Ensemble Learning

This page describes the Ensemble Learning row of the Papers.md matrix: tree ensembles and model averaging, covering random forests, gradient boosting and XGBoost, AdaBoost, and stacked ensembles. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

Trees, bagging or boosting are what put a paper here as the primary learner, which separates the row from SVM (kernels and margins), Linear & Regularized Models (no trees) and Chemometrics (latent-variable projection). An ensemble assembled over networks that already hold their own architecture row does not additionally earn this one: what the row records is the learner, not the combination rule placed on top of it. The taxonomy’s own phrasing is the most useful thing about it: this is “the default strong baseline for tabular prediction”. A paper appears here because someone reached for the thing that usually works on a few hundred rows of features, and very often it did.

That makes it the largest row on the method axis, and a heavily concentrated one, with the great majority of its references in Sensory Prediction, where the data is almost always a table of molecular descriptors against a panel-derived label.

Read the membership of this row as provisional. Because it is the default baseline, papers land here whose principal model is arguably something else, and a re-audit of matrix placements is open work; a reference whose headline model turns out to be, say, a standalone Gaussian process rather than a tree ensemble would move out. Re-derive the current set from the matrix rather than treating the list below as closed.

Media Optimization

#169 (Hashizume & Ying 2025, New Biotechnology): the ensemble surrogate (gradient boosting, support-vector regression, k-nearest neighbors and a neural network) inside the 57-component CHO-K1 serum-free campaign. Described in Active Learning.
#170 (Gangwar et al. 2024, Applied Microbiology and Biotechnology): an end-to-end framework for CHO media component selection and prediction of a critical quality attribute, which is the pairing that matters industrially since a medium that grows cells well but shifts glycosylation is not an improvement. Fifteen model families are run across a five-stage pipeline of preprocessing, feature selection, optimization, model development and evaluation, with random forest, extra trees, CatBoost and gradient boosting alongside the linear and kernel alternatives. The paper’s framing is explainability: the point of the framework is to unravel which media components drive which attribute, not only to predict.
#213 (Grzesik & Warth 2021, Frontiers in Bioengineering and Biotechnology): a machine-learning pipeline layered onto design of experiments for advanced T-cell media, motivated by a specific failure of traditional DoE on primary cells: donor variability is large enough that a design fitted to one donor does not predict another. Cells were CD3+ T cells from healthy human donors, activated with CD3/CD28 beads in the presence of IL-7 and IL-15, with two to three repeats per condition. The relevance to cultivated meat is direct, since primary satellite cells carry the same animal-to-animal variability that makes a single fitted response surface untrustworthy.

Bioprocess & Scale-Up

#32 (Roell et al. 2022, Biochemical Engineering Journal): random forests were one of the two families that generalized to unseen syngas fermentation conditions. Described in K-Nearest Neighbors.
#208 (Xu et al. 2025, Biotechnology Progress): the random-forest regressor is the model carrying the reported 31.7% accuracy improvement over PLS for real-time CHO viability from Raman and capacitance spectra. Described in Chemometrics.
#207 (Behdani et al. 2024, Biotechnology Progress): a soft sensor for total sialic acid, a critical quality attribute affecting drug efficacy, replacing a wet chemical assay that takes hours because it requires isolating sialic acid from the protein, sample preparation and chromatographic separation. Separate models were built per measurement day (8, 10 and 14), with later days’ data withheld when training the earlier model so each is genuinely predictive rather than interpolating. Inputs are process scale, set points and real-time bioreactor parameters. The pattern, replacing a slow release assay with a model over data the reactor already produces, is the most portable idea in this column.
#27 (Du et al. 2025, npj Science of Food): predicts volatile-compound profiles in engineered S. cerevisiae fermentation simulating canned meat, sitting in both this column and Sensory Prediction. It is one of the clearest precision-fermentation-to-flavour references in the matrix: an engineered yeast strain deposited at CCTCC, grown in parallel bioreactors, with structured volatile data from canned meat and fermented products compiled to train the models. Code at illuminate6060/ML_S.cerevisiae.

Scaffolding

#214 (Nair et al. 2021, Applied Physics Reviews, Cambridge): random-forest-assisted optimization of ice-templated collagen scaffolds, and the paper in this row that uses trees for what trees are best at, feature importance over a large messy parameter space. Two cases are studied: primary and secondary drying across two temperatures and three chamber pressures, and solvent-solute interactions across four collagen concentrations and three solvents over 33 conditions. The stated problem is that the fabrication parameter space is large and complex enough that interpreting results and choosing parameters is hard in practice, which is a curation problem before it is a modelling one.
#20 (Rafieyan et al. 2024, Biofabrication): 3D-(bio)printed scaffold quality prediction over a hard-filtered literature dataset. Described in Deep Learning.
#217 (Serpe et al. 2025, ACS Omega): ultrasound imaging plus machine learning for nondestructive sensing in bioreactors, and it names cellular agriculture explicitly as a motivation alongside cell therapy. It characterizes cellular growth and decellularization in spinach-leaf scaffolds, a decellularized-plant substrate cultivated-meat groups use for vasculature, imaging fresh leaves, decellularized leaves and recellularized ones across a five-day SDS treatment. Ultrasound is the interesting choice: unlike optical methods it penetrates an opaque construct, which is what a thick tissue needs.
#309 (Dahl et al. 2025, Food Hydrocolloids): predicts rheological parameters of plant-protein biopolymer mixes from composition alone, over formulations spanning 14 to 43% protein from yellow pea and faba bean isolates and concentrates, mixed with 0 to 13% of maize starch, pectin, cellulose or carrageenan. Rheology governs how a mixture behaves in hydrothermal structuring, so predicting it from composition short-circuits an iteration that otherwise needs a rheometer per formulation.
#171 (Kircali Ata et al. 2023, Foods): XGBoost among the models predicting hardness and chewiness of meat analogues from proximate composition, evaluated leave-one-group-out. Described in Linear & Regularized Models.

Sensory Prediction

The row’s centre of gravity. Almost all of it is structure-to-taste or structure-to-odour prediction over curated compound databases, which is worth stating plainly because it means these papers share not just a method but a substrate: BitterDB, SuperSweet, SweetenersDB, ChemTastesDB, OlfactionBase and the Good Scents Company database recur throughout, and much of the apparent progress across a decade is progress in curation as much as in modelling.

Bitterness

#102 BitterPredict (Dagan-Wiener et al. 2017, Scientific Reports, Niv lab): the lineage’s starting point, an AdaBoost classifier over molecular descriptors flagging a compound as bitter, correctly classifying over 80% of held-out compounds. The preparation is documented in detail and sets the pattern the rest of this cluster follows: largest fragment retained, salts and peptides removed, 3D structures and protonation states generated at pH 7.0 ± 0.5, original chirality kept, lowest-energy conformer extracted, and both forms retained where a molecule has two protonation states in range.
#103 BitterIntense (Margulis et al. 2021, Computational and Structural Biotechnology Journal): moves from bitter-or-not to how bitter, an XGBoost classifier reaching over 80% accuracy. Its handling of ground truth deserves attention: labels come from a rat brief-access taste aversion assay, and the paper reasons explicitly about the offset between species and between measures, using 3 mM IC50 in rat and a 0.1 mM human recognition threshold as the two thresholds, noting rats are slightly more tolerant of bitterness than humans and that an IC50 and a recognition threshold are not the same quantity. That kind of stated correction is rare and makes the labels reusable.
#105 BitterMasS (Ziaikin et al. 2024, Journal of Agricultural and Food Chemistry): predicts bitterness directly from mass spectra, bypassing structural assignment entirely, which is what lets bitterness be tracked through an untargeted MS workflow as a sample changes across processing. Trained on BitterDB plus BitterIntense compounds, with several molecules previously mislabelled non-bitter (hydrogen peroxide, hydrogen cyanide, phenylthiourea among them) excised, a small correction with an outsized effect on a binary classifier.
#332 e-Bitter (Zheng et al. 2018, Frontiers in Chemistry): consensus voting across several machine-learning methods, and notable for what it fixed in the field’s data rather than its model. Earlier bitterant work used non-bitterant sets that were wholly or partly hypothetical; this paper assembles a fully experimental set of 707 bitterants and 592 non-bitterants under four stated criteria (no disconnected structures such as salts, only common elements, exclusion of compounds labelled with conflicting taste qualities, and de-duplication across sources). Any comparison against pre-2018 bitterness numbers should account for that difference in negative class.

Sweetness

#346 (Chéron et al. 2017, Food Chemistry): SweetenersDB, 316 compounds across seventeen chemical families spanning relative sweetness 0.20 to 225,000. Described in SVM.
#345 (Bouysset et al. 2020, Food Chemistry): updates SweetenersDB and builds a model on open-source molecular features with a public web server, then does the thing most of this cluster does not: validates in vitro, showing cellular functional assays in which the sweet-taste receptor is activated by a newly identified natural-compound scaffold. A prediction confirmed at the receptor is a different class of evidence from a held-out accuracy.
#336 (Goel et al. 2023, Computers in Biology and Medicine): regression rather than classification, over a manually curated set of 671 sweet molecules with experimental sweetness values spanning 0.2 to 22,500,000. Gradient Boost and Random Forest regressors came out best at correlation 0.94 and 0.92. Predicting degree of sweetness is what a low-calorie-sweetener search actually needs.
#337 e-Sweet (Zheng et al. 2019, Frontiers in Chemistry): applies the same consensus-voting protocol as e-Bitter, over KNN, SVM, GBM, RF and a deep network with ECFP descriptors, adapted from classification to regression to predict both sweetener identity and relative sweetness, over 530 sweeteners curated from SuperSweet and SweetenersDB.
#344 (Yang et al. 2022, Food Chemistry): a multi-layer sweetness evaluation system built to work across chemical spaces and categories that are usually modelled separately, natural against artificial, carbohydrate against non-carbohydrate, nutritive against non-nutritive. MOE2d descriptors plus Atompairs, ECFP4 and MACCS fingerprints, with highly correlated (threshold 0.95) and zero-variance descriptors removed.

Bitter versus sweet, and multi-taste

#268 BitterSweet (Tuwani et al. 2019, Scientific Reports, Bagler lab): models the bitter-sweet axis together rather than as two independent problems, curating from nine sources including Polya’s Biochemical Targets of Plant Bioactive Compounds, BitterDB, Fenaroli’s Handbook of Flavor Ingredients, SuperSweet, TOXNET and the Good Scents database, specifically to span the chemical space rather than one corner of it. Code at cosylabiiit/bittersweet; catalogued in Software.md.
#331 BitterSweetForest (Banerjee & Preissner 2018, Frontiers in Chemistry): a random-forest binary classifier over molecular fingerprints from 1,202 compounds (517 sweeteners, 685 bitterants), distributed as an open KNIME workflow, with standardization steps (water removal, aromatization, charge normalization, explicit hydrogens) and InChIKey-based de-duplication documented.
#343 (Maroni et al. 2022, Current Research in Food Science): an explainable sweet-bitter classifier over 3,130 compounds (1,764 sweet, 1,366 bitter) pooled from nine prior databases, with the explicit aim of enriching understanding of the molecular basis rather than only raising accuracy, which the authors note prior classifiers already achieved.
#342 (Androutsos et al. 2024, npj Science of Food): multi-objective prediction across sweet, bitter, umami and other. Described in Genetic Algorithms.

Umami

#328 UMPred-FRL (Charoenkwan et al. 2021, IJMS): umami-peptide prediction by feature representation learning on the same UMP-TR and UMP-IND split as iUmami-SCM, 140 umami and 304 non-umami peptides. The shared benchmark across #269, #271, #322 and this paper makes the umami cluster unusually comparable, with the standing caveat that its negative class is bitter peptides.

Odour and aroma

#80 (Keller et al. 2017, Science, DREAM Olfaction Prediction Consortium): the crowd-sourced challenge that set the field’s benchmark, with teams predicting sensory attributes from chemoinformatic features against psychophysical data from 49 individuals rating 476 molecules. Odour intensity and pleasantness were predicted accurately, and 8 of 19 rated semantic descriptors successfully. The methods section documents what was deliberately excluded from the challenge data (familiarity and edibility ratings, recognition, free-text descriptions, and four molecules), which matters for anyone reusing it. This is the substrate the later graph-based work in GNN is measured against.
#323 (Tyagi et al. 2024, Journal of Biomolecular Structure and Dynamics): an XGBoost odour-prediction model over 1,278 odorants, restricted to the seven odour descriptors appearing at least 50 times in OlfactionBase. That threshold is the paper’s implicit finding: the descriptor vocabulary has a long tail that no model trained on this data can reach, which is the same limit #338 approaches by clustering descriptors instead.
#329 (Wang et al. 2021, RSC Advances): predicts both flavour class (aromatic, bitter, sulfury, other) and retention index for beer compounds from structure, using NIST retention indices on non-polar SE-30 and polar Carbowax 20M columns. Predicting retention index is the practical half: it is what lets a peak in a GC trace be tentatively assigned before a standard is run, which is the routine bottleneck in any volatile-profiling workflow.
#28 (Sun et al. 2026, Journal of Food Engineering): random forest and gradient boosting among the six algorithms compared on algal odour compounds. Described in K-Nearest Neighbors.
#11 (Shen et al. 2024, Food Chemistry): the dynamically weighted fusion of base classifiers whose ablation showed the ensemble step was indispensable to generalization. Described in GAN / VAE.
#27 (Du et al. 2025, npj Science of Food): volatile-profile prediction in engineered yeast fermentation, described under Bioprocess & Scale-Up above.

Texture and whole-food properties

#72 (Colantonio et al. 2022, PNAS): metabolomic selection for fruit flavour, predicting consumer panel ratings from targeted metabolomics in tomato and blueberry. Described in SVM.
#171 (Kircali Ata et al. 2023, Foods): meat-analogue hardness and chewiness from composition, with carbohydrates, fat and targeted moisture content the dominant features. Described in Linear & Regularized Models.
#341 (Oppen et al. 2023, Food Research International): the most unusual input in the row. Rather than measuring texture by deformation force or a panel, it builds a pipeline over mastication physics, the measured dynamics of 11 panellists’ jaw movement, to identify samples differing in particle size and fibrousness. The argument is that meat and meat alternatives differ in more than texture, which makes an isolated texture impression hard to extract by conventional means, so measuring how the food is actually chewed sidesteps the confound. Directly relevant to cultivated and hybrid products, where the eating experience is the acceptance criterion.

Food Safety Prediction

#290 AlgPred 2.0 (Sharma et al. 2021, Briefings in Bioinformatics, Raghava lab): predicts allergenic proteins and maps IgE epitopes, trained on a balanced set of 10,075 allergens and 10,075 non-allergens assembled from COMPARE, AllergenOnline, the original AlgPred and AllerTOP datasets and Swiss-Prot, plus 10,451 experimentally validated IgE epitopes for identifying antigenic regions, with non-standard-character and short sequences filtered out. Allergenicity assessment is a regulatory requirement for novel food proteins and recombinant growth factors, so this and #289 AllerTrans in Deep Learning are two routes at the same gate. The tool is catalogued in Software.md.

Adjacent methods

SVM, K-Nearest Neighbors, Linear & Regularized Models, Chemometrics: the classical families sharing a comparison table with tree ensembles in nearly every reference here.
Deep Learning, CNN, GNN: the neural alternatives, which win on the larger odour datasets and rarely on the smaller taste ones.
Active Learning and Bayesian Optimization: where a tree ensemble is used as the surrogate inside an experiment-selection loop rather than as a final predictor.
Genetic Algorithms: the search layer over these surrogates.