Key Takeaways
- A two‑step sample‑preparation workflow (ultrafiltration → acetone precipitation) concentrates and cleans viral particles, improving MALDI‑TOF ionization and enabling culture‑free analysis of low‑biomass samples.
- Ten machine‑learning and deep‑learning models were tested; the Extra Trees Classifier, Support Vector Classifier, and a 1‑D CNN achieved near‑perfect classification on the internal dataset, correctly separating bacteria from viruses and identifying Gram type.
- When validated against an external RKI MALDI‑TOF database that used different inactivation (TFA) and extraction protocols, model performance dropped, highlighting over‑fitting and the impact of inter‑laboratory variability.
- Persistent confusion between Bacillus cereus and B. subtilis spectra stemmed from limited intra‑genus diversity in the training set.
- The study demonstrates feasibility of MALDI‑TOF + AI for rapid, agent‑agnostic pathogen identification but stresses the need for larger, heterogeneous datasets and robust preprocessing to ensure real‑world reliability.
Viral Sample Preparation Strategy
The authors designed a workflow that first enriches viruses using 10 kDa polysulfone ultrafiltration membranes, a step shown previously to retain > 4–5 log₁₀ of viral particles (> 99.9 %). “Our choice of ultrafiltration as the first step was based on our previously published work demonstrating that polyethersulfone membranes (PES) with a 30 kDa cutoff achieved > 4–5 log₁₀ viral retention (> 99.99%) for MS2 and AcNPV.” The retained particles are then subjected to acetone precipitation, which removes salts, lipids and detergents while concentrating viral proteins. This combined approach is culture‑independent, works with low‑biomass clinical material, and yields cleaner spectra than direct‑cell or chemical‑lysis methods that often introduce host‑derived contaminants or interfere with ionization.
Advantages Over Conventional MALDI‑TOF Prep
Compared with chemical lysis, solvent‑based extraction, or direct analysis of infected cells, the ultrafiltration + acetone protocol provides a pre‑analytical concentration step followed by organic‑solvent cleanup. The authors note that it “offers several advantages compared to previously reported methods. It is culture‑independent providing a pre‑analytical concentration of viral particles present in the sample followed by sample clean up through organic solvent precipitation, which is particularly advantageous for analyzing low‑biomass clinical samples.” By reducing matrix effects and nonspecific background, the workflow improves ionization efficiency and produces more reproducible spectral peaks essential for downstream AI classification.
Internal Dataset and Model Training
Using this preparation, the team generated spectra for seven bacterial strains and five viruses (AAV2, MMLV, HIV‑1‑derived lentivirus, AcMNPV, bacteriophage MS2). After standard preprocessing (baseline subtraction, Savitzky‑Golay smoothing, normalization, and binning to 5 000 features), they trained ten classifiers—eight traditional ML models (Random Forest, Linear SVC, Ridge Classifier, kNN, Extra Trees, SVM, Logistic Regression, XGBoost) and two DL baselines (1‑D CNN and a denoising autoencoder + Ridge). Five‑fold stratified cross‑validation was employed to guard against over‑fitting given the modest number of spectra per class.
Classification Performance on Internal Data
Several models achieved perfect or near‑perfect scores. “Seven models consistently delivered perfect classification results for distinguishing bacteria from viruses, identifying Gram type, and reliably classifying the panel of 12 bacterial and viral samples studied into their respective species.” Notably, the Extra Trees Classifier, Support Vector Classifier, and 1‑D CNN produced the highest average accuracy and F1‑score, correctly separating bacterial from viral spectra and discriminating Gram‑positive versus Gram‑negative organisms. The ability to differentiate closely related Bacillus species was also evident, although some confusion persisted (see later).
External Validation and Performance Drop
To test generalization, the authors challenged the top three models with spectra from the Robert Koch Institute (RKI) MALDI‑TOF database, which employed trifluoroacetic acid (TFA) for inactivation and different extraction procedures. Performance declined, especially for the SVM. As the paper states, “A major and a common source of misclassification was the consistent confusion of B. cereus and B. subtilis spectra.” The decline was attributed to protocol heterogeneity: “More specifically, trifluoroacetic acid (TFA) was used to inactivate bacteria and prepare samples of the RKI external database, whereas we used direct transfer (DT), extended direct transfer (eDT), and standard protein extraction (PE) recommended by Bruker for producing spectra of our internal dataset.” Variations in instrument calibration and sample handling further increased inter‑laboratory variability, causing the models to over‑fit to the internal spectral signatures.
Sources of Misclassification
The persistent B. cereus/B. subtilis mix‑up arose because the internal dataset contained only three Bacillus classes with a limited number of spectra each, making it difficult for the models to learn subtle genus‑level variations. When confronted with the broader RKI collection, the models misassigned spectra from these closely related species. The authors note that “the persistent misclassification between B. cereus and B. subtilis can be attributed to the fact that the internal dataset comprised three classes belonging to the Bacillus genus, which, in combination with an adequate but not extensive number of spectra for model training, reasonably leads to a non‑negligible rate of misclassification during the evaluation with the external database among species within this genus.”
Limitations and Real‑World Relevance
The study acknowledges several constraints. Spectra were obtained from reference strains under controlled conditions, not from complex matrices such as blood, urine, or environmental samples that can introduce ionization‑suppressing contaminants. Moreover, the total dataset comprised only 255 spectra, raising concerns about over‑fitting despite cross‑validation. The authors advise that “addressing these limitations will require future studies focused on the acquisition of MALDI‑TOF spectra from clinical and environmental samples, the use of larger and more heterogeneous datasets, and the systematic evaluation of preprocessing strategies designed to mitigate matrix effects for bacterial and viral pathogens.”
Future Directions and Broader Impact
Despite the limitations, the work illustrates a promising path toward rapid, agent‑agnostic pathogen identification. Expanding the viral panel, incorporating aerosol samples, and refining deep‑learning architectures (e.g., attention‑based 1‑D CNNs) could improve discrimination among strains and enhance robustness. Advances in MALDI matrices that improve analyte‑matrix stoichiometry, combined with growing open spectral libraries, lay the groundwork for deploying AI‑augmented MALDI‑TOF in diagnostics, biodefense, and outbreak management—especially when rapid decision‑making is critical.
Conclusion
By integrating a stringent ultrafiltration‑acetone preparation pipeline with a suite of machine‑learning and deep‑learning classifiers, the researchers demonstrated that MALDI‑TOF MS can reliably differentiate bacteria from viruses and classify specific agents in a controlled setting. The drop in performance when tested against externally generated spectra underscores the necessity of training on diverse, protocol‑variant data to achieve real‑world applicability. Continued efforts to enlarge and diversify spectral repositories, refine preprocessing, and develop more generalizable AI models will be essential to translate this platform from bench‑proof to routine clinical and field use.
https://www.nature.com/articles/s41598-026-54426-y