Chemometrics turns biosensors from simple signal transducers into intelligent analytical systems

Biosensors have become remarkably good at detecting tiny amounts of clinically important molecules, pathogens, pollutants and food contaminants, with modern devices pushing detection limits into the sub-femtomolar range thanks to advances in nanomaterials, surface engineering and signal amplification. Yet a persistent gap separates laboratory prototypes from robust real-world devices. In the messy conditions of actual samples, measurement reliability is systematically undermined by signal drift, surface fouling, non-specific adsorption, cross-reactivity and matrix interferences, while device-to-device and batch-to-batch variability drive up recalibration requirements. A new perspective article, published on 15 June 2026 in the open-access journal Biosensors and Biomolecular Electronics, argues that the deepest obstacle is not chemical at all: it is the habit of reducing inherently multidimensional sensor signals to a single number.

The article, written by Wanda Cimmino, Alessandra Glovi, Antonella Miglione, Umberto Malapelle and Stefano Cinti, is a perspective rather than a report of new experiments, synthesizing representative case studies across electrochemical, optical and hybrid platforms. Its central claim is that chemometrics, the discipline of extracting chemically meaningful information from complex multivariate data, should be treated not as a post-processing accessory but as an enabling architectural component of next-generation biosensors. The authors introduce the concept of a sensor–model system, in which analytical performance emerges from the combined interaction between physical transduction and model-based information extraction, rather than being determined exclusively by molecular recognition or hardware sensitivity.

The authors identify five recurring mechanisms through which data-driven modeling enhances biosensing. First, multivariate calibration and pattern recognition, such as principal component analysis (PCA) and partial least squares (PLS) regression, can improve selectivity in complex matrices by resolving overlapping and non-specific responses. Second, modeling the full evolution of a signal rather than an isolated feature can compensate for drift and sensor-to-sensor variability. Third, multiplexed quantification can be extracted from composite responses without adding new receptors. Fourth, detection limits can be extended beyond apparent hardware constraints through multivariate optimization and full-signal exploitation. Fifth, data-driven models support smart, autonomous systems capable of real-time classification and decision-making.

The case for informational selectivity is illustrated by striking examples. Dip-type colorimetric paper-based sensor arrays for water discrimination do not rely on selective ion binding; instead they generate multidimensional color response patterns encoding each sample’s chemical fingerprint. When analyzed with PCA and linear discriminant analysis (LDA), weakly selective dye responses yielded highly accurate classification of bottled and tap water, with cross-validated sensitivities and specificities around 97 percent. A similar strategy using hydrophobic dye ink sensor arrays reliably differentiated complex multi-component perfume vapors composed of hundreds of constituents. In biological matrices, colorimetric sensor arrays based on nanozymes discriminated structurally similar monoamine neurotransmitters in human serum, achieving high classification accuracy even in the presence of serum interferents, despite the limited selectivity of each individual catalytic reaction.

Electrochemical platforms show the same principle. A voltammetric strategy developed by Yin and colleagues achieved simultaneous detection of dopamine, epinephrine, norepinephrine and serotonin in human serum, four monoamines whose strongly overlapping oxidation peaks defeat conventional peak-based analysis. By processing the full voltammetric profiles with Tchebichef moment feature extraction followed by multivariate regression, the workflow enabled accurate simultaneous quantification with nanomolar detection limits and satisfactory recovery values. In a related demonstration, a portable screen-printed electrochemical strip developed by Romanò and colleagues evaluated drug encapsulation in lipid nanovectors with accuracies between 88 and 95 percent compared with spectrophotometric methods, and PCA-LDA analysis of the voltammograms additionally discriminated liposomes built from different lipid matrices, expanding the device’s role from simple quantification toward quality control and formulation discrimination.

Drift compensation represents another frontier. In a chemometric-assisted cholesterol biosensor combining enzyme immobilization with molecularly imprinted polymers, adsorption phenomena and surface aging distorted peak intensity and position during differential pulse voltammetry. Rather than calibrating on individual peak currents, a three-way multivariate calibration applied to the complete voltammetric data matrices isolated stable analytical information from drift-influenced variations, improving reproducibility and long-term stability in serum samples. In food analysis, second-order data generated by systematically varying the scan rate, combined with baseline correction, alignment and Multivariate Curve Resolution-Alternating Least Squares calibration, allowed reliable quantification of gallic acid directly in red wine with a detection limit of 5.9 mg per liter, a root mean square error of prediction of 15.4 mg per liter and a relative error of prediction of 9.6 percent, despite overlapping ascorbic acid interference. For electronic noses, whose sensor arrays drift over weeks or years, a semi-supervised domain-adaptive convolutional neural network proposed by Jiang and colleagues quantified source-target similarity at multiple feature levels and improved performance in both long- and short-term drift benchmarks without requiring labels for the target domain.

Multiplexing without added chemical complexity emerges as a third theme. Pérez-Ràfols and colleagues modeled complete UV–Vis spectral profiles from 234 to 320 nanometers with PLS regression, enabling simultaneous quantification in two binary mixtures, pyrocatechol/isocotinamide and pyrocatechol/isoniazid, even under dynamic shaking conditions that typically amplify instability. He and colleagues combined constant-energy synchronous fluorescence, second-derivative processing and support vector regression to simultaneously quantify two benzimidazole pesticides, thiabendazole and fuberidazole, directly in red wine, with support vector regression performing best among the tested models. A multicolor gold nanoparticle colorimetric array developed by Huang and colleagues discriminated 14 metal ions with 100 percent accuracy using PCA-LDA and PLS-DA, while PLS regression quantified toxic ions including thallium, lead and mercury in real samples with correlation coefficients up to 0.9991 and low prediction errors, moving selectivity from the recognition level to the data level.

Detection limits themselves can improve through smarter experimental design rather than new hardware. Cimmino and colleagues applied a D-optimal design to simultaneously optimize manufacturing variables such as gold nanoparticle loading and probe concentration together with measurement conditions including ionic strength, hybridization time and square-wave voltammetry settings, achieving a fivefold improvement in the detection limit for miRNA-29c, from roughly 1 nanomolar to roughly 0.2 nanomolar, in both buffer and human serum, with only 30 experiments instead of the hundreds required for exhaustive univariate optimization. A disposable platform for mercury detection using a NiO-PSS carbon nanocomposite reached a detection limit of 0.021 micrograms per liter in real water matrices after Plackett–Burman screening and central composite response surface optimization. And a flexible, self-cleaning SERS platform coupled with deep learning quantified thiram residues directly on apples with a prediction correlation coefficient of about 0.996 and a detection limit of 0.020 milligrams per liter, while allowing sensor reuse over multiple runs. Beyond the laboratory, a portable electronic nose coupled with an artificial neural network discriminated lung cancer patients from healthy controls using global breath fingerprints, achieving accuracy around 80 percent with robust performance in blinded validation cohorts, and a machine-learning-assisted wearable gesture sensor based on graphene and carbon nanotubes translated high-dynamic strain signals into recognized sign language with accuracy exceeding 99 percent.

The authors are careful to stress that these benefits are conditional on methodological discipline. Overfitting is flagged as the most pervasive issue, particularly when high-dimensional signals are modeled with small datasets; nonlinear machine learning approaches such as artificial neural networks and convolutional networks require substantially larger datasets than data-efficient linear methods like PLS or PCA-LDA. Data leakage and inadequate validation design can artificially inflate performance when replicate measurements or temporally correlated samples are spread across training and validation sets, and apparently excellent metrics may fail dramatically on independent data, new sensor batches or unseen operating conditions. The authors call for independent test sets, batch-wise and temporal validation, validation across different sample matrices, transparent reporting of preprocessing pipelines and hyperparameters, and explicit attention to model transferability. They also note practical constraints for edge deployment, where lightweight architectures such as PLS and LDA suit low-power devices while heavier models may require compression or hybrid edge-cloud workflows. If chemometrics is developed with the same rigor applied to sensor fabrication, the authors conclude, biosensing is likely to shift from a hardware-dominated paradigm to an information-centric one, in which coupled sensor–model systems co-determine analytical performance and signal complexity becomes a strategic resource rather than a limitation.

Subject of Research: The role of chemometrics and machine learning in transforming biosensors into intelligent analytical systems

Article Title: Chemometrics reshapes the role of biosensors: from signal transducers to intelligent analytical systems

Article References: Cimmino, W., Glovi, A., Miglione, A., Malapelle, U., & Cinti, S. (2026). Chemometrics reshapes the role of biosensors: from signal transducers to intelligent analytical systems. Biosensors and Biomolecular Electronics, 1(1), Article 5. https://doi.org/10.1186/s44510-026-00005-4

Image Credits: AI Generated

DOI: 10.1186/s44510-026-00005-4

Keywords: biosensors, chemometrics, machine learning, multivariate calibration, electrochemical sensing, colorimetric sensor arrays, drift compensation, multiplexed detection, detection limit, electronic nose, point-of-care diagnostics, validation