The generalizability of the model is often diminished as a result of the limited sample data size (Weis et al., 2020). The true prevalence of drug resistance can be more comprehensively captured by large-scale datasets. Consequently, extensive datasets are often necessitated by state-of-the-art DL architectures to achieve optimal performance. However, particular attention is warranted by the phenomenon of high-dimensional datasets (Nguyen et al., 2024). It is noteworthy that it has been demonstrated by Weis et al. that a significant impact on the performance of predictive models is exerted by the temporal origin of the samples, specifically the year of collection. The predictive accuracy of the models can be markedly improved by employing spectral data derived from samples collected in more recent years (Weis et al., 2022). High-quality MALDI-TOF MS data needs to be utilized to optimize model training.

Pure bacterial cultures grown on solid media are typically employed in MALDI-TOF MS analysis. The phenotypic characteristics of bacteria under different culture conditions can be significantly influenced by the growth temperature. Degraded mass spectra due to bacterial lysis may be caused by frozen and lyophilized cultures, whereas clearer and more reproducible mass spectral results are generally obtained from fresh cultures (Topić Popović et al., 2023). It was demonstrated by GATO et al. that the expression of peptides and proteins during bacterial growth is significantly impacted by the nutritional composition of different culture media. In their study, the prediction outcomes for carbapenemase - producing Klebsiella pneumoniae were notably influenced by this variation (Gato et al., 2023). The accuracy and reproducibility of experimental results are underscored by the essential role of consistent culture conditions. Additionally, a variety of pre-analytical processing methods are available for cultures. In the direct deposition method, colonies are simply spotted onto the target plate and then covered with the matrix. This method is characterized by low labor intensity and rapid operation. If an additional protein extraction step is implemented, it would involve bacterial lysis, followed by centrifugation, washing, and concentration. While a more purified protein profile is achieved by this approach, thereby enhancing data quality, its use in clinical settings is typically precluded by its labor-intensive, time-consuming, and costly nature. Discrepancies in peak detection may be led to by variations in pre-analytical processing methods. In the Escherichia coli gentamicin model examined by Maureen Feucherolles et al., a lower model performance, with an Area Under The Precision-Recall Curve (AUPRC) of 0.23, was resulted in by the ethanol/acetonitrile extraction method (EtOH/ACN). In contrast, a higher performance, with an AUPRC of 0.92, was achieved by the direct deposition method (Feucherolles et al., 2021). Consequently, optimal results are more likely to be achieved when mass spectra are acquired under standardized culture conditions and sample preparation protocols. Drug resistance was not directly assessed by using antimicrobial agents by Margot Delavy et al., but was rather indirectly evaluated by inducing differential protein expression. In Candida albicans, the calcineurin signaling pathway, which is responsible for mediating the fungal stress response and regulating drug resistance, may have distinct activities exhibited between fluconazole-resistant and susceptible strains. The normal function of this pathway can be disrupted by cyclosporin A, a calcineurin inhibitor. Consequently, distinct protein expression profiles between resistant and susceptible strains can emerge following treatment with cyclosporin A (LaFayette et al., 2010). Without directly employing fluconazole, differences in protein expression between fluconazole-resistant and susceptible strains following treatment with cyclosporin A were detected by Margot Delavy et al. using MALDI-TOF MS. By integrating ML algorithms for rapid classification, rapid (4h) and accurate prediction of fluconazole resistance was successfully achieved (Delavy et al., 2019).

Owing to their limited resolution, proteins or molecules that fall outside the conventional mass - to - charge ratio (m/z) range of 2000–20000 cannot be accurately identified by entry - level MALDI - TOF MS instruments. Certain AMR - associated proteins with low expression levels may be obscured by background noise in the conventional spectra, thereby impeding their effective detection (Yoon and Jeong, 2021).

The intricacy of algorithmic processing can be exacerbated, the precision of predictive modeling can be diminished, and the duration required for model training can be significantly extended by the considerable noise and extraneous information contained within raw datasets. Before ML techniques are employed for the analysis of MALDI-TOF MS data, data preprocessing must be conducted to enhance data quality and analyzability. Intensity variance stabilization, smoothing, baseline removal, total ion current normalization, mass spectral trimming, and the association with corresponding resistance labels are included in the preprocessing steps. The quality of peaks (features) utilized in ML models is influenced by the distinct preprocessing methods employed (Chung et al., 2023a). Moreover, the selection of preprocessing methods is determined by the underlying data structure and the architecture of the anticipated model.

Supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning are typically encompassed by ML algorithms Figure 1 (Goodswen et al., 2021). A model is trained on a training dataset that includes input features and corresponding target labels by supervised learning. The relationship between these features and labels is learned by the model, which enables it to generalize and predict the output for unseen data. Various learning tasks, including classification and regression, are widely applied in supervised learning (Jiang et al., 2020; Wang et al., 2023). Prior knowledge is not relied upon in the operation of unsupervised learning. The intrinsic structure, patterns, and distributional relationships within the data are autonomously uncovered by the algorithm through processing unlabeled data. This approach is well-suited to learning tasks such as clustering and dimensionality reduction (Bröker et al., 2024). A small amount of labeled data along with a large quantity of unlabeled data is employed by semi-supervised learning. The model’s performance is enhanced by utilizing the structural information within the unlabeled data. In contrast, pre-labeled data is not required by reinforcement learning. Feedback from the environment, in the form of rewards or penalties, is relied upon to guide the learning process. The objective is for an optimal policy for acting within the environment to be learned in order to maximize cumulative rewards (Rani et al., 2023).

The integration of MALDI-TOF MS with ML for AMR prediction is fundamentally based on supervised learning methodologies (Jiang et al., 2020). In the preprocessing stage, such as in feature selection and dimensionality reduction, unsupervised learning is often employed. When training data are limited, simple algorithms with regularization, such as linear regression and logistic regression with penalty terms, are usually chosen. The training set may have limited fitting capacity by these algorithms, resulting in higher error rates (high bias), but better generalization ability (low variance) is typically exhibited. When training data are abundant and complex feature interactions may be involved by the underlying algorithms, instance-based methods (e.g., k-nearest neighbors) or tree-based algorithms (e.g., stochastic gradient boosting or random forests) may be performed well, characterized by low bias and high variance (Deo, 2015).

The intrinsic complex nonlinear relationships inherent in MALDI-TOF MS data have been proven to be effectively addressed by the gradient boosting technique utilized by the CatBoost algorithm. As reported by López-Cortés et al., in the prediction of ciprofloxacin-resistant Escherichia coli, an Area Under The Receiver Operating Characteristic Curve (AUROC) and an AUPRC of 0.91 were achieved by the algorithm, with balanced accuracy and F1 score values of 0.81 and 0.78, respectively. Similarly, in the prediction of oxacillin-resistant Staphylococcus aureus, an AUROC of 0.86, an AUPRC of 0.73, balanced accuracy of 0.77, and an F1 score of 0.61 were demonstrated by the algorithm (López-Cortés et al., 2025). Models with strong generalizability and adaptability may be trained in the future, thereby enhancing the capability for detecting AMR Table 1.

Hyperparameters are parameters that should be specified prior to model training. The optimization objective is to have the optimal combination of hyperparameters identified to maximize model performance on the validation set (Greener et al., 2022). Typically, after the hyperparameter grid is constructed, cross-validation is conducted to assess the performance of each hyperparameter configuration. The optimal hyperparameter combination is selected based on the cross-validation performance. Subsequently, the model is retrained on the entire training dataset using the selected optimal hyperparameter combination (Bzdok et al., 2018; Pfob et al., 2022).

The annotated mass spectrometry features are fitted to the “drug-resistant/susceptible” labels in the process of model training. The underlying patterns and regularities in the data are learned thereby, and outcomes for unseen data are enabled to be accurately predicted by the model. In the study by Nguyen HA et al., five base models-Logistic Regression, Random Forest, Support Vector Machine, LightGBM, and Multilayer Perceptron-were applied with different feature extraction methods (e.g., dynamic binning and latent embedding) to develop models (Nguyen et al., 2024).

The effectiveness and robustness of ML models are what the clinical applicability is substantially contingent upon. Consequently, a pivotal role is assumed by model validation and evaluation. During the model training phase, model validation is predominantly executed, and following model deployment, it can also serve as a continuous monitoring and validation mechanism. The primary objectives of model validation include the optimization of model hyperparameters being facilitated, overfitting or underfitting being averted, the model’s generalization capability being assessed, and robust performance on unseen data being ensured. Model validation is typically implemented using the validation set approach or cross-validation techniques. Validation accuracy, validation loss, learning curves, and cross-validation metrics are commonly employed as validation metrics.

The model’s overall performance is provided with a comprehensive and systematic assessment after the completion of model training through model evaluation, and its suitability for practical application is thereby determined. The test set approach, Leave-One-Out Cross-Validation (LOOCV), and cross-validation are commonly used as model evaluation methods. In addition to these methods, a variety of assessment metrics are involved in model evaluation, including sensitivity, specificity, accuracy, positive predictive value, negative predictive value, AUROC, AUPRC, and F1 score. Among these metrics, interpretable summaries of the model’s overall performance are offered by AUROC, AUPRC, and F1 score. Although still somewhat affected by class imbalance, these metrics’ ability to recognize minority classes is more accurately evaluated than that of traditional metrics, and thus the model’s performance on imbalanced data is better reflected by them. In cases of class imbalance, it is recommended to use a combination of multiple metrics to comprehensively evaluate model performance. Based on these evaluation metrics, continuous optimization of the model can be achieved, and model performance can be progressively enhanced through iterative refinement.

The aforementioned cross-validation can be involved in various stages within the process (Bradshaw et al., 2023). In the context of feature selection, the influence of different feature subsets on model performance can be evaluated by cross-validation, thereby allowing the most informative features to be selected. The comparison of various algorithms on the same dataset can also be facilitated by cross-validation, which allows for the identification of the algorithm best suited for a specific problem and dataset. Furthermore, model performance across different hyperparameter settings can be assessed by cross-validation, and the optimal hyperparameter combination can be identified thereby. During model evaluation, a more stable and reliable performance estimate can be provided by cross-validation through assessing the model on multiple training/testing splits (Bradshaw et al., 2023). K-fold cross-validation, LOOCV, and nested cross-validation are encompassed by common cross-validation techniques.

The decisions and output results of a model can be clearly understood and explained, which is determined by the model’s interpretability (Allgaier et al., 2023). SHAP are derived from the concept of Shapley Value in game theory. The SHAP value for each feature can be calculated, and the contribution of each feature to the model’s prediction can be quantified, with both local and global interpretations being provided. A variety of visualization tools, such as Bar Plots and Bee Swarm Plots, are offered by SHAP to display the SHAP values of features, and the transparency and interpretability of ML models are thereby enhanced. The UniProt database is commonly used to identify proteins that are directly or indirectly related to AMR mechanisms. If certain peaks are not matched to known proteins, this suggests that unexplored MALDI-TOF MS data is indicated (Table 2).

The dissemination pathways of resistance genes can be tracked by MALDI-TOF MS through the detection of proteins that are encoded by these genes. Bacterial resistance can be caused by spontaneous mutations within their own genomes and by resistance genes that are acquired from other bacteria via horizontal gene transfer. For instance, the PSM-mec peptide, which is encoded by the mecA gene-a key determinant of oxacillin resistance in staphylococci, can be detected by MALDI-TOF MS as utilized by Weis et al (Weis et al., 2022). The PSM-mec peptides are detected in diverse bacterial species, which suggests that the mecA gene may be harbored by these bacteria. The mecA gene could have been acquired through horizontal gene transfer mechanisms, such as plasmid dissemination. Indirect inferences regarding the dissemination of antibiotic resistance genes can be made based on this observation. While a ciprofloxacin resistance model was being investigated, a protein associated with the characteristic bin 2,025 was identified by researchers. Although this protein was not annotated as a protein of Staphylococcus epidermidis in the UniProt database, it was identified as a member of the pathogenicity island family in Staphylococcus aureus at a molecular weight of 8,077 Da (Ren et al., 2024). The capacity of MALDI-TOF MS to identify pertinent biomarkers is underscored by this.

These biomarkers are frequently associated with known resistance mechanisms, either directly or indirectly. However, considering the complex nature of protein functions and interactions, the proteins that confer resistance to microorganisms may not necessarily be the identified biomarkers. Rather, they may be encoded by the same gene or gene cluster that encodes the proteins responsible for resistance (Feucherolles et al., 2021). The accurate identification of resistance-associated proteins can be complicated by overlapping mass spectral peaks, which can be caused by the highly similar molecular weights of distinct proteins in MALDI-TOF MS data (Nguyen et al., 2024). Nonetheless, the precise identification of resistance-associated proteins or genes can be achieved by employing molecular docking simulations (Yu et al., 2022)or by integrating genomic sequencing data (Jiang et al., 2019). Additionally, molecular weights that exceed the conventional detection range of MALDI-TOF MS (approximately 2,000 - 20,000 Da) may be possessed by certain resistance-determining proteins. In such instances, surrogate biomarkers may be detected to serve as an indirect means for resistance to be inferred. For example, in Staphylococcus aureus, the molecular weight of the mecA protein, which confers resistance to methicillin, is beyond the conventional detection range. However, its smaller product, PSM-mec (with a molecular weight of 2,415 Da), remains within the detectable range and can be utilized to identify strains harboring the mecA gene (Rhoads et al., 2016). To ensure the veracity and robustness of predictive interpretations, the biological functions of identified features should be prudently elucidated.

Significant challenges in infrastructure, finance, and training may be posed by the integration of MALDI-TOF MS and ML into clinical practice. To justify the investment, it is essential to conduct rigorous cost - effectiveness studies. Additionally, the current insufficiency in the availability of open - source data and code can be addressed by establishing large - scale, publicly available datasets and analytical codes. Model validation and optimization by other researchers would be facilitated by this approach, thereby enhancing the reproducibility of studies (Weis et al., 2020). MALDI-TOF MS data of AMR are highly geographically specific. Pretrained models, which are trained on other datasets, cannot be directly applied and need to be adjusted according to local data. Moreover, the resolution, peak intensity, and noise of mass spectra may be varied by different instruments. Pretrained models need to be fine - tuned for a specific setting to be effective (López-Cortés et al., 2025). Therefore, the standardization of data and processes should be focused on by future research to enable transfer learning across different regions and instruments. The development of models that are adaptable to diverse clinical environments and the improvement of the accuracy of AMR prediction will be facilitated by this. Given that AMR is dynamically evolving, it is essential that continuous data monitoring and dynamic model updating be carried out to ensure the validity and clinical value of these models. A multimodal feature repository can be established by integrating proteomic fingerprints with patient-derived genomic profiles and clinical metadata, and its online, continuously updated framework can be advanced in parallel with privacy-compliant information sharing. The feasibility and effectiveness of using MALDI-TOF MS integrated with ML for AMR prediction can be further enhanced by future research through expanding the range of species studied, extending MALDI-TOF MS datasets, integrating multimodal data, and consolidating clinical information (Peiffer-Smadja et al., 2020).