SVM

This page describes the SVM row of the Papers.md matrix: support vector machines and support vector regression, kernel methods for classification and regression on modest, well-featurized datasets. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The kernel and the margin are what put a paper here, which separates the row from Linear & Regularized Models (no kernel), Ensemble Learning (trees and boosting) and Chemometrics (latent-variable projection). “Modest, well-featurized” in the taxonomy is doing real work: nearly every reference here has a few hundred samples with hand-built or spectral features, which is precisely the regime where an SVM is a sensible default and a deep network is not.

That makes the row a good place to look for the field’s honest small-data evidence. Several of these papers set out to test whether a neural network beats an SVM on their data, and report that it does not. It also means the row overlaps heavily with its neighbours: most references appear in a comparison table alongside them.

A second, less obvious use recurs here and is worth naming separately: the one-class SVM, trained only on normal data and used to flag anything unlike it. Two of the bioprocess references use it for contamination detection, which is a different job from prediction and one directly relevant to a cultivated-meat facility.

Media Optimization

#21 (Xu et al. 2014, BioMed Research International): optimizes a serum-free medium for suspension-adapted CHO cells by pairing a Plackett-Burman screening design with support vector regression, with a genetic algorithm tuning the SVM. The two-stage structure is the point: the Plackett-Burman design screens seven supplements at two levels each to find which matter, deliberately trading the ability to resolve interactions for the ability to eliminate factors cheaply, and the SVR model then confirms the same ranking independently. ZnSO₄, transferrin and BSA came out as the important ones. The cells were adapted to serum-free conditions by halving serum every three passages until below 0.1%, the same gradual-adaptation discipline the Cosenza work uses.

Bioprocess & Scale-Up

#31 (Zhang et al. 2020, Bioresource Technology): microbial lipid production from cellulosic ethanol wastewater by Rhodotorula glutinis, reporting directly that on a small experimental dataset the SVM fitted and generalized better than a back-propagation neural network, with a genetic algorithm then optimizing over the SVM to reach 11.87 g/L biomass and 2.18 g/L lipid. Also in Genetic Algorithms.
#32 (Roell et al. 2022, Biochemical Engineering Journal): benchmarks seven algorithm families on syngas fermentation, where support vector machines and random forests generalized best to unseen gas compositions (test R² ≈ 0.85) while neural networks overfit. Described in K-Nearest Neighbors; also in Ensemble Learning. Code at garrettroell/SyngasMachineLearning.
#208 (Xu et al. 2025, Biotechnology Progress): the CHO Raman and capacitance monitoring comparison, in which the SVM sits among the advanced ML methods beating PLS regression on real-time viability prediction. Described in Chemometrics; also in Ensemble Learning.
#259 (Pandi Chelvam et al. 2025, Scientific Reports, MIT and A*STAR): detects microbial contamination in cell-therapy manufacturing from UV absorbance spectra using a one-class SVM, trained on sterile spectra so anything unlike them is flagged. Seven organisms were spiked into mesenchymal-stromal-cell supernatant from six commercial donors, and contamination was detected at inoculums as low as 10 CFU, with mean true-positive and true-negative rates of 92.7% and 77.7%, the latter rising to 92% once certain donor samples were excluded. It is label-free, needs under 1 mL, and returns an answer in under 30 minutes, against the days a conventional sterility test takes. The mechanism is chemical rather than optical trickery: LC-MS confirmed contaminated cultures shift the nicotinamide-to-nicotinic-acid balance. Directly transferable to cultivated-meat quality control, where contamination is a batch-loss event and the current test is slow.
#262 (Nguyen et al. 2025, Bioprocess and Biosystems Engineering): the same problem at industrial scale, over 246 real fermentation batches from Novonesis of which 23 were contaminated, comparing a one-class SVM against autoencoders. Reported recall up to 1.0 for contaminated batches while keeping precision at 0.96 and specificity at 0.99 on healthy ones, which is the balance that matters when false alarms cost a discarded batch. The paper spends unusual effort on two things worth copying: hyperparameter optimization run in parallel with Optuna using Bayesian optimization with Hyperband, and a data-preprocessing account of genuinely messy industrial data, where batches have different durations, different variables, different timestamp columns and assorted missing and invalid values. Most published fermentation ML does not confront that.

Sensory Prediction

#26 (Sun et al. 2023, Food Chemistry): off-flavour profiling of cultured salmonids from near-infrared hyperspectral imaging, over 107 fish spanning Atlantic salmon and rainbow trout from recirculating aquaculture systems, natural reservoirs and imports. Four feature-extraction algorithms feed four classifiers, PLS-DA, least-squares SVM, extreme learning machine and a 1D CNN. The comparison across both feature selection and model family, rather than one held fixed, is what makes it useful. Directly relevant to cultivated seafood, where off-flavour is a known risk and panel testing is expensive. Also in CNN.
#334 (Cao et al. 2023, Foods): hyperspectral texture prediction through the skin of 387 live common carp, where least-squares SVM was one of three best models alongside BP-ANN and PLS regression. Described in Chemometrics.
#11 (Shen et al. 2024, Food Chemistry): salted-goose breed discrimination, where the SVM is one of the base classifiers whose predictions are fused by dynamic weighting. Described in GAN / VAE; also in CNN and Ensemble Learning.
#28 (Sun et al. 2026, Journal of Food Engineering): algal odour-compound classification, with the SVM among six compared algorithms. Described in K-Nearest Neighbors; also in Ensemble Learning.
#72 (Colantonio et al. 2022, PNAS, Florida): predicts consumer sensory perception of fruit flavour from targeted metabolomics, across 147 tomato genotypes (68 sugars, acids and volatiles, 209 samples) and 63 blueberry genotypes (55 compounds plus firmness, 244 samples), with ground truth from panels averaging around 80 participants drawn from a diverse population. The problem it solves is the one every sensory-prediction programme has: flavour phenotyping is so expensive that a breeding programme cannot measure it at the scale it selects at, so a model that predicts liking, sweetness, sourness, umami and flavour intensity from chemistry lets selection happen anyway. That is the same bottleneck a cultivated-meat formulation programme faces. Also in Ensemble Learning. Code at Resende-Lab/metabolomic_selection_for_enhanced_fruit_flavor.
#271 VirtuousUmami (Pallante et al. 2022, Scientific Reports): predicts umami taste from a molecule’s SMILES rather than from an amino-acid sequence, which is the stated advance: every prior umami predictor was sequence-based and therefore could only score peptides. Built on the same UMP442 dataset as iUmami-SCM, with sequences converted to SMILES via RDKit and standardized through the ChEMBL Structure Pipeline, 352 ligands for training and 90 held out. Note the same negative-class caveat as the bitter-peptide work: UMP442’s non-umami peptides are the bitter peptides of BTP640. Code at lorenzopallante/VirtuousUmami.
#327 iBitter-Fuse (Charoenkwan et al. 2021, IJMS): bitter-peptide prediction by fusing multi-view features, amino-acid composition (20 dimensions), dipeptide composition (400), and physicochemical properties drawn from 544 AAIndex entries, on the BTP640 benchmark split 8:2. The contribution is the feature fusion rather than the classifier, and it is a clear example of the row’s characteristic move: on 640 sequences, engineering better features beats reaching for a bigger model.
#346 (Chéron et al. 2017, Food Chemistry): a QSAR model for sweetness over SweetenersDB, 316 compounds across seventeen chemical families with relative sweetness spanning 0.20 to 225,000, log-transformed because of that skew. Molecules were described in both 2D (SMILES) and 3D, with 3D descriptors Boltzmann-weighted over the three lowest-energy conformers and protonation set at salivary pH 6.5. Beyond the model it reports a structural generalization, that potent sweeteners tend to combine a hydrophobic scaffold with fewer than 4 hydrogen-bond donors and 10 acceptors at molecular weights of 350 to 450 g/mol, and screens the largest natural-compound database for sweetness, bitterness and toxicity together, finding most predicted natural intense sweeteners carry saponin or stevioside scaffolds. Also in Ensemble Learning.

Adjacent methods

Ensemble Learning, K-Nearest Neighbors, Linear & Regularized Models: the classical models the SVM shares a comparison table with in almost every reference here.
Chemometrics: PLS and PLS-DA, the linear-projection baseline on the same spectra.
Deep Learning and CNN: the comparison that these datasets repeatedly fail to justify, and the two rows where the hyperspectral work goes when it does.
Genetic Algorithms: the optimizer layered on top of an SVM surrogate in #21 and #31.

SVM

Scope boundary

Media Optimization

Bioprocess & Scale-Up

Sensory Prediction

Adjacent methods

Further reading