Mass Spectrometry
Online ISSN : 2186-5116
Print ISSN : 2187-137X
ISSN-L : 2186-5116
Review
Recent Applications of Artificial Intelligence and Related Technical Challenges in MALDI MS and MALDI-MSI: A Mini Review
Ali FarhanYi-Sheng Wang
Author information
JOURNAL OPEN ACCESS FULL-TEXT HTML

2025 Volume 14 Issue 1 Pages A0175

Details
Abstract

Artificial intelligence (AI) has provided viable methods for retrieving, organizing, and analyzing mass spectrometry (MS) data in various applications. However, several challenges remain as this technique is still in its early, preliminary stages. Critical limitations include the need for more effective methods for identification, quantification, and interpretation to ensure rapid and accurate results. Recently, high-throughput MS data have been leveraged to advance machine learning (ML) techniques, particularly in matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) MS and MS imaging (MSI). The accuracy of AI models is intricately linked to the sampling techniques used in MALDI and MALDI imaging measurements. With the help of artificial neural networks, traditional barriers are being overcome, accelerating data acquisition for different applications. AI-driven analysis of chemical specificity and spatial mapping in two-dimensional datasets has gained significant attention, highlighting its potential impact. This review focuses on recent AI applications, particularly supervised ML in MALDI-TOF MS and MALDI-MSI data analysis. Additionally, this review provides an overview of sample preparation methods and sampling techniques essential for ensuring high-quality data in deep learning-based models.

1. INTRODUCTION

In recent years, artificial intelligence (AI) has been increasingly applied to various mass spectrometry (MS) applications. AI-based techniques are particularly valuable for analyzing large and complex datasets, such as those generated in matrix-assisted laser desorption/ionization (MALDI) MS and MS imaging (MSI). The combination of AI and MS has demonstrated tremendous capabilities for various applications. Figure 1 illustrates the rapid rise in AI-driven MALDI-MS studies, emphasizing the field’s growing reliance on AI techniques over the last decade. Moreover, this is prompted by the fact that AI algorithms can learn parameters, perform automated identification, analyze data, and convert input features into predictive insights, making them especially useful in biomedical and clinical research.13) Furthermore, AI techniques for clinical data analysis can detect lethal diseases more accurately compared to traditional diagnostic methods in clinical examinations. As AI continues to evolve, understanding its progress and current status in MS is crucial for advancing research, improving analytical techniques, and unlocking new possibilities in biomedical applications.

Fig. 1. Number of publications employing ML and DL techniques for MALDI and MSI data analysis from 1997 to 2024. Data were retrieved from the PubMed database in February 2025. DL, deep learning; MALDI, matrix-assisted laser desorption/ionization; ML, machine learning; MSI, mass spectrometry imaging.

In the context of biomedical analysis, AI techniques are tailored to handle or interpret high-dimensional spectral data from analytical techniques like MS and spectroscopy. Figure 2 illustrates the application of MS and spectroscopy in diagnosing abnormal cell proliferation in cancer, demonstrating the advantages of AI-driven biomarker analysis.3) Among the AI approaches, machine learning (ML) models are widely used for analyzing complex biological data and making predictive decisions. ML is a subset of AI that uses algorithms to learn patterns and make predictions based on MS data. Analysts apply ML to tasks such as feature extraction, noise reduction, compound classification, and quantification to automate and enhance the accuracy in complex analyses. Since histopathological examination of tissue samples by manual inspection often presents a risk of misinterpretation, there is a growing need for advanced ML algorithms that can automate labor-intensive processes traditionally dependent on human expertise. ML models are generally categorized into four main types based on their ability to solve complex problems. They include clustering algorithms,4) dimensionality reduction techniques,5) regression analysis,6) and classification algorithms.7) These methods have been widely applied in clinical and experimental data analysis, particularly in biomedical, biomolecular, proteomics,8) and other biological research fields. In MS applications, ML plays an essential role in identifying intricate patterns within large datasets, improving accuracy in MS data interpretation.9) Recently, advanced unsupervised ML methods have demonstrated the drawbacks of manual peak selection in assessing drug distribution.10)

Fig. 2. Analysis of complex signals in in vitro diagnosis can be effectively addressed using clustering and regression algorithms powered by AI. The AI-driven techniques facilitate automatic processing and analysis of big data acquired using mass spectrometry and optical spectroscopic methods. Reprinted by permission from Chen et al.3) AI, artificial intelligence.

Similar to ML algorithms, deep learning (DL) models are also used for training, testing, and evaluation on a given dataset. However, DL has a remarkable advantage over ML algorithms when substantially large datasets are used for prediction. DL is a subset of ML that leverages deep neural networks to extract more complex relationships from multidimensional MS data, especially large datasets. The hierarchy of AI, ML, and DL forms an effective system to handle different aspects of experimental datasets, and together they transform MS by enhancing data interpretation, speeding up analysis, and uncovering insights that might be missed using traditional methods. In the context of MALDI-time-of-flight (TOF) datasets, most studies show that DL algorithms deliver more comprehensive results in peak analysis due to their architecture.11) DL is particularly valuable when dealing with large datasets where traditional ML methods might be insufficient to capture intricate, non-linear relationships. The conceptual differences between ML and DL approaches are given as follows1214):

  • •  DL is an extension of ML in which enhanced or multiple layers of artificial neural networks are designed to extract hierarchical features from the data.
  • •  Typical ML algorithms include a variety of methods such as linear regression, decision trees, support vector machines, and k-nearest neighbors.
  • •  DL models are primarily designed using convolutional neural network (CNN) layers, as well as recurrent neural networks (RNNs) and transformers for natural language processing.
  • •  In ML, feature selection depends on manual identification in relevant data, while in DL, CNNs layers automatically learn to extract features from raw data with less manual incorporation.
  • •  ML requires manually guided fine-tuning method for optimization, whereas DL models incorporate and fine-tuning based on inherent data structure.

ML models extract and transfer information through multiple layers, including the input, hidden, and output layers. In contrast, while DL models retain the basic hierarchical structure of ML, they incorporate additional components such as convolutional layers, pooling layers, and fully connected layers, enabling more advanced feature extraction and pattern recognition. Key steps involved for ML and DL include data collection, data preprocessing (cleaning, splitting), selection of model, training of datasets, evaluation of output metrics, and optimization when outcomes are unsatisfactory or fail to achieve high accuracy on blind data. The application of ML and DL algorithms varies based on the type of problem, such as classification, segmentation, or object tracking. The method for handling MALDI-TOF and MSI datasets follows the same concept. Take the ML method as an example, the process typically begins with preprocessing, where raw spectra are normalized, the noise is removed, and the key features are extracted to prepare the data for ML models. Following this, model training is performed, where supervised ML models learn patterns from labeled datasets. While training can occur on standard CPUs, graphics processing units significantly accelerate execution time,15) particularly for DL applications. Classification in supervised ML relies on well-labeled training data, and testing/validation subsequently ensures that the model generalizes well to new data.

This review highlights key aspects of MALDI-MS and MSI that are critical for AI using ML and DL applications. Recent advancements in the use of AI techniques in biomedical studies are also included. Since MS data quality is a key determinant of AI algorithm effectiveness, proper sample preparation to obtain suitable data quality is also an essential factor to be considered. Even minor errors in sample handling can lead to significant variations in detected analyte abundance.16) For that reason, important principles of the sample preparation process are also discussed. Due to the rapidly evolving nature of this field, this review does not aim to provide an exhaustive discussion of all related studies. Instead, it serves as a brief introduction to the field and a collection of works for researchers who are interested in AI-based MALDI-MS analysis.

2. MALDI AND MALDI-MSI TECHNIQUES IN BIOCHEMICAL AND CLINICAL APPLICATIONS

Advancements in MALDI and MALDI-MSI applications have been driven by improvements in instrumentation, software design, computational power, and data analytics. MALDI-MS is a popular method for identifying proteins, molecular markers, and microorganisms because of its sensitivity, speed, and efficiency in compared to other techniques. MALDI was first developed in the 1980s for ionizing large biological molecules using a pulsed laser beam.17) While Tanaka first demonstrated the concept of matrix in the soft laser desorption/ionization process,18,19) Karas and Hillenkamp further advanced the field by using organic matrices.20) The original MALDI concept involved depositing sample mixtures containing matrix materials onto a surface to form co-crystals, where the matrix absorbs laser energy and facilitates the desorption and ionization of biological samples, in which the ions were detected using a TOF mass spectrometer.18) The complex ionization process has been extensively described in the literatures.2129) Due to its efficiency in ionizing biological samples, MALDI quickly became the method of choice in MS-based proteomics research.3032)

Since the 1990s, Caprioli et al.33) and Spengler and Hubert34) have made significant contributions to MALDI-MS by introducing imaging technology. MALDI-MSI35) is now an established tool for visualizing biomolecules at the tissue level. Unlike conventional MS approaches that require homogenization, extraction, and pre-separation of molecules, MSI involves acquiring mass spectra across a tissue sample in a grid pattern. These spectra are then displayed as intensity maps for each mass-to-charge (m/z) feature, preserving their spatial information. Spatial features can also be identified by selecting specific fragment ions that represent components of the sample. Fragment ions, produced when a molecular ion disintegrates in the ion source, provide insights into the chemical composition of a compound. Currently, MSI is widely used to analyze the spatial distribution of peptides,36) lipids,37) glycans,38) and metabolites39) in various fields, including clinical,40) plant science,41) food science,42) and pharmaceutical research.43) An interesting emerging topic is the application of MALDI-MSI in hair analysis to detect the history of drug abuse.44)

Nowadays, MALDI-MS and MALDI-MSI remain reliable methods for identifying microorganisms and protein interactions, as well as for understanding different biological processes, including sublethal membrane molecules and various biomolecular pathways.45,46) Since TOF mass spectrometers offer superior speed, resolution, and accuracy, as well as good sensitivity across a wide m/z range, MALDI-TOF MS is particularly beneficial for identifying viral and bacterial strains associated with diseases such as hepatitis, pneumonia, and influenza.4750) They also benefit from other high-resolution instruments, such as when combined with Fourier-transform MS.51,52) In recent years, the combined benefits of MALDI-TOF and MALDI-MSI, coupled with AI integration, have generated significant interest among researchers, as will be detailed next.

3. SAMPLE PREPARATION METHODS TO ENSURE DATA QUALITY

High-quality data is essential for AI model training as it directly impacts the accuracy, reliability, and predictive outcomes, ensuring that the model can effectively learn meaningful patterns while minimizing errors and biases. In fact, data reproducibility and spatial resolution are common challenges in MALDI and MALDI-MSI data analysis. These two factors are critical concerns as they can significantly impact the predictive performance of ML models.53,54) Optimizing the MALDI-MSI pipeline requires improvements in sample preparation, instrumentation to optimize sensitivity, as well as data analysis workflows. A quality control framework is beneficial for addressing common challenges in mass analysis, such as ion suppression and sample heterogeneity.55) In MALDI-MS and MSI, sample preparation and laboratory protocols can significantly impact data quality.56) A known issue with MALDI samples is that samples are usually inhomogeneous, leading to unpredictable data quality and insufficient reproducibility. Furthermore, MALDI-MS and MALDI-MSI also require distinct sample preparation methods.

For effective MALDI ionization, sample/matrix co-crystal formation is critical. MALDI facilitates biomolecule ionization without degradation by embedding analytes within organic matrix crystals. Matrix selection and application method play a significant role in sample preparation. Commonly used MALDI matrices include 2,5-dihydroxybenzoic acid,57) α-cyano-4-hydroxycinnamic acid,58) and sinapinic acid.59) Typically, the analyte and matrix are dissolved separately in different solutions, and they are mixed in a 1 : 1 (v : v) ratio.60) To obtain the sample crystal, an aliquot of the mixture, typically a few microliters, is deposited and dried on MALDI sample plates. Environmental factors, such as substrate temperature and drying conditions, can influence heterogeneity, which may confound the results.60) As a result, many new sample preparation protocols have been proposed to improve the outcomes of measurements, including incorporating nanomaterials, sample re-engineering, or enhancing hydrodynamic flows.6163)

The requirements in MALDI-MSI sample preparation differ from conventional MALDI-MS. In biomedical applications,64) MSI is typically used to examine molecular distributions in tissue sections. Hence, it is difficult to extract the molecules from their original location in tissues. Since a successful MALDI reaction requires the incorporation of matrix compounds around biomolecules, modified sample preparation methods without complex molecular extraction must be used. In biological specimens, the choice of matrix and its application technique is usually a compromise between preferred analyte classes and extraction efficiency. Pneumatic spraying is an effective and popular technique for applying the matrix.65)

One key factor in tissue analysis is surface wetness during matrix application, which affects analyte extraction efficiency.66) For instance, excessively moist applications may lead to molecular diffusion, and wet methods usually result in larger matrix crystals. However, dry methods can prevent molecular diffusion, but they usually require an additional incorporation spray step to enhance analyte extraction,67) since overly dry matrix solutions may fail to extract surface molecules efficiently. The choice between “wet” and “dry” application methods also affects crystal size, and subsequently the spatial resolution of images. That is, smaller crystals improve spatial resolution and data reproducibility. It may also minimize data fluctuation across different laboratories. Literature also shows that additional spray is helpful to enhance sensitivity. An example is that ammonium citrate washes and pre-spraying with cyclohexane significantly improved clozapine extraction from rat brain tissue sections.68) To improve homogeneous matrix application, researchers have also explored alternative deposition techniques. Electrospray deposition, often performed using custom-built systems, has been particularly effective in controlling matrix crystal size and improving spatial resolution for imaging applications.69) Applying an electric field during matrix deposition has proven beneficial, leading to the production of smaller matrix crystals, which enhance spatial resolution.70)

Sublimation is another highly efficient method known for its ability to produce small and uniform matrix crystals.71) This technique is widely recognized for improving image resolution in MALDI-MSI. In the sublimation process, samples and powdered matrix are placed inside a vacuum chamber, with the sample surface positioned upside-down above the matrix powder. A heated bath inside the chamber sublimates the matrix, forming a uniform, pure matrix coating. Such homogeneous samples reduce variability in MSI data and enhance spatial resolution. Applications of sublimation-based methods have been demonstrated in mapping lipid distributions in brain tissues, high-resolution imaging of cancer markers, and drug distribution analysis in preclinical studies. Literature shows that the homogeneous samples produced quality images, as demonstrated for the analysis of phospholipids, neutral lipids, and proteins.7274)

4. DATA ACQUISITION FOR AI MODELS TRAINING

Recent advancements in raw data acquisition and data mining techniques underscore that a sophisticated understanding of MALDI-TOF and MALDI-MSI is highly beneficial for AI applications. MALDI-MSI enables complex biological studies by providing detailed insights into molecular interactions in biological samples, structural characteristics of chemical compounds, and protein behavior and post-translational modifications. These applications have positioned MALDI-MSI as a critical tool in computational biology. Figure 3 illustrates the workflow of raw data acquisition, preprocessing, and peak classification, with the aid of ML in recent years.75)

Fig. 3. Common applications of ML in MS. (A) Gradient-boosted decision trees applied for preprocessing in classification of peaks containing isotopic clusters; (B) prediction of structural features in analytes using feed-forward neural networks for spectral analysis; (C) prediction of peptide ion abundances by DL-based DeepScp pipeline; (D) classification of tissue samples analyzed through mass cytometry; (E) CNN propagated for MSI data analysis. Reprinted by permission from Beck et al.75) CNN, convolutional neural network; DL, deep learning; ML, machine learning; MS, mass spectrometry; MSI, mass spectrometry imaging.

One of the major challenges in biomolecular data acquisition is ensuring accurate tumor cell proliferation analysis via biomarker identification. A fundamental requirement for achieving high accuracy in ML and DL models is the availability of large, high-quality training datasets. For biomarker analysis, typical training datasets consist of data from normal and patient samples, with total spectra of roughly 10,000. Tissue sections (brain, liver, skin cross-sections, 5–20 μm thickness), and microbial colonies and smears are all popular samples in biomedical research. To validate the model, a testing dataset from normal and patient samples with each of a few thousand spectra, is necessary. Abdelmoula et al.76) conducted tissue-based MALDI-MSI experiments on 5 distinct intracranial glioblastoma multiforme (GBM) patient-derived xenograft models. Figure 4 presents the results for GBM,77) with data collection methodologies previously documented by the same research team. The datasets include hematoxylin and eosin staining-based images78) annotated by pathologists, which were manually propagated for training and testing in the MSI dataset.77) The spatial resolution used in such experiments is typically 100 μm. The MSI data were downloaded from SCiLS Lab 2020a (Bruker, Bremen, Germany) in imzML format, a standardized data format for MS imaging.79) The authors extracted the data in a readable format before performing preprocessing to ensure the data quality for AI training.

Fig. 4. Training and testing of DL models on tissue sections from several intracranial GBM PDX models. (A) Distribution of eight tissue sections for training and testing of DL models; (B) the number of spectra from normal and tumor regions and representative images for model training and testing; (C) H&E annotated tissue sections and a representative mass spectrum. Reprinted by permission from Abdelmoula et al.77) DL, deep learning; GBM, glioblastoma multiforme; H&E, hematoxylin and eosin; PDX, patient-derived xenograft.

Additionally, AI has the potential to integrate data from various sources, including MS and optical spectroscopy, providing a more comprehensive understanding of molecular interactions. However, the high computational requirements of existing stochastic approaches remain a barrier to their widespread use in dynamic studies. In MS, several data acquisition techniques are employed to collect informative and high-quality MS data, including multiplexed data-dependent or independent MS/MS acquisition.80) However, ensuring accurate and consistent MALDI and MSI datasets is still crucial and challenging. Particularly, hyperspectral MSI data is inherently complex, as it consists of numerous pixels, each associated with a detailed mass spectrum. So, a complete MSI measurement typically consists of thousands of spectra, depending on the raster step size (distance between adjacent sampling points) and sampling area. Typical MSI measurements require hours and considerable computational resources. AI-powered automated data collection can help reduce the time and labor required for data analysis.

Automating data acquisition in MALDI applications remains challenging not only in sample preparation, but also in acquisition methods. Klaila et al.35) introduced an algebraic topological framework that extracts intrinsic information from MALDI data to reflect topological persistence. The approach compresses MALDI data for optimizing computational time. The framework’s effectiveness was demonstrated on both real-world and synthetic MALDI datasets. Another recent advancement is the use of low-rank feature extraction of complex high-dimensional hyperspectral data. Xie et al.81) proposed a method to increase the speed of Fourier-transform ion cyclotron resonance MSI by leveraging these features, achieving a 10-fold improvement in throughput. Similarly, comparative analyses suggest that gradient averaging techniques provide superior resolution for high-density MSI data compared to low-resolution fingerprinting of biomarkers.81)

5. DATA ANALYSIS AND DL MODELS TRAINING

AI models, including ML and DL, are applied to MALDI-TOF and MALDI-MSI with various types of datasets, ranging from sequential data to two-dimensional (2D) images. Figure 5 illustrates the evolution of AI-based transformations, including ML and DL approaches applied to MSI data analysis from the 1960s to the present.82,83) Since AI technology has wide applications, here we highlight supervised ML and DL approaches used for MALDI-TOF and MADLI-MSI data analysis. DL models have powerful neural network layers capable of learning complex patterns from the given dataset, demonstrating remarkable potential in MS analysis compared to ML approaches.8486) Several DL models are widely used in MS analysis, including CNNs, ResNet,87) LSTMs,88) and RNNs.89) These models improve the depth and objectivity of MS data analysis, such as enhancing peak-picking efficiency, reducing experimental bias, and streamlining data processing workflows. Notably, DL models have also contributed to significant improvements in MALDI-MSI analysis, particularly in detecting tumors and abnormal cell growth in cancer.90)

Fig. 5. Chronological development of AI, ML, and DL in relation to algorithm/model design and their integration into technological innovations for MS. This schematic illustrates how advancements in AI methodologies have progressively influenced various phases of MS and MALDI-TOF and MALDI-MS data acquisition and analysis. The transition from basic algorithmic approaches to sophisticated DL models has facilitated a shift from molecular-level analytics to comprehensive applications in clinical diagnostics. AI, artificial intelligence; DL, deep learning; MALDI, matrix-assisted laser desorption/ionization; ML, machine learning; MS, mass spectrometry; MSI, mass spectrometry imaging; TOF, time of flight.

In DL studies, CNNs are particularly effective for classifying spatial features in biomedical datasets. CNN architectures offer multiple trainable parameters, allowing them to handle overfitting efficiently. Their ability to process 2D datasets makes CNNs particularly suitable for MALDI-TOF analysis, which involves 2D data. DL models have been deployed to enhance the accuracy of peak detection in MALDI-TOF and MSI data analysis by using large and cleaned datasets.91) In metabolite prediction, for instance, DL models have demonstrated high sensitivity, improving identification accuracy and retention time estimation. Another notable approach involved training CNNs on large MALDI-MSI spectra datasets, as demonstrated by Seddiki et al.92) Prior to training DL models using MALDI-TOF or MALDI-MSI datasets, standard experimental protocols must be followed, including data preprocessing steps such as normalization, resizing, noise reduction, splitting, and feature engineering. However, challenges persist, such as variations in sample quality, improper calibration, and excessive noise, which can hinder AI model performance due to the poor interpretability of the obtained dataset.93,94) When misinterpretation occurs in DL models, fine-tuning, data augmentation, early stopping,95) or in some cases ensembling,96) can be used for selecting alternative models.

Prior to training an ML model, several essential preprocessing steps are typically undertaken to ensure data quality and compatibility. These include data cleaning, outlier detection, normalization, transformation (e.g., converting categorical strings to numerical values), and optimization. In the specific context of MALDI-TOF and MALDI-MSI data, ML and DL models are commonly trained using selective preprocessing protocols. It is important to note, however, that not all preprocessing steps are universally applicable; their implementation depends on the nature and complexity of the dataset. For 1D and 2D MALDI spectra, common preprocessing techniques for noise reduction include TopHat filtering96) and ALS.97) The selection of these methods depends on the underlying data structure and the architecture of the intended model.98) In certain cases, post-training evaluations are used to reassess and redefine features based on the model’s feature extraction capacity. Following model training, evaluation is employed to assess performance, which may lead to feature redefinition, hyperparameter tuning, or the application of data augmentation techniques to enhance generalizability and accuracy. Among DL architectures, CNNs, transformers,99) and autoencoders98) have shown high efficacy for spectrometric data analysis. These models generally require input tensors and are implemented using popular DL frameworks such as PyTorch100) and TensorFlow101) frameworks.

Spatial distribution and spatial resolution are tightly interconnected within ML- and DL-based feature selection frameworks. Several technical challenges in applying DL to spatial resolution and spatial information in MSI data have been highlighted previously by Zhang et al.102) In particular, the abundance of biomolecules directly distributed on tissue sections contributes to the complexity of spatial data visualization, which in turn introduces significant computational burden associated with DL model training.103) These challenges include limitations in hardware and memory storage, especially when processing the spectral datasets generated in high-throughput MALDI-MSI experiments. A possible solution is parallel processing104) and concurrency.105) Recently, the roles of ML and DL in mitigating spatial resolution constraints in MALDI-MSI datasets have also been categorized by Zhang et al.106) To enhance spatial coherence, Tang et al.107) introduced a DL-based interpolation method that aligns tissue images in their respective orientations, thereby improving downstream analysis.108) Furthermore, Yee and Drum109) noted that RNN architectures, compared to CNN, demonstrate fewer constraints in certain data processing scenarios.

In the realm of spectral complexity, an emerging issue is the development of AI methodologies to address challenges associated with large-scale datasets, especially to resolve the spectral complexity problem without compromising essential spectral features. A fundamental challenge remains in effectively managing complex data structures, which often undermines the performance of ML models in terms of accuracy and dimensionality reduction, as mentioned above. For example, the msiPL method introduced by Abdelmoula et al. employs a DL-based variational autoencoder110) for data dimensionality reduction.76) Moreover, other functionalities such as molecular annotation, clustering, and segmentation are highly desirable for interpretability in biological contexts.

DL models have played a significant role in advancing new algorithms for MSI data acquisition and analysis.111113) Since 2021, there has been a steady increase in AI applications with the help of ML and DL used for MSI data analysis.113) These advancements have greatly enhanced tumor classification, protein behavior analysis, and other molecular studies.113) Recent research has highlighted the impact of AI,75) computer vision, ML, and DL techniques in analyzing various MALDI-MS data types. Over the past decade, DL has garnered significant attention, further advancing the field of MSI data analysis. Recent studies on how supervised and unsupervised learning methods employing robust ML and DL models contribute to the analysis of MALDI-TOF and MALDI-MSI datasets are summarized in Table 1.

Table 1. Recent studies highlighting ML and DL applications in MALDI-TOF and MALDI-MSI data analysis.

Applications ML approach Algorithm Reference
MALDI-TOF Supervised learning ML and DL López-Cortés et al.114)
Single-cell/MALDI-MSI N/A ML and DL Ngai et al.115)
Single-cell/MALDI MSI Supervised/unsupervised learning ML and DL Zhang et al.116)
MALDI-MSI Supervised learning DL Golpelichi and Parastar117)
Single-cell/MALDI-MSI Supervised learning ML Moore and Charkoftaki118)
Single-cell/MALDI MSI Unsupervised learning ML and DL Krestensen et al.119)
MALDI-MSI Unsupervised learning ML and DL Jetybayeva et al.113)

DL, deep learning; MALDI, matrix-assisted laser desorption/ionizatinon; ML, machine learning; MSI, mass spectrometry imaging; N/A, not applicable; TOF, time of flight.

5.1 Necessary steps while using DL for MALDI-MSI data analysis

One of the most basic and important steps for ML or DL model training is feature selection. In the context of biomarker detection using MALDI-MSI high-density data may sometimes fail to achieve high accuracy in classification tasks. DL provides powerful tools for biomarker detection in MALDI-MSI data by automatically learning spatial and spectral features. CNNs and other architectures are highly effective for detecting and classifying biomarkers in complex tissue samples, offering a valuable solution in medical diagnostics and research. Therefore, extracting features becomes essential when biomarkers exhibit unique characteristics, which requires a precise training of DL models. Other key considerations when selecting DL or supervised ML approaches for MSI datasets include reproducibility, data augmentation strategies, and gradient averaging techniques for optimizing sample size interactions.120) The reproducibility issue is largely determined by sample morphologies and the data acquisition techniques, as already mentioned previously. Data augmentation, on the other hand, helps reduce computational complexity by selectively extracting new data points from existing data. This technique significantly reduces the number of pixels required for imaging and reduces acquisition time.81) Recent studies have focused on enhancing MSI spatial resolution and reducing experimental duration. For example, Hu et al.121) introduced an effective DL-based dynamic sampling technique to reconstruct molecular images using a limited number of data samples. The method also predicted biologically significant regions for targeted MS sampling.

5.2 MALDI-TOF and MALDI-MSI peak classification using AI algorithms

Peak structure, peak position, peak alignment, and peak intensity are the common features to be considered when performing typical peak(s) classification with the help of AI models. DL algorithms play a critical role in peak classification and spectral annotation. Traditional supervised ML classification methods often focus on analyzing individual peak positions within training datasets. However, the use of deep neural networks for MALDI-TOF data analysis remains challenging, and no single standardized method has been established to date.121) Recent studies have employed CNNs to improve peak classification accuracy in MALDI-MS data analysis.122) Remarkably, CNNs exhibit high performance compared to other models in the case of MALDI-TOF datasets. By fine-tuning neural networks, it is possible to achieve higher accuracy in multi-classification tasks using large MALDI-MSI datasets. To demonstrate this, rat brain MSI data was analyzed, as shown in Fig. 6. The figure also compares principal component analysis (PCA) and ML models,123) demonstrating the power of AI since the ML/DL models are capable of extracting more subtle and intricate features. This allows for improved deconvolution of overlapping signals and a deeper understanding of the underlying molecular patterns, leading to clearer, more informative analyses of the MSI data.

Fig. 6. Hyperspectral modeling results using a RGB color-coding scheme to visualize rat brain images in MALDI-MSI analysis. (A) Schematic of the rat brain anatomy that was subjected to MALDI MSI with a 2-mm scale bar for reference; (B) randomly selected ion images of three distinct m/z values; (C) RGB color intensities overlaid for each pixel using the three images in (B); (D) the result in PCA space, where the position of each pixel determines the corresponding RGB intensity values; (E) the pixels within the highlighted box in the PCA plot (D) shown in the same color; (F) the image showing pixels with RGB color-coding in (D); (G) the result in a 20 × 10 × 5 3D SOM space based on unsupervised artificial neural network124) in which the level of the RGB colors determines their locations in the SOM; (H) the color pixels corresponding to the marked section in (G) in the original image; (I) full rat brain image reconstructed using SOM-based color-coding; (J) t-SNE model used to produce a scatter plot showing clusters in which pixels are color-coded in RGB according to their positions; (K) pixels highlighted in the boxed region in (J) visualized as corresponding colors; (L) final image output showing RGB color coding derived from t-SNE manifold learning. MALDI, matrix-assisted laser desorption/ionization; MSI, mass spectrometry imaging; PCA, principal component analysis; RGB, red–green–blue; SOM, self-organizing map; t-SNE, t-distributed stochastic neighbor embedding.

In Fig. 6, the authors provide a comparative analysis with the original study,125) where a different color-coding strategy was used to map rat brain images using clustering algorithms and ML techniques. In addition to PCA, 3D self-organizing maps (SOMs), and t-distributed stochastic neighbor embedding (t-SNE), RGB color channels were used to encode pixels. For example, the color codes are used to show the first three components in PCA, whereas in t-SNE visualization, colors are assigned to pixels based on their location within the 3D embedded space. In contrast, hyperspectral similarity maps were generated for the SOM analysis by assigning each neuron an RGB value corresponding to its x, y, and z coordinates in the 3D SOM. Each image pixel was then colored according to its most similar neuron, thereby preserving the topological relationships between spectral features and spatial localization. Collectively, the integration of ML techniques proved instrumental not only in automating data processing but also in achieving an unbiased, interpretable visualization of high-dimensional MSI data.

6. DISCUSSION

The seasonal impact of AI applications in MALDI-TOF and MALDI-MSI has been steadily increasing in recent years. AI technology is demonstrating significant improvements in accuracy and predictive capabilities when applied to high-end MALDI-TOF and MALDI-MSI datasets, particularly in histopathological classification. In the future, integrating AI with MALDI and MSI is expected to enhance biomedical and clinical research, providing more precise molecular insights. Most significantly, DL methods, including CNNs and other advanced neural networks, have demonstrated advantages in MSI-based spectral identification and pattern recognition. However, the accuracy of MALDI-MSI spectral analysis still faces challenges. For example, Mohammad et al.95) identified key limitations in distinguishing clonal microbial strains using DL methods.

Notably, training DL models requires large-scale datasets compared to ML, which are often difficult to collect. Data augmentation is often necessary to compensate for limited training samples in imaging-based data analysis. Similar concerns also pose challenges in tumor classification, as features included in imaging data are not always well-defined, even with fine-tuning. Overfitting and data bias can also occur if the dataset is too small or lacks diversity.126) Therefore, instead of analyzing only a small subset of peaks in MALDI-TOF and MALDI-MSI, it is crucial to process entire MALDI-MSI datasets to avoid overlooking relevant ions and oversimplifying complex biological systems. On the other hand, manual peak selection may cause subjective selection of spectral peaks and introduce bias, particularly when handling large datasets. A lack of understanding of sample composition and fragmentation patterns can complicate data interpretation. Key examples include clinical sampling data, as histopathological examination of tissues by manual inspection may show common problems of misinterpretation. For example, studies involving liposomes containing the antibiotic prodrug cefditoren pivoxil, analyzed via ToF-SIMS, identified a significant C5H9O2- fragment ion generated from both prodrug and lipid fragmentation.123,127) This finding suggested potential issues with drug distribution within the liposomes. In contrast, overlapping spectra in MALDI-MSI make it difficult to capture subtle variations, reducing classification accuracy.

Overall, this review has highlighted some technical challenges associated with the application of ML and DL in MALDI-TOF and MALDI-MSI data analysis. In addition to outlining current advancements, we have also discussed limitations that persist within the computational environment, inappropriate datasets, and model selection. In summary, AI technologies continue to enhance the speed, robustness, and automation of data analysis and annotation, showing clear advantages over conventional computational approaches.126)

7. PERSPECTIVES

MALDI-TOF and MALDI-MSI continue to be indispensable tools in biomedical and biochemical research. Their integration with AI-driven computational biology holds great promise for advancing molecular analysis. Continued improvements in data acquisition, preprocessing, and DL algorithms are essential to overcome current limitations and fully realize the potential of AI applications in MALDI-based techniques. While this review has outlined notable AI-driven advancements in MALDI-TOF and MALDI-MSI data analysis, several challenges persist, particularly related to data quality and limited dataset sizes. In addition, to enhance peak classification and ML model performance, key issues must be addressed. These include optimizing feature selection strategies to improve classification accuracy, addressing sample-related limitations such as small sizes and high biological variability (especially pertinent to MALDI-MSI), and improving the robustness of DL models by mitigating overfitting through careful dataset curation, which is often caused by high-throughput data. Advanced preprocessing methods are also needed to extract meaningful information from inherently noisy MALDI-MSI data. Additionally, computational and memory constraints remain a significant concern, as processing large MSI datasets requires substantial computational resources. Future developments should also prioritize the creation of efficient multimodal imaging workflows that offer higher spatial resolution. The future of MALDI-TOF and MALDI-MSI, combined with ML and DL, holds significant promise for rapid diagnosis by reducing false-negative results. The novelty of DL models lies in their ability to identify potential biomarkers and develop effective drug therapies as well.

ACKNOWLEDGMENT

This work is supported by the Academia Sinica (Grant No. AS-IA-113-L01) and the Ministry of Science and Technology of Taiwan, the Republic of China (Contract No. 113-2113-M-001-021).

Notes

Mass Spectrom (Tokyo) 2025; 14(1): A0175

REFERENCES
 
© 2025 Ali Farhan and Yi-Sheng Wang

This article is licensed under a Creative Commons [Attribution-NonCommercial 4.0 International] license.
https://creativecommons.org/licenses/by-nc/4.0/
feedback
Top