Sensory Prediction

The prediction of sensory attributes is a critical challenge for cellular agriculture and alternative-protein development. Consumer acceptance is driven in large part by taste, smell, and texture — and bridging the gap from cell-level engineering (cell type, media formulation, bioprocess conditions, scaffolding choices) to the final organoleptic experience is one of the central commercial barriers in the field. The GFI 2024 State of Alternative Proteins and the NECTAR Taste of the Industry 2024 sensory benchmark both flag taste parity as the binding constraint for plant-based and cultivated meat products. AI offers one promising toolkit for compressing the experiment-and-iterate cycle by modeling the relationship between molecular composition and human perception.

Applied AI for cell-ag and alt-protein flavor

Recent work demonstrates the practical application of ML methods to cultivated and alternative-protein flavor problems. Sun et al. (2023, ref #26) developed a CNN to profile off-flavors in cultured salmonids using hyperspectral imaging, providing a quality-control pipeline that bypasses expensive panel work (see Datasets/Fish.md for finfish data resources). Shen et al. (2024, ref #11) combined chemometric methods with a GAN-based integrated deep-learning framework to discriminate salted goose breeds — illustrating how generative architectures can extend limited sensory training data. Du et al. (2025, ref #27) used ensemble machine learning to predict volatile compound profiles in Saccharomyces cerevisiae fermentation simulating canned meat, a directly precision-fermentation-relevant application. Sun et al. (2026, ref #28) leveraged ML for odor control in algal foods. Together these papers establish that ML approaches — CNN, GAN/VAE, ensemble, and k-NN — work on the kinds of metabolomic and sensory data that alt-protein labs already collect. ML is also reaching the texture and appearance dimensions of the eating experience: Kircali Ata et al. (2023, ref #171) predicted the hardness and chewiness of plant-based meat analogs from proximate composition (comparing several models, including ridge regression, XGBoost, and a neural network), identifying carbohydrate, fat, and targeted moisture as the dominant levers for tuning texture, while Ulucan et al. (2019, ref #195) trained a convolutional neural network to classify red-meat freshness from RGB images — an image-based quality-control proxy transferable to non-destructive monitoring of cultivated product. The open training data for that kind of classifier is now available for beef specifically: Gyening et al. (2025, ref #196) released MeatScan, an 11,000-image collection of fresh and spoiled raw cow meat photographed in real-world Ghanaian markets, butcher shops, and cold storage, deliberately built under uncontrolled lighting to train robust fresh-versus-spoiled classifiers (the dataset itself is inventoried in Datasets/Cow.md). At the product-design end, Tac, Gardner & Kuhl (2026, ref #236) treat the human palate as a high-dimensional probability distribution learned from large-scale recipe data and sample it to generate burger recipes optimized jointly for deliciousness, sustainability, and nutrition; in a blinded 101-participant tasting their generated burgers matched or beat a Big Mac benchmark, while a mushroom variant carried more than 10× lower environmental impact and a bean variant roughly double the nutritional score. The system is recipe- and ingredient-level rather than cell-cultured, but it is a concrete demonstration of generative AI optimizing the sensory endpoint that cultivated and alternative products ultimately have to hit.

Foundational AI for olfactory perception

A second line of work develops AI methods for olfactory perception itself, with the goal of replacing or augmenting expensive human sensory panels. Lee et al. (2023, ref #14) introduced the Principal Odor Map using a graph neural network trained on >5,000 molecule-odor pairs (Monell dataset); the model outperformed human panels on unseen molecules and generalized to detection thresholds and cross-species olfaction. Qian et al. (2023, ref #36) dissected the Principal Odor Map to show that metabolic activity organizes olfactory representations — a follow-up that bridges molecular chemistry, evolutionary biology, and ML interpretability. Going further back, Keller et al. (2017, ref #80) launched the DREAM Olfaction Prediction Challenge with crowdsourced ML models predicting human olfactory perception from chemical features — a foundational benchmark dataset. On the breeding-and-flavor side, Colantonio et al. (2022, ref #72) used 18 ML models on tomato and blueberry metabolomes paired with consumer-panel ratings, capturing up to 56% of variance in consumer liking from volatiles alone — the closest published reproducible metabolome-to-flavor pipeline to date.

Computational prediction of taste and off-flavor

Distinct from aroma, a focused line of work predicts taste — especially bitterness, the off-flavor most likely to limit acceptance of novel protein ingredients — directly from chemical structure or spectra. A coherent lineage from the Niv lab builds this out: Dagan-Wiener et al. (2017, ref #102) introduced BitterPredict, an AdaBoost classifier over molecular descriptors that flags a compound as bitter or not, correctly classifying over 80% of a held-out test set. Margulis et al. (2021, ref #103) extended this to bitterness intensity with BitterIntense, an XGBoost classifier reaching over 80% accuracy, and Margulis et al. (2022, ref #104) added receptor specificity with BitterMatch, a gradient-boosting model predicting which TAS2R bitter-taste receptors a ligand activates (~80% precision at ~50% recall). Most relevant to a metabolomics workflow, Ziaikin et al. (2024, ref #105) trained BitterMasS, a random-forest classifier, to predict bitterness directly from mass spectra (test precision 0.83, recall 0.90) — bypassing structural assignment, so bitterness can be tracked straight from untargeted MS data as a cultivated or alt-protein sample changes across processing. These models, anchored in the BitterDB compound-and-receptor database, give cell-ag formulators an in-silico screen for the single most common aversive taste.

Where that lineage scores small molecules, a parallel line targets bitter peptides: the form of bitterness that arises when proteins are hydrolyzed, and thus directly relevant to cultivated-meat and alternative-protein ingredients. Srivastava et al. (2024, ref #286) introduced BitterPep-GCN, a graph convolutional network that classifies a peptide as bitter or non-bitter from its sequence, learning amino-acid embeddings via graph convolution rather than relying on hand-engineered descriptors. Steuer et al. (2026, ref #285) then closed the loop from prediction to design: they fine-tuned ZymCTRL, a conditional protein language model, to generate de novo peptide candidates and used BitterPep-GCN to filter them, then validated 31 of the resulting candidates by expert taste panel (25 correctly classified: 15 confirmed bitter, 10 confirmed non-bitter). Together they extend the in-silico bitterness screen from single tastants to the peptide space that protein hydrolysis releases, the setting where unpredictable bitter-peptide formation limits acceptance of novel protein ingredients.

Sensomics methodology and reference work

The Schieberle / Hofmann school at TU Munich and the Leibniz-Institute for Food Systems Biology established the molecular sensory science paradigm: from the ~10,000 volatiles found in foods, only ~226 “Key Food Odorants” (KFOs) account for the aroma of ~230 documented foods, identifiable through GC-O + Aroma Extract Dilution Analysis + stable-isotope dilution + Odor Activity Value (OAV) ranking + recombination / omission tests. Nicolotti, Mall & Schieberle (2019, ref #73) introduced SEBES (Sensomics-Based Expert System), automating this workflow with GC×GC-TOF-MS + GC Image + OAV computation — though the implementation is not released as open-source software.

For cell-ag specifically, the foundational sensory-science work is Lew, Yuen, Zhang, Fuller, Frost & Kaplan (2024, ref #75) — the first GC-MS + GC-O + descriptive-panel characterization of cultivated porcine adipose tissue as a flavor enhancer for meat alternatives (Tufts group; see Datasets/Pig.md for porcine data resources). Zhou et al. (2025, ref #193) complements that tissue-level work at the cell level: spontaneously immortalized porcine myoblasts and fibroblasts adapted to suspension culture, processed via freeze-thaw flavor-precursor isolation followed by optimized Maillard and lipid thermal-degradation reactions, deliver a hybrid cultivated meat at 1.2% w/w cell incorporation with 78.5% sensory similarity to pork and an 80% production-cost reduction; myoblasts outperform fibroblasts on aroma fidelity. On the precision-fermentation side, Spaccasassi et al. (2024, ref #74) demonstrate microbial starter screening for pea-protein-based beverages using UHPLC-TOF-MS plus sensory profiles. And O’Neill et al. (2022, ref #77) report spent-media analysis suggesting cell-ag media will require species- and cell-type-specific optimization — directly tying sensory-relevant metabolite profiles to media formulation strategy.

Tools and data

The analytical stack for sensory prediction draws on mass-spectrometry preprocessing, chemometrics, and curated flavor databases:

Mass-spec tooling: OpenMS / pyOpenMS and the broader Mass Spectrometry & Chemometrics section in Software.md.
Chemometrics: ropls (PCA / PLS / OPLS / OPLS-DA), used as the multivariate engine in Workflow4Metabolomics.
Flavor databases: FlavorDB / FlavorDB2, BitterDB, and Pherobase — see the Flavor & Taste Compound Databases section in Databases.md.
Species metabolomes: HMDB and the Bovine Metabolome Database for species-specific metabolite profiling that grounds OAV calculations in physiologically relevant concentrations.

Open challenges for cell-ag

Several gaps remain conspicuous. The 2024 organoleptic-analysis landscape (see also #76 Wang, #78 Mittermeier, and #79 Alasi reviews) flags: no purpose-built reproducible workflow-manager pipeline (Nextflow / Snakemake / CWL) for organoleptic analysis; no standardized sensory data exchange format (no mzTab equivalent for sensory panels); a structural “two cultures” gap between sensometrics (food-science / Pangborn tradition) and the bioinformatics workflow-manager community; vendor-software dominance (Agilent, Thermo, LECO, Bruker, Waters) that constrains open-pipeline adoption; and confidentiality of industry data (Givaudan, Symrise, IFF, dsm-firmenich, Mane) that limits public benchmarks. For cell-ag specifically there is no equivalent of the BiGG Models / Cellosaurus / Human Cell Atlas canonical home for sensory-and-flavor data from cultivated tissues — most cultivated-meat sensory data is locked behind company R&D walls. A reproducible Snakemake or Nextflow pipeline that fuses GC-MS volatile data, microbiome data, and trained-panel scores into an OAV-ranked sensomics report is a substantial open opportunity, and one that this repo’s curated set of tools, databases, and reference work would directly support.