Engineered nanozymes and explainable machine learning enable sensitive bacterial detection across complex conditions. The system uses three distinct signals and delivers transparent, verifiable results.
(Nanowerk Spotlight) The stakes of pathogen detection are higher than ever. A single contaminated food shipment or a missed diagnosis in a clinical setting can trigger outbreaks with global consequences. Yet the tools used to identify bacterial threats like Salmonella typhimurium remain hindered by trade-offs that have proven stubbornly difficult to overcome.
Systems that are sensitive tend to be too complex for widespread use. Simpler methods, in turn, often lack the accuracy needed in noisy, real-world environments. The result is a persistent gap between what biosensors promise in the lab and what they deliver at scale.
Efforts to bridge this gap have increasingly turned to nanozymes—synthetic nanomaterials designed to mimic the behavior of natural enzymes. These materials offer durability and tunability, but suffer from limited catalytic performance, especially when built from noble metals like platinum.
At the same time, multi-signal biosensing platforms—those that combine color, light, and heat outputs—are gaining attention for their ability to cross-validate signals and filter out noise. But as signal complexity increases, so does the challenge of interpretation. Many systems rely on opaque machine learning models that make accurate predictions but offer little insight into how those predictions are formed.
These two bottlenecks—catalytic inefficiency and algorithmic opacity—have largely developed in parallel. But a new approach suggests they might be solvable together. By treating materials design and signal interpretation as interdependent problems, researchers are now beginning to rethink how biosensors are built. This shift opens the door to platforms that don’t just detect pathogens more effectively, but also explain how and why they did so—an essential step if such tools are to earn trust in clinical, food safety, or public health contexts.
Their platform integrates a chemically reengineered platinum-based nanozyme with a machine learning algorithm that not only improves detection performance but also provides clear explanations for its decisions. The system is designed to detect S. typhimurium using a combination of optical and thermal outputs, coupled with interpretable data analysis. The work represents a methodical advance in how biosensors can be constructed for both high accuracy and practical reliability.
A,B) The development of AFRNBs@PtNPs nanozymes with the favored electron transfer pathway, C) XGBoost learning algorithm-assisted nanozyme-based multisignal biosensors, and D) the SHAP interpretable frame analysis-driven feature optimization in biosensing. (Image: Reprinted from DOI:10.1002/advs.202505712, CC BY) (click on image to enlarge)
At the core of the sensing system is a redesigned platinum nanoparticle catalyst. The team began by addressing the electronic limitations of traditional Pt-based nanozymes. In catalytic systems, a property known as the d-band center helps determine how easily a metal can interact with molecules involved in a reaction—in this case, hydrogen peroxide (H₂O₂), which is commonly used in biosensing assays. If the d-band center sits too far below the material’s Fermi level, the interaction with H₂O₂ becomes inefficient.
To correct this, the researchers anchored platinum nanoparticles to specially fabricated nanobowls made from aminophenol-formaldehyde resin (AFRNBs). This structure facilitates a directional flow of electrons from nitrogen atoms in the resin, through the platinum, and toward oxygen—described as an N→Pt→O transfer pathway.
This interfacial electron transfer was confirmed using spectroscopy and density functional theory calculations. The result was a measurable shift in the d-band center of platinum toward the Fermi level, enhancing its ability to adsorb and activate H₂O₂. Compared to unmodified PtNPs, the new nanozyme achieved a 3.4-fold increase in peroxidase-like activity.
Kinetic tests showed it had improved substrate affinity and lower activation energy. Electron spin resonance data indicated a higher generation rate of hydroxyl radicals, a key reactive intermediate. These enhancements allow the catalytic reaction to proceed more efficiently, and the engineered structure remained stable for at least 30 days at room temperature.
To translate this improved catalysis into readable signals, the researchers turned to a compound known as TPEN, a type of aggregation-induced emission luminogen (AIEgen). TPEN is notable for emitting fluorescence when unoxidized, while its oxidized form produces a visible color and heat. This allowed the team to construct a triple-output system: colorimetric (absorbance at 600 nm), fluorescent (emission at 525 nm), and photothermal (measured by heat conversion).
The different signal modes arise from distinct physical states of the molecule, which respond differently to the presence and concentration of the pathogen. Importantly, the reactions reached completion within 15 minutes, and the photothermal efficiency of the oxidized form reached nearly 40 percent—sufficient for use in thermal readouts.
With this system in place, the team designed a biosensor for detecting S. typhimurium based on a classical sandwich immunoassay format. As the concentration of the target pathogen increased, the colorimetric and photothermal signals rose, while fluorescence decreased, due to TPEN being consumed in the oxidation process. This inverse correlation among signal types provided a built-in mechanism for cross-validation.
Tests showed the system could detect bacterial concentrations as low as 204 CFU/mL, outperforming conventional enzyme-linked immunosorbent assays (ELISAs), which had a detection threshold of around 1600 CFU/mL. The system also showed high specificity when exposed to a range of other common bacterial species and maintained consistent performance in spiked milk samples, with recoveries ranging from 82 to 118 percent.
To manage the complexity of interpreting these three simultaneous signals, the researchers implemented the XGBoost machine learning algorithm. This model builds a series of decision trees that progressively refine their predictions based on error correction. When applied to the biosensor data, it achieved a 95.8 percent overall accuracy and correctly identified all positive samples above 100 CFU/mL. However, unlike most machine learning applications in biosensing, this system did not stop at prediction.
Using SHAP (SHapley Additive exPlanations), a framework grounded in cooperative game theory, the team was able to quantify the contribution of each signal mode to each classification decision. This analysis revealed a clear pattern: at low pathogen concentrations, the colorimetric signal had the most influence on detection; at medium levels, the photothermal response played a larger role; and at high concentrations, fluorescence became the dominant factor.
These findings aligned with the physical properties of the TPEN-based signal generation process and helped verify the internal logic of the algorithm. The use of SHAP provided a layer of interpretability rarely seen in biosensor data analysis, offering a path to more transparent diagnostics.
The study demonstrates that by modifying the electronic properties of nanozymes and combining them with explainable machine learning, it is possible to build biosensing systems that are both analytically powerful and operationally transparent. Each part of the system—from catalyst to signal readout to algorithm—was designed with the others in mind. This coordinated design approach allowed the team to overcome limitations that have held back the field, including inefficient catalysis, signal interference, and unexplainable model behavior.
Rather than separating materials development from computational analysis, the researchers treated them as components of a unified system. This integration produced a sensor that not only detects pathogens with high sensitivity but also provides evidence for how each detection decision was made.
As biosensing technology moves into more complex environments, such as point-of-care diagnostics or field-based monitoring, this type of transparent, cross-validated platform offers a model for how performance and trust can be engineered together.
If this article was useful, support our independent nanotechnology reporting with any amount.
Your contribution funds the next explainer and keeps Nanowerk open for everyone.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
