performance is the objective, which is the right objective for manufacturing.
#216 (Golbabaei et al. 2024, RSC Advances): predicts morphological properties of electrospun scaffolds from both processing parameters and polymer chemistry, encoding the materials as SMILES so chemical composition enters the model directly rather than as a categorical label. Covers matrix polymers (PVA, PCL, chitosan, gelatin), conductive polymers (polypyrrole, PEDOT, PANI), solvents and dopants. The SMILES encoding is what lets the model generalize to a polymer it has not seen.

Sensory Prediction

#333 (Nozaki & Nakamoto 2016, PLOS ONE) and #338 (Nozaki & Nakamoto 2018, PLOS ONE): a two-paper progression worth reading in order. The first predicts odour impression from mass spectra with a nine-layer feed-forward network, applying dimensionality reduction to both the spectral and the sensory side before training to avoid the curse of dimensionality. The second changes the target rather than the model: instead of predicting individual descriptors, it clusters them by natural language processing into groups of larger granularity, so “rose”, “violet” and “lavender” fall together as something like “flower”. The motivation is that traditional aroma and fragrance wheels, built by experienced panellists, use at most a few tens of descriptors, so the raw descriptor space is finer than the perceptual one.
#325 Umami-MRNN (Qi et al. 2023, Food Chemistry): a merged MLP and recurrent network for umami-peptide prediction over the UMP-499 dataset with six feature representations, published with a web server. The MLP-plus-RNN hybrid handles fixed-length descriptors and variable-length sequence together.
#326 (Bo et al. 2022, Food Research International): three structure-taste models, bitterant versus non-bitterant, sweetener versus non-sweetener, and bitterant versus sweetener, across CNN, MLP-descriptor and MLP-fingerprint architectures, over BitterDB, SuperSweet and FlavorDB. The curation is instructive: mixtures of two or three compounds were removed, and so were compounds reported as both bitter and sweet, which is a real phenomenon being excluded to make the label space clean.
#330 (Sharma et al. 2021, JCIM): dense and random-forest models on physicochemical properties and fingerprints, alongside a CNN on 2D chemical images. Described in CNN.
#236 (Tac et al. 2026, npj Science of Food, Kuhl group): learns the structure of the human palate from large-scale recipe data and generates novel burgers optimized for deliciousness, sustainability or nutrition. The architecture handles a real awkwardness in the data: a recipe is a hybrid discrete-continuous object, needing both ingredient selection and ingredient quantification, so a two-stage diffusion framework decouples them, a multinomial diffusion model generating the binary ingredient mask and a score-based model generating weights conditional on that mask. As a sanity check the model rediscovers the Big Mac without supervision. It is the row’s clearest example of generative design aimed at the eating experience rather than at a molecule.

Metabolic Modeling

#145 MetaGEM (Xiao et al. 2026, arXiv): reconstructs genome-scale metabolic models bottom-up by predicting enzyme-metabolite interactions, framing GEM reconstruction as deep link prediction between an enzyme sequence space and a metabolite chemical space. Its argument against the conventional route is precise: top-down reconstruction leans on sequence homology and leaves unknown enzymes and metabolic dark matter unresolved, while going directly from metabolomics is an ill-posed inverse problem with combinatorial ambiguity and spurious networks, so enzymes are used as physical anchors between the two. The dataset was curated under a homology-aware protocol specifically to avoid inheriting sequence-alignment bias. Relevant to any cell-ag species without a curated GEM.

AI Tooling / Methodology

#4 D-GEX (Chen et al. 2016, Bioinformatics): infers the expression of roughly 21,290 target genes from 978 landmark genes, the computational half of the NIH LINCS cost-reduction strategy, trained on 129,158 GEO expression profiles. A cost-per-sample argument rather than a modelling one, and the same trade a cell-ag lab makes when deciding how deeply to profile each condition. Code at uci-cbcl/D-GEX.
#197 Epicure (Radzikowski & Chen 2026, arXiv): skip-gram ingredient embeddings trained from scratch on 4,135,189 recipes aggregated from 11 datasets across seven languages, normalized through an LLM-augmented pipeline to 1,790 canonical ingredients, then used to seed three Metapath2Vec variants over a 203,508-edge ingredient co-occurrence graph and an 80,019-edge FlavorDB ingredient-compound graph. Note the placement: Metapath2Vec is a non-learned random-walk embedding rather than a trained message-passing network, so this sits here and not in GNN, which is precisely the distinction that row’s definition draws. Useful for ingredient pairing and cuisine navigation in flavour design.
#289 AllerTrans (Sarlakifar et al. 2025, Biology Methods and Protocols): predicts protein allergenicity from sequence, extracting embeddings with the ESM-2 and ProtT5 protein language models and training a classifier on them, which keeps it in this row rather than a foundation-model one since the pretrained models supply features to a task-specific network. Directly relevant to cellular agriculture: recombinant growth factors, scaffold proteins and novel food proteins all need an allergenicity assessment, and laboratory testing is slow and expensive. A published correction applies to this paper (10.1093/biomethods/bpaf076); read it alongside the original, since it revises a methodological detail of the reported validation. Code at faezesarlakifar/AllerTrans.

Adjacent methods

CNN, GNN, GAN / VAE: the architecture-specific rows a paper moves to when one of them describes its core model.
The four foundation-model rows (Next-Token Prediction, Masked Language Modeling, Cell-State & Perturbation Prediction, LM + Biological Priors): where pretrained-and-transferred models go.
Chemometrics, SVM, Ensemble Learning: the classical comparisons, which win more often in this corpus than the volume of deep-learning papers would suggest.
Reinforcement Learning and Bayesian Optimization: the control and experiment-design layers built on top of these networks as surrogates.