Chemometrics

This page describes the Chemometrics row of the Papers.md matrix: multivariate statistics for extracting quantitative information from chemical measurements, overwhelmingly spectra. The workhorses are partial least squares regression, PLS-DA, PCA and OPLS. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The taxonomy separates this row from its neighbours by mechanism: latent-variable linear projection, not learning by kernels (SVM), trees (Ensemble Learning) or backpropagation (Deep Learning). That matters because chemometric papers almost always benchmark against those alternatives on the same spectra, so a reference here usually appears in one of those rows too, and the placement records which method the paper’s contribution actually rests on.

For cellular agriculture this row is the backbone of spectroscopy-based Process Analytical Technology: measuring glucose, lactate and biomass in a running bioreactor without pulling a sample, and predicting quality attributes of the product without destroying it. It is the oldest and least fashionable body of work on the method axis, and the most likely to already be installed in a manufacturing facility.

Media Optimization

#287 (Wang et al. 2026, Journal of Agricultural and Food Chemistry): untargeted metabolomics on conditioned medium from an 80 L stirred-tank bioreactor running cultured-meat production, sampled at days 0, 3, 6, 9 and 12. This is a rare thing in the matrix: nutrient consumption measured at pilot scale in an actual cultivated-meat process rather than in a well plate. Of 1,761 detected metabolites, leucine and arginine were picked out by variable-importance-in-projection scores, the standard PLS-derived ranking, together with pathway enrichment; targeted supplementation of both was then validated in 2D monolayer and 3D suspension systems for proliferation and stemness. The VIP score is what places this in the chemometrics row rather than a modelling one: the multivariate projection is being used to decide which of 1,761 variables to act on.

Bioprocess & Scale-Up

#7 (Tamburini et al. 2014, Sensors): on-line and in-line near-infrared monitoring of cell, substrate and product concentrations across aerobic and anaerobic bacterial fermentations, with the calibration models then used for automated process control rather than only for reporting. The paper is unusually explicit about why bioprocess calibration is hard, naming both the chemical problem (ill-defined complex medium) and the physical one (a multiphase matrix), which is exactly the situation in a cultivated-meat reactor with serum-free medium and microcarriers. It also tests whether in-line NIR can discriminate between differently shaped bacteria, which is the same question as distinguishing cell states from a spectrum.
#208 (Xu et al. 2025, Biotechnology Progress): the modern comparison, and the reason this row’s boundary is worth stating carefully. Over eight batches of recombinant CHO culture in 2-L bioreactors, with online Raman (785 nm, full 100 to 3425 cm⁻¹ range) and capacitance spectra (0.3 to 1000 MHz) and 120 offline reference samples, it benchmarks advanced ML against PLS regression, the traditional linear baseline. The advanced algorithms win, with a reported 31.7% accuracy improvement for a random-forest regressor. The problem being solved is real and specific: viable cell density is already monitorable in real time, but viability is not, and that is the gap. Also in Ensemble Learning and SVM. Read this alongside #7 to see the decade-long shift from PLS as the method to PLS as the baseline.

Sensory Prediction

#335 (Liu et al. 2003, Meat Science): the earliest reference in the row, predicting colour, instrumental texture and sensory attributes of beef steaks from visible and near-infrared reflectance over 400 to 1080 nm, across 24 carcasses aged 2, 4, 8, 14 and 21 days post mortem. It is worth reading for its calibrated honesty about what spectroscopy can and cannot reach: colour predicted well (R² 0.78 to 0.90), while tenderness, Hunter L, chewiness and juiciness sat at 0.49 to 0.55, and tenderness prediction ranged from 0.22 to 0.72 depending on aging day. Reframed as a two-class tender-versus-tough decision, the PLS model classified 83% correctly. The lesson that a hard regression becomes a workable classification is directly transferable to cultivated-meat quality control.
#347 (Chen et al. 2020, Journal of Food Engineering): Raman prediction of texture in frozen and thawed raw beef, over 130 subsamples from 16 M. longissimus dorsi muscles under continuous freezing or repeated freeze-thaw cycles. PLS models reached R²p of 0.81 to 0.82 for tenderness, chewiness, firmness and hardness, and weaker performance for springiness. It also reports a process finding independent of the model: texture changed significantly only once freeze-thaw cycles exceeded three. Relevant to any cultivated product that will be frozen in distribution.
#334 (Cao et al. 2023, Foods): hyperspectral imaging of 387 live common carp, scanning skin at four body regions (400 to 1000 nm, 2.5 nm resolution) to predict eight texture indicators of the underlying muscle. Predicting muscle properties non-invasively through the skin of a live animal is a harder problem than scanning a fillet, and the paper compares six models, with backpropagation ANN, PLS regression and least-squares SVM identified as the best. Also in SVM. The cell-ag reading is about the readout rather than the fish: a hyperspectral scan of a growing construct is the same measurement.
#348 (Ma et al. 2017, LWT): hyperspectral prediction of texture in grass carp fillets during vacuum freeze drying, across 112 fillets dried for 3 to 36 hours to generate seven texture levels. The methodological contribution is wavelength selection: PLSR is fitted on the full range, regression coefficients identify the important wavelengths for each of shear force, hardness, gumminess and chewiness separately, and those are then integrated into one shared group so a single sensor can predict several attributes at once. That is the step that turns a lab hyperspectral camera into a plausible in-line instrument.
#340 (Rojas et al. 2017, Frontiers in Chemistry): a QSTR expert system for sweetness in which PLS-DA is the linear half of a consensus with a nearest-neighbour classifier, applied to the region of fingerprint space where taste assignment is hardest. Described under K-Nearest Neighbors.

AI Tooling / Methodology

#198 ropls (Thévenot et al. 2015, Journal of Proteome Research): both a metabolomics study and the reference implementation of an OPLS workflow. It characterizes how the human urinary metabolome varies with age, BMI and gender across 183 adults by LC-HRMS, with the explicit purpose of establishing baseline physiological variation so cohort studies do not mistake it for signal. The curation is careful in a way worth noting, excluding five high-BMI volunteers as statistical outliers with too few subjects to model robustly. The comprehensive univariate-plus-OPLS workflow was released to the community through Workflow4Metabolomics and as the Bioconductor package ropls, catalogued in Software.md. Any spent-media or off-flavour metabolomics analysis in cell-ag is likely to run through this or its equivalent.
#199 mixOmics (Rohart et al. 2017, PLoS Computational Biology): an R package for multivariate feature selection and integration across several omics layers at once, addressing the gap that most methods find a molecular signature within a single omics type and treat features independently. Eighteen projection-based methods are implemented, twelve of them variants of the (mint).(block).(s)pls(da) family built as wrappers over one generic sGCCA-based routine, plus PCA, sparse PCA, IPCA, rCCA and rGCCA. Catalogued in Software.md. The relevance for cellular agriculture is direct wherever transcriptomic, proteomic and metabolomic readouts of the same culture need to be analysed together rather than in sequence.

Adjacent methods

SVM, Ensemble Learning, Deep Learning: the three families PLS is benchmarked against on the same spectra, and increasingly beaten by when enough data exists.
K-Nearest Neighbors: the local half of the consensus model in #340.
CNN: where hyperspectral work goes when the spatial structure of the image, rather than the mean spectrum, is what carries the signal.
Foundation Models (other modalities): SpectraLLM and the end-to-end spectrum-to-structure models that bypass the latent-variable step entirely.