Introduction

The number of cancer-related deaths has increased globally over the past few decades, with hepatocellular carcinoma (HCC) ranking among the top three in mortality rates. Although effective diagnosis and treatment have made significant progress in the early stages of HCC, the overall survival rate after 5 years is still low (only 50–70%). Clinically, the gold standard method for diagnosing HCC is histopathological observations performed with haematoxylin and eosin (H &E) or immunohistochemical staining, which is time-consuming and subjectivity from pathology specialists. Therefore, the rapid and accurate diagnosis methods for HCC are needed. Mass spectrometry (MS)-based proteomics as a high sensitivity analysis technique has a powerful capacity to quantify numerous proteins in complex biological samples in several hours, which is feasible for the examination of isolated tissue samples in a short time1,2.

The mainstream method of traditional proteomics analysis is to first identify peptides based on secondary mass spectrometry (MS2), using MS analysis software such as MaxQuant3 and MSGFPlus4 to search databases, and then identify potential tumor biomarkers5,6 and determine the probability of cancer through machine learning methods7,8,9. For example, Zhang et al. found five urine biomarkers associated with lung cancer using random forest algorithm based on urine from 231 patients10. Sun et al. employed deep neural networks (DNN) to identify 19 biomarkers of thyroid cancer from 288 thyroid tissue samples11. Zhu et al. performed a quantitative analysis in 19 patients with HCC to quantify 2579 swissProt proteins with high confidence, in which 16 promising candidate tumor markers were highlighted using a cluster heat map approach12.

However, since certain errors occur in the process of peptide identification and protein, many researchers focus on the application of raw MS data, mainly using machine learning models to mine raw data features to promote clinical application of tumor prediction. Giordano et al. firstly adapted multivariate statistical analysis to collected features of ion peaks from 222 full-scan MS data, and the random forest model subsequently was trained based on features of ion peaks to classify tumor tissues and adjacent non-tumor tissues13. Deep learning, as a branch of machine learning14,15,16, has the ability to handle large-scale, high-dimensional data17,18,19,20, which makes it an ideal tool for processing raw MS data. In recent years, deep learning models have exhibited exceptional capabilities in various medical diagnostic applications, encompassing fields such as radiomics and proteomics21,22,23. There are also some studies that have leveraged deep learning approaches to analyze raw mass spectrometry data for the purpose of tumor classification. Wang et al. used a deep learning network to build the MSpectraAI platform based on public MS data of six types of tumor24. Zhang et al. converted the MS2 data of DIA into a (cycle, window, m/z) Tensor data structure, which was feed into the ResNet model to obtained a HCC classifier25. The above deep learning classification methods all convert raw MS data into 2D or 3D heatmaps. However, due to different settings of some parameters (such as chromatographic gradients) in different batches, the retention time (RT) and intensity information of the same peptide in different batches have great differences, which bring some difficulties for HCC diagnosis and biomarker discovery.

In the field of deep learning, Transformer26 as a neural network model for sequence modeling27,28,29,30 has been widespread used in natural language processing (NLP)31,32,33 and achieves remarkable performance in various tasks such as machine translation34,35 and information extraction36,37 due to its ability to capture contextual dependencies between sequences. In this work, we developed a new model architecture named MS1Former (an improved transformer encoder model) for HCC disease prediction in humans from MS1 data. Here, we accumulated MS1 spectra along RT dimension to implement data dimensionality reduction from 2D heatmaps to 1D sequences. The input m/z feature sequence of our method is similar to a text sequence, so the HCC diagnosis task can be regarded as a text classification problem. In this paper, we further explored the capacity of MS1Former to learn highly heterogeneous peptide features and the robustness for the high-resolution mass spectrometry obtained from Orbitrap series instruments.

Results

Study design and analysis workflow

Figure 1 presented the overall pipeline of MS1Former. Wherein, Fig. 1(A1) showed the data were obtained based on the LC-MS/MS method. In detail, the samples were hydrolyzed into peptide fragments by trypsin, and then analyzed by Orbitrap series mass spectrometers equipped with Nano-ESI. Each mass spectrometer adopted different LC conditions, such as gradient time and flow rate, as well as different instrument parameters, including MS1 range. Detailed parameters can be seen in Table 1. Subsequently, we used the ThermoRawFileParser tool to convert raw data (.raw/.wif) into mzXML or mzML format, and the data set (.mzML) was parsed to obtain MS1 spectra(RT, m/z, intensity). Additionally, MS1 spectra were visualized as heatmap images (as seen in Fig. 1(A2)), and the MS1 data collected by the four Orbitrap mass spectrometers does not have a uniform pattern in the RT dimension (as shown in the red box). Figure 1(A3) showed the pattern detection for malignant and normal samples in m/z dimension. The heatmaps showed that the peak intensities distribution in malignant samples is much higher than those in normal samples(as shown in the red box), which means that the peaks of the proteome are more unstable in tumor. The density plots of the m/z are much more consistent and centralized in normal samples while the intensity distribution in malignant samples is dispersed and inconsistent (seen in the red arrow). Therefore, there is a significant difference in the raw data distribution between malignant and normal samples in the m/z dimension.

Table 1 Summary of LC conditions and mass spectrometer parameters.
Full size table
Fig. 1
figure 1

Overall pipeline of MS1Former. (A1) Tissue samples are hydrolyzed into peptide using trypsin, MS data was acquired using four different mass spectrometry instruments containing nano-electrospray ionization (Nano-ESI), (A2) The MS data was visualized, and (A3) Pattern detection for tumor and non-tumor(normal) samples across m/z dimension, the heatmaps show the distribution of the intensities of spectra data, where the color from deep viridis to deep yellow means that the intensity values change from low to large. The density graphs show the distribution of each channel’s intensities, for example, the right density plots show the intensity distribution of 340-360 across the m/z dimension. (B) Data processing, the front tail and rear tail (highlighted in two red boxes) was removed using the adapted noise removal method; subsequently, binning and normalization on m/z were performed to obtain the processed MS data. (C) MS1Former model was trained to diagnose HCC based on input binned m/z sequences. By using the combination of 1d-convolutional neural network (1d-CNN) and transformer encoding module, it achieve an accurate prediction for HCC, which was further verified through the distinguishable features (m/z, intensity), as seen in the red box. The MS data visualization results are generated using ProteoWizard 3.038, and the overall graph is created using Figma online.

Full size image

Figure 1B showed the detailed implementation of the data process in MS1Former. Before the liquid phase peak stabilizes, the interference peaks close to the starting time are removed from the data, and the short signal peaks within a short period of time before the end of the peak are removed from the data to reduce the interference of these impurity peaks on the input signal data. We applied an adapted process of removing noise to remove tailing data treated as noise at the beginning and end of the MS acquisition process (highlighted in two red boxes). Then, the binning and normalized process of m/z was performed. Herein a window across the dimension m/z (the range between two red dashed lines is referred to as one window or a bin, i means total window number) was split and all peak intensities in each window were summed. After summation, the intensities in each window are normalized by dividing the maximum intensity of all windows in a spectra data24. Intensity values of binned m/z were accumulated along the RT dimension to obtain the m/z sequence data (m/z, intensity), which served as input for subsequent modeling.

Finally, we developed a classification model framework with an improved transformer encoder model in MS1former to diagnose malignant and normal HCC tissue in humans (as seen in Fig. 1C). The MS1Former framework mainly consists of a CNN layer, a transformer encoder module, and a feed forward neural network block for classification. The encoder module of transformer is mainly composed of the multi-head attention mechanism that transform every position into the input sequence by computing a weighted sum across the representations of all other positions in the sequence, feed-forward neural network, residual connections and layer normalization. Here, the binned m/z sequences obtained from the data processing were subjected to weighted combinations through CNN kernels to obtain a local feature embedding for each m/z. The transformer encoder module further performed encoding on the hidden vector, which captures long-range dependencies between each m/z and the rest of the sequence. Due to the encoder model directly interacts with all other positions in the sequence, information can transform better between distant elements. So the model can effectively gather information from all relevant regions of the m/z sequence to improve the classification accuracy of malignant and normal tissues. Additionally, raw MS1 spectra of malignant and normal tissues were exhibited to further verify the high-performance of classification models through the distinguishable part of raw MS data (as seen in the red box).

Model performance evaluation and comparison

We performed five-fold cross-validation using the dataset PXD006512 (80% for training, 20% for validation) to evaluate the performance of the model. We obtained an mean accuracy of 0.934, mean precision of 0.926, mean recall of 0.930, and mean F1 score of 0.929 (Table 2). The ROC curve of four different model on PXD006512 dataset is shown in Fig. 2A. Figure 2B shown a UMAP plots indicating the separation between malignant and normal groups through latent space of m/z for PXD006512 dataset. To test whether MS1Former is generic, the model is applicable to other external test datasets including WL-2023, WL-Fast, PXD002171 and PXD021979. As seen in Table 3, for each dataset, the accuracy was over 0.84, and the highest accuracy was 0.95 on the PXD021979. The precision of the four test datasets was all over 0.89 and the Recall was over 0.82, meanwhile, the F1-score was over 0.85 and AUC was over 0.90 for four test datasets. The WL-Fast test datasets exhibited lowest assessment on accuracy, precision, recall, F1-score and AUC among those datasets because of the lack of information when adopting the full scan spectra with fastest gradient(only 39 minutes). The detailed classification results of the four external test datasets are shown in Fig. 2C.

Table 2 5-fold cross-validation results of MS1Former on the PXD006512 dataset.
Full size table
Table 3 Performance of MS1Former on four testing sets consisted of WL-2023, WL-Fast, PXD002171 and PXD021979.
Full size table

We further compared the classification performance of MS1Former with other deep learning-based methods containing MSpectraAI24, MaxQuant+RandomForest(MaxQuant+RF), and ResNet-1825 based on the four external test datasets including WL-2023, WL-Fast and PXD002171 and PXD021979. The MSpectraAI, MaxQuant+RF, and ResNet-18 was likewise trained with PXD006512, notably, for the MaxQuant+RF, the protein identification was firstly performed to obtain potential relative proteins, and then the classified model was developed based on those potential proteins, the detail method was described in25. Our results showed that MS1Former outperformed other methods, as evaluated by the accuracy and AUC, nevertheless, the ResNet-18 indicated a worst generalization performance on the four test datasets (Fig. 2D, E). Particularly, the WL-Fast dataset was full scan data and there is no MS2 to identify protein via MaxQuant, so the MaxQuant+RF was not suitable to the WL-Fast. From the above result analysis, we can see that MS1Former has good performance in tumor and non-tumor samples classification.

Fig. 2
figure 2

MS1Former performance validation and testing in independent datasets. (A) ROC plot of four different machine learning models on PXD006512 dataset. (B) The UMAP plots showing the separation between malignant and normal groups in the PXD006512 dataset with latent space of m/z features. (C) Confusion matrix on four external test datasets including WL-2023, WL-Fast, PXD002171, and PXD021979, wherein, a showed the confusion matrix of WL-2023; b showed the confusion matrix of WL-Fast; c showed the confusion matrix of PXD002171; d showed the confusion matrix on PXD021979. (D, E) Performance comparison of Ours(MS1Former), MSpectraAI, MaxQuant+RF and ResNet-18 on four external test datasets including WL-2023, WL-Fast, PXD002171 and PXD021979. (F) The performance of the different binned size for the HCC prediction using MS1Former on dataset PXD006512.

Full size image

In most cases, the range of ion m/z scans and the number of peaks in each spectrum were first discretized19. We employed a binning method that the whole m/z range was divided into equal bin (window), and the bin size (window size) here can be designed freely according to the complexity of data to improve the model’s performance (Detailed in Methods). We used Ten bin sizes containing 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0 for the samples to determine the peak-intensity distribution that affect model’s performance as evaluated by accuracy, F1-score and AUC based on PXD006512. As seen in Fig. 2F, the model’s performance gradually improved as the the bin size decreases.The accuracy, F1-score, and AUC of MS1Former was highest while bin size set as 0.1, the model learn high-resolution specific characteristic distribution of peptide while the peak-intensity distribution become more refined as the decrement of bin size. So, our results showed bin size of 0.1 or even higher should be determined to improve model’s performance.

Fig. 3
figure 3

The result of Local Interpretable Model-Agnostic Explanations (LIME) for MS1Former. (A) LIME values shown for the top ten binned m/z index using WL-2023 that had the highest contribution to a correct HCC prediction. The average correlation corresponds to whether the feature is up- (red) or down- (blue) regulated. (B) Take binned value 3333 as an example, we listed some of the raw m/z values. The similar RT and m/z values are divided into one group, and different groups can be identified as different peptides.

Full size image
Fig. 4
figure 4

Beeswarm distribution of important features for WL-2023 and corresponding raw mass spectra data. The certain binned m/z (binned size was set as 0.1) intervals is corresponding to the raw mass spectra data, as seen in red box.

Full size image

Model interpretability

LIME (Local Interpretable Model-Agnostic Explanations)39,40 analysis was used to identify the metabolites and the corresponding mass spectrometry-based method that contributed the most to the prediction of HCC using the WL-2023 data set. LIME mainly uses ablation methods to identify which binned m/z has more important influence on the classified results, meanwhile, give the importance value for the binned m/z. We have listed the top ten m/z binning index using WL-2023 that had the highest contribution to a correct HCC prediction (Fig. 3A). When obtaining key classification features, we restored the m/z information in the RT time dimension to provide data information for peptide analysis. The raw m/z values corresponding to the top ten m/z binning indexes is also given in the Supplement A. Some of the raw m/z values corresponding to binned value 3333 are shown in Fig. 3B. Similar RT and m/z values are divided into one group, and different groups can be identified as different peptides, as described in DirectMS141,42. Later, we can further study the segmented merging of adjacent RT times to better improve the classification prediction ability of the model.

Additionally, Fig. 4 showed beeswarm plot43 of important binned m/z features for WL-2023 and corresponding mass spectrometry data, the numbers of important features for malignant tissues was larger than that for normal tissues. The significant differences of binned m/z features shown in the beeswarm distribution between malignant and normal tissues was mapped with the raw MS. For example, as the range of binned m/z was 3400-4100, the significantly remarkable difference of the m/z distribution between HCC malignant and normal tissues, besides, the corresponding raw mass spectrometry also present a notable differences (seen in red box). Therefore, we observed that MS1Former is capable of effectively learning distinguishable m/z feature distributions between normal and malignant tumor tissues, which can better guide the inference of the model.

Discussion

In recent years, deep learning models have demonstrated remarkable performance in various medical diagnostic applications, including radiomics and proteomics. For instance, a systematic review by Martinino et al. highlighted the potential of artificial intelligence in enhancing the detection and characterization of HCC through the analysis of imaging data, such as CT, MRI, Ultrasound (US).23 In this study, we focused on the use of mass spectrometry data for HCC diagnosis.

We established an end-to-end framework named MS1Former allowing us to directly analyze raw data with serialized LC-MS featurlization. This framework involves the featurlization and classification of raw MS data based on the deep learning method. The potential of MS1 data for HCC diagnosis is confirmed. The input m/z feature sequence of our method is similar to a text sequence, so the HCC diagnosis task can be regarded as a text classification problem. Additionally, our method bridges DDA and DIA proteomics directly to deep-learning technology such as a transformer encoder network, which can handle DDA, DIA, and full scan data. We foresee the potential of MS1Former framework in rapid clinical diagnosis, which could be applied to the diagnosis of many tumor types.

Several prior investigations have explored the integration of mass spectrometry data with machine learning methodologies for disease diagnosis. However, many of these approaches are limited by their reliance on additional data preprocessing or specific requirements for mass spectrometry data. In contrast, our proposed methodology, MS1Former, offers distinct advantages. For example, the MaxQuant+RF approach requires the identification and quantification of proteins before applying random forest models for prediction. Conversely, our method enables direct classification without the need for these additional steps. The ResNet-18 architecture was initially designed for image data analysis. Researchers have adapted it for mass spectrometry data by transforming MS2 data into three-dimensional tensors indexed by circle, m/z, and window, and then using the ResNet-18 framework for classification. In comparison, MS1Former is capable of directly modeling based on MS1 data, thereby simplifying the workflow and avoiding the need for complex data transformations. While MSpectraAI also employs MS1 data for disease diagnosis, it does not explicitly account for the relationships among different scans within the same sample, classifying each scan independently. In contrast, MS1Former integrates all scan information from a sample, facilitating end-to-end diagnostic predictions that capture the comprehensive characteristics of the data. Through evaluations on various HCC datasets, we demonstrate that MS1Former consistently outperforms MaxQuant+RF, MSpectraAI, and ResNet-18 across multiple classification metrics, underscoring its superior performance and versatility in HCC diagnosis.

Overfitting and underfitting are inevitable risks in the field of deep learning. Especially, in the study of small sample classification, overfitting is a noteworthy problem. Here, we avoid overfitting in our work by adopting a relatively larger dataset to increase the sample size. Therefore, the MS data (PXD006512) of 220 HCC samples consisting of 1488 raw files was used as the training set. In this data set, tumor and paired non-tumor tissues in clinically early HCC patients were collected and analyzed by Orbitrap Fusion mass spectrometer. However, the MS data collection just through Obitrap series instruments, so in the future, more types of MS datasets were collected to validate our model. Additionally, during the training process, we implemented a combination of techniques, including dropout, L2 regularization, and early stopping, to further alleviate the overfitting problem.

Binning is a significant step to discretize data suitable for the deep learning method. As the bin size decreases, the precision of identifying peptides increases, which leads to a significant improvement in the model performance (seen in Fig. 2F). Due to the memory and time-consuming limitations of the training machine, we only conducted experiments with bin sizes between 0.1 and 1. Additionally, although we can discover significant binned m/z features through the interpretability method, and then obtain a lot of important m/z features, it’s not that precise. In the future, we can try to carry out more experiments on MS2 and identify biomarkers as a further research direction.

In order to further evaluate the ability of our model to judge the disease course and identify the key discriminant components in the model. It can be found that the differences in the pathological tissue images can be well captured by the model (seen in the corresponding feature heatmaps), and the characteristic differences are shown on the raw mass spectral peaks, the results are shown in Fig. 5. The figure shown that the model learned the m/z differences of different HCC tissue differentiations. In the future, we also can further train the model for the judgment of pathological staging and identify the key biological influencing factors that affect classification based on the model.

Fig. 5
figure 5

Differences between different tumor differentiation types in pathological images, heatmap feature images and raw mass spectrometry data. The differences in the pathological tissue images can be well captured by the model (seen in the corresponding feature heatmaps), and these differences can be verified on the distinguishable features (m/z, intensity) on the raw mass spectrum peaks. The MS data visualization results are generated using ProteoWizard 3.038, and the overall graph is created using Figma online.

Full size image

In conclusion, despite the promising results, our approach has several limitations that need to be addressed in future work:

  1. 1.

    Generalizability: The datasets used for training and testing were primarily generated using Orbitrap series mass spectrometers, and the generalizability of the model to other types of mass spectrometers and experimental conditions has yet to be fully established. To assess the performance of our method on non-Orbitrap series mass spectrometers, we conducted additional evaluations using a dataset (PXD004837) obtained from a TripleTOF 5600 mass spectrometer. Our model achieved an AUC of 0.8442 on this dataset. For a more detail analysis, please refer to Supplement B. Although this AUC is lower than that observed on the Orbitrap series datasets, it still indicates a certain degree of generalizability of our model. Even though, we acknowledge that further investigation is necessary to fully determine the model’s performance across a broader range of mass spectrometry platforms. In future work, we plan to expand the diversity and size of the datasets to rigorously evaluate the robustness and generalizability of the model. Additionally, we will consider extending our approach to other tumor types, thereby enhancing the versatility and applicability of our framework.

  2. 2.

    Biological Interpretation: While the interpretability framework LIME has been instrumental in identifying significant m/z bins, the biological relevance of these features remains inadequately explored due to the lack of MS2 data. Future work will focus on integrating MS2 data to perform peptide sequencing and to identify specific proteins and peptides corresponding to the significant m/z bins. This integration will enhance the biological interpretability of the model and facilitate the identification of potential biomarkers. To achieve this, we plan to leverage existing research on predicting proteins and peptides from MS2 spectra and incorporate these advancements into our framework. For example, the recently proposed MonoMS1 method44 attempts to use peptides identified by MS2 spectra as labels and train a peptide recognition model based only on MS1 features, which provides a valuable reference for our future research. By doing so, we aim to bridge the gap between the identified m/z features and their corresponding biological entities, thereby enhancing the translational potential of our findings.

By addressing these limiting, we believe the MS1Former framework can be further refined and applied to the clinical diagnosis of various tumor types, contributing to the advancement of precision medicine.

Methods

Hepatocellular Carcinoma dataset. The publicly available dataset PXD006512, PXD002171, and PXD021979 were obtained from Proteomics Identifications Database (PRIDE), more details are provided in previous publication45,46,47. Furthermore, all tissue samples for WL-2023 and WL-Fast were collected from the Shulan (Hangzhou) hospital, Hangzhou, China. This study was approved by Research Ethics Committee of Shulan(Hangzhou) Hospital(KY2023033). The study has obtained informed consent from the subjects or their guardians, and all subject information has been desensitized. All methods were performed in accordance with the relevant guidelines and regulations. For details on the processing of MS data for WL-2023 and WL-Fast, see Supplement B. The details of the five data sets were shown in Table 4.

Table 4 Summary of five datasets used in our experiments.
Full size table

Adapted process of removing noise. To reduce the interference of impurity peaks on the input signal data, an adapted function for elimination of tails was proposed, as seen in Fig. 1B. The steps as follow:

  1. 1.

    Firstly, summation of the intensities of all m/z at each RT point.

  2. 2.

    Subsequently, \(D_{value}(r_i,r_{i+1})\) is the peak summed intensities difference between two adjacent retention time points \(r_i\), \(r_{i+1}\), where i represents the RT, R represents the maximum of RT, \(0 \leqslant i \leqslant R\), and \(D_{value}\) as follows:

    $$D_{value}(r_i,r_{i+1})=\left| \sum \limits _{j=0}^{M} \,intensity_{i,j} - \sum \limits _{j=0}^{M}\,intensity_{{i+1},j} \right|$$
    (1)

    where j represents as m/z and M is the maximum of m/z.

  3. 3.

    Finally, \(D_{value}(r_i,r_{i+1})\) compare with the cutoff value, if it is greater than the cutoff value, then remove the summed intensity values before i. In this paper, the cutoff value is set as \(3\times 10^{7}\), which is counted from the raw MS data.

Binning of m/z. Binning index: \(I = round(\frac{m/z-m_{min}}{\gamma })\); \(\gamma\): bin size 0.1; \(m_{min} = 260\); \(m_{max} = 1800\); The intensity value corresponding to each binning index is the maximum intensity value within the interval; if there is no value in bin index, the intensity value is 0.

Model framework. The MS1Former framework consists of three parts: a 1-dimensional CNN with pooling, which is mainly used to perform word packet training operations in natural language process (NLP) to capture information between adjacent tokens within a local range; a transformer encoder block with relative position vector encoding was used to capture long-range interactions across m/z sequences48. Finally, the feedforward neural networks is used for HCC classification. The detailed model framework is shown in Fig. 6, and the settings for the key hyperparameters of the model are given in Supplement B.

Fig. 6
figure 6

MS1Former network structure consisted of 1d-CNN block, transformer encoder and feed forward neural network block. The number of trainable parameters for different parts of MS1Former are shown on the right side of the blocks.

Full size image

Multihead Attention layers was used to share information across the sequence and model long-range interactions. Each head has an independent set of weights \(w_{q} \in R^{C \times K}\), \(w_{k} \in R^{C \times K}\), \(w_{v} \in R^{C \times K}\), which transform the input sequence \(X \in R^{L \times C}\) into query \(q_{i} = X_{i}w_{q}\), keys \(k_{i} = X_{i}w_{i}\),and values \(v_{i} = X_{i}w_{v}\). The input consists of queries and keys of dimension \(d_{k}\). We compute the sum of the matrix of outputs and relative positional encoding \(R_{ij}\) as attention matrix, which is determined as \(a_{ij} = softmax\left(\frac{q_{i} k_{j}^T}{\sqrt{d_{k}}}v_{i} + R_{ij}\right)\), where the entry \(a_{ij}\) represents the amount of weight query placed at position i on the key at position j. Values represent the information that each position will propagate to the ___location that attends to it. Each single attention head computes its output as a weighted sum across all input positions. This allows each query position to use information across the whole sequence. The multiple heads compute with independent parameters, and we concatenate the outputs from each head to form the final layer output followed by a linear layer to combine them. Our layers used 8 heads, a value size of 24, and a key/query size of 48.

For injecting positional information, relative positional49 encodings \(R_{ij}\) were added into \(q_{i}k_{j}^{T}\). A parameterized baseline for how actively two positions in the sequence influence mutually during the layer’s transformation as a function of their pairwise distance was provided by relative positional encodings. The funcions term as formulated in Enformer48,50.

MS1Former training details. The 1488 HCC MS samples from public PXD006512 were divided into 5 groups. Four groups were treated as the training set, while the remaining one group was treated as the validation set. Additionally, three sets of DDA-MS (WL-2023, WL-Fast, PXD002171)and one set of DIA-MS (PXD021979) were used for testing. To differentiate spectra from tumor and non-tumor tissue areas, an active function softmax(·) is used to transform the output of the previous connection layer into a probability output, with a leaning rate of 0.001. To prevent overfitting, a dropout layer was employed with a rate of 50%. Additionally, cross-entropy is then used to define the loss function:

$$loss = -\frac{1}{m} \sum \limits ^{m} y_{true} \log y_{pred} + (1-y_{true}) \log (1-y_{pred})$$
(2)

m is the batch size, which is set to 8, \(y_{true}\) is the true label of the input x, and \(y_{pred}\) is the predicted score of the input x. During the training process, the model was iteratively trained for 200 times on the data set with Adam optimizer. The MS1Former was implemented in Pytorch and Python 3.8.

Conclusions

We present MS1Former, a powerful deep learning model that integrates CNN and Transformer blocks to facilitate HCC diagnosis using raw MS1 spectra. MS1Former heralds a novel era in the diagnosis of HCC through its ability to directly analyze raw MS1 spectra, offering a promising alternative to conventional, time-consuming diagnostic methods. The high accuracy and efficiency of our model, coupled with its potential for clinical applicability, underscore its potential for HCC diagnostics. In the future, we will focus on further exploring the efficacy of MS1Former in classifying multiple cancer types. Additionally, we plan to integrate MS2 spectra data into the model to enhance the predictive power and interpretability of the model, which is meaningful for advancing our understanding of HCC pathogenesis.