Прогнозирование аутоиммунных заболеваний на основе глубокого обучения.

Introduction

Autoimmune diseases, a group of disorders encompassing a wide range of conditions, are centrally characterized by an imbalance in self-tolerance, which triggers an immune response against one’s own tissues, leading to pathological changes and clinical symptoms 1 . These diseases are prevalent worldwide and their high morbidity and mortality pose a significant threat to public health 2 . In the last two decades, the population affected by autoimmune diseases has grown significantly, especially in economically developed countries 3 – 5 . The prevalence varies markedly from region to region, ranging from less than five cases per 100,000 population (e.g., uveitis 6 , Wegener’s granulomatosis 7 ) to more than five hundred cases (e.g., rheumatoid arthritis 8 and ankylosing spondylitis 9 ). More than 80 autoimmune diseases have been identified 10 , and their etiologic mechanisms are extremely complex, involving genetic factors, environmental influences, and abnormalities in immune cell function. Autoimmune diseases are usually preceded by mild inflammation 11 , disease-specific autoimmune markers, and nonspecific symptoms 12 . Significant differences in the autoantigens involved in different types of autoimmune diseases pose significant challenges in the assessment and diagnosis of the disease.

The autoimmune process in the human body mainly involves innate immune cells such as granulocytes and macrophages, as well as mechanisms of autoantibody complexes and aberrant presentation of T-helper cells and aberrant activation of T-eff cells. The etiology of autoimmune disorders is extremely diverse and complex and may include variations in Human Leukocyte Antigen (HLA) genes, aberrant expression of antigens and dysregulation of T-cell function. These conditions may be triggered by infection with external pathogens (e.g., Prevotella, EBV, cytomegalovirus, etc.) or by mutations in internal immune cell genes. Dysregulation of any of these can lead to immune abnormalities. In this paper, we will focus on some of the specific changes in the TCR experienced during the pathogenesis of autoimmune diseases. T cells, as the central regulators of the immune system, play a key role in the immune response by specifically recognizing antigens through TCRs on their surface. In response to diverse pathogens, the human adaptive immune system has constructed a large library of TCRs. During their developmental process, each T cell generates a unique TCR with specific antigen specificity through gene rearrangement, comprising either and chains or and chains. The variable region (V-region) of the TCR chain contains three highly variable loop regions, i.e., the complementary determining regions, CDR1-3 13 . CDR1 and CDR2 are encoded by the V-gene, and are essential for the interaction of the TCRs with the Major Histocompatibility Complex (MHC) complexes. CDR3, on the other hand, as the main cloning region of T cells, is encoded by the connecting region between V and J or D and J regions, which has a high degree of variability and directly determines the antigen-binding specificity of TCR. As the part of the TCR in direct contact with antigen, CDR3 plays a crucial role in the interaction of the TCR with peptide-MHC complexes.

Numerous studies have shown an increasing association between the TCR and a variety of autoimmune diseases, including Rheumatoid Arthritis (RA), Type 1 Diabetes Mellitus(T1D), Multiple Sclerosis (MS), and Idiopathic Aplastic Anemia (IAA).

RA, an organ-specific autoimmune disease, is characterized by symmetric peripheral polyarthritis, often with complications such as systemic inflammation 12 . RA affects 0.5–1% of the global population, with a higher prevalence in regions far from the equator 14 , 15 . The prevalence of RA in women is significantly higher than that in men, with a ratio of approximately 23:1 16 . The disease is categorized as an autoimmune disorder, in part due to the presence of autoantibodies such as rheumatoid factor (RF) and anti-citrullinated protein antibody (ACPA) in most patients, which are strongly associated with a more severe course of the disease 17 , 18 . Abnormal T-cell immune responses play a central role in the pathogenesis of rheumatoid arthritis, in which the TCR is a key factor. In the joint tissues of patients with RA, a large number of activated T-cells recognize and bind to RA-associated antigens through their TCR. When these RA-associated antigens, which are usually self-antigens, stimulate inflammation in the joints, the T cells are activated to release inflammatory mediators such as cytokines and chemokines, which exacerbate the inflammatory response in the joints, leading to joint damage and pain. In addition, aberrant activation of TCR and T cell-mediated immune responses play an important role in the pathogenesis of RA. Studies have shown that there are specific TCR subpopulations and clonal expansions in the T-cell population of RA patients, pointing out that abnormal T-cell immune responses and aberrant TCR activation may be one of the key mechanisms in the development of RA. Similar to diseases such as rheumatoid arthritis, the incidence of T1D is influenced by latitudinal gradients and migration patterns. Studies have shown that when populations migrate from areas of low prevalence to areas of high prevalence, their risk of T1D increases 19 , 20 . In addition, genetic factors play an important role in the risk of T1D. the risk of developing the disease in the immediate family members of T1D patients is 15–20 times higher than in the general population, increasing from about 0.4% in the general population to 25–50% in identical twins 21 . Abnormal T-cell responses of the immune system are also key factors in T1D, with helper CD4+ T cells and cytosolic CD8+ T cells playing a dominant role in driving pancreatic -cell destruction 22 , 23 . This destruction ultimately leads to loss of pancreatic -cell function and insulin deficiency. These aberrant T-cell responses may be associated with abnormal TCR function and immunoregulatory imbalance. MS, an inflammatory demyelinating disease of the central nervous system (CNS) 12 . Among the genetic risk factors, HLA alleles play a major role, especially the HLA-DRB1*1501 haplotype, which is significantly associated with the development of MS 24 , 25 . MS is more common in women, with a female to male prevalence ratio of approximately 2–3:1 26 . In the pathogenesis of MS, autoimmune-reactive T cells attack and destroy myelin sheaths in the central nervous system, a process that is thought to be closely related to aberrant activation of the TCR and T cell dysfunction. IAA is a rare autoimmune bone marrow failure disorder characterized by a decrease in the number of bone marrow hematopoietic stem cells leading to a decrease in the number of whole blood cells 27 . Krell et al. found that the T-cell repertories of children with IAA were predominantly dominated by public T-cell receptor clones, which was different from that found in normal children and adults, where the proportion of public clones was only 0.2% or 1.9%. In addition, the CDR3 sequence ASSGVGFSGANVLT was highly prevalent in three patients with hepatitis-induced anemia (15–32% of all sequences). In patients with IAA, 15 high-frequency sequences were found to be unique, suggesting that T-cell responses play a key role in the pathogenesis of IAA 28 .

In recent years, deep neural networks have made remarkable achievements in fields such as computer vision by automatic layer-by-layer feature extraction, combined with the localization and translation invariance features of convolutional neural networks 29 , 30 . In 1997, Sepp Hochreiter and Jürgen Schmidhuber pioneered the Long Short-Term Memory Network (LSTM) 31 , which is a kind of network that effectively mitigates the long-term dependency problem in recurrent neural networks. As its variant, Bidirectional Long Short-Term Memory Network (BiLSTM) consists of two LSTMs that process the input sequences forward and backward respectively and stitch the results to enhance sequence processing 32 . The emergence of these techniques has reduced the need for complex feature engineering in traditional machine learning, lowered the barrier to entry for practitioners, and enabled deep learning to be widely used in numerous fields, including immunology. In immunology, deep learning has become an important tool for neoantigen identification in immunotherapy. In the field of cancer prediction, T cells also play a central role in cancer immune response 33 . For cancer patients, some T-cells often undergo cancer-specific changes during tumor progression 34 , and the corresponding TCRs is referred to as cancer-associated TCRs (caTCRs). It has been suggested that some caTCRs may share common biochemical characteristics 35 , 36 . Therefore, distinguishing caTCRs from normal TCRs has become an important topic. To address the challenge of cancer detection, Xiong et al. employed TCR sequence analysis to predict the tumor status of patients 37 . In addition, Beshnova et al. developed a deep learning-based model called DeepCAT 38 , which is capable of predicting cancer-related T-cell receptors from blood samples for noninvasive cancer detection. DeepCAT performed excellently in a variety of early-stage cancer predictions with an area under the curve (AUC) of more than 0.95, which laid the groundwork for noninvasive peripheral-blood TCR library-based cancer detection. Xu et al. proposed the DeepLION model 39 , which further improves the prediction accuracy by employing a multi-instance learning approach, considering TCRs of different lengths, and using convolutional filters of different sizes and 1-max pooling operations to process TCR sequences. DeepLION has shown high accuracy, sensitivity, and specificity on a number of cancer cohorts. Especially on thyroid and lung cancer cohorts, its AUC reached 0.97 and 0.90, respectively. MINN-SA, a multi-instance neural network model based on the sparse attention mechanism proposed by Young hoon Kim et al 40 , used the sparsemax function to sparse the attention scores, thereby adaptively focus on key instances in the samples. MINN-SA performs well in processing datasets of multiple tumor types, outperforming traditional machine learning and other deep learning methods in terms of accuracy and efficiency.

This study aims to fill the gap in the application of deep learning algorithms for the prediction of autoimmune diseases. Four autoimmune diseases, namely “Type 1 Diabetes”, “Rheumatoid Arthritis”, “Multiple Sclerosis” and “Idiopathic Aplastic Anemia”, are selected as research targets and two prediction models with different architectures are proposed. The first is the AutoY model based on convolutional neural network, which designs multiple convolutional kernels for different CDR3 residue regions in the CDR3 sequence and applies global maximum pooling to efficiently extract features that are closely related to autoimmune diseases. The second is the LSTMY model based on a bidirectional LSTM network, which employs the attention mechanism to filter and aggregate the effective features output from the bidirectional LSTM layer. We trained using TCR data from healthy people and patients with autoimmune diseases. The aim was to construct predictive models that would be able to distinguish patients with autoimmune diseases from healthy people. This study aims to provide new reference information for the diagnosis and treatment of autoimmune diseases. We hope that these methods will provide new ideas and tools for the early diagnosis and treatment of autoimmune diseases and promote the research and clinical practice of autoimmune diseases.

Results

We conducted 100 rounds of five-fold cross-validation for each autoimmune disease to evaluate the performance of the AutoY and LSTMY models on four autoimmune disease datasets (see Tables 1 and 2 ). Standard deviations are in parentheses.

Table 1 Average performance of the AutoY model on four autoimmune disease datasets. Types of autoimmune diseases Accuracy Sensitivity Specificity AUC RA 0.8764 (0.0225) 0.7115 (0.0629) 0.9426 (0.0125) 0.9375 (0.0007) T1D 0.9850 (0.0060) 0.9609 (0.0169) 0.9968 (0.0033) 0.9991 (0.0034) MS 0.9763 (0.0313) 0.9713 (0.0434) 0.9803 (0.0282) 0.9961 (0.0083) IAA 0.9279 (0.0158) 0.7637 (0.0626) 0.9801 (0.0091) 0.9750 (0.0099)

Average performance of the AutoY model on four autoimmune disease datasets.

Table 2 Average performance of the LSTMY model on four autoimmune disease datasets. Types of autoimmune diseases Accuracy Sensitivity Specificity AUC RA 0.8228 (0.0229) 0.5540 (0.0669) 0.9308 (0.0207) 0.8780 (0.0239) T1D 0.9641 (0.0120) 0.9237 (0.0287) 0.9837 (0.0118) 0.9932 (0.0069) MS 0.9814 (0.0085) 0.9731 (0.0098) 0.9881 (0.0128) 0.9963 (0.0022) IAA 0.9149(0.0202) 0.7068(0.0765) 0.9782(0.0110) 0.9533(0.0165)

Average performance of the LSTMY model on four autoimmune disease datasets.

Analysis of the results of AutoY

The predictive performance of the AutoY model for four different autoimmunities (i.e., RA, TID, MS, and IAA) is presented in Table 1 . Among them, the performance on T1D and MS is protruding, with AUC values of 0.9991 and 0.9961, respectively. The AutoY model not only has higher prediction accuracies (accuracy, sensitivity, and specificity of 0.9850, 0.9609, and 0.9968 for T1D, and accuracy, sensitivity, and specificity for MS, respectively, of 0.9763, 0.9713, and 0.9803, respectively), and showed high robustness in multiple cross-validation (standard deviation of the AUC values was 0.0034 for T1D and 0.0083 for MS). These results demonstrate the potential of AutoY in the clinical diagnosis and detection of these two autoimmune diseases. AutoY performs well on the IAA with an AUC of 0.9750. For the other evaluation metrics of the model, the accuracy and specificity values are high at and 0.9279 and 0.9801, respectively, but the sensitivity is comparatively low at 0.7637. The main reason for the low sensitivity of the model is the imbalance in the IAA dataset, where the negative samples have a size of 214 while the positive sample size is only 65. This can be achieved by decreasing the threshold for prediction, allowing the sensitivity to be increased to an appropriate level without significantly decreasing the accuracy and specificity. In addition, the imbalance in the training dataset is mitigated by adding new positive sample data, which in turn further trains AutoY to obtain a more practical model for clinical applications. Finally, the model performed moderately well on RA, with an AUC of 0.9375. the accuracy was 0.8764, while the sensitivity and specificity were 0.7115 and 0.9426, respectively. it should be noted that the variability of the sensitivity was significantly high, with a standard deviation of 0.0629, and the standard deviations of the other assessment metrics of the model were also high. One of the reasons for the low sensitivity and stability of the model is that the RA dataset is also unbalanced, with a positive sample size of 86. Another important reason is that the TCRs of RA patients have a large diversity, and it is difficult for the model to capture the features therein, resulting in insufficient generalization ability.

Figure 1 contains the ROC curves of the 5-fold cross-validation experiments of AuotY on these four autoimmune diseases. The experimental results show that AutoY has notable predictive performance on T1D and MS, with AUC values of 0.9992 and 0.9999, respectively, which are almost perfectly close to 1. On IAA, although the performance is not as good as that of T1D and MS, it can still reach an AUC value of 0.9740. In contrast, the model performs poorly on RA, with an AUC value of 0.9371. This suggests that AuotY can effectively differentiate between patients with certain autoimmune diseases, such as T1D and MS, and healthy individuals. This Fig. 2 shows the output of the AutoY model, demonstrating how AutoY categorizes healthy individuals and diseased patients: In the figure, the horizontal axis indicates the categorization of disease patients and healthy people, and the vertical axis indicates the Predicted Values. The red dotted line indicates the classification threshold of 0.5 for the Predicted Values, i.e., it is the Predicted Values greater than 0.5 will be considered as the disease categorized group, and vice versa for the healthy categorized group. The light green portion indicates the health categorical group and the light-yellow portion indicates the disease patient categorical group. Violin plot for RA. The predictive values for healthy samples range from 0 to 1, with a median near 0.1, below the threshold line region. This indicates that the model has high specificity and is good at correctly identifying healthy samples, with only a very small number of healthy samples misclassified as RA. The predictive values for RA have a much wider range of predictive values, ranging from 0 to 1, with a median near 0.8, and a larger number of patients have predictive values beyond the classification threshold line, misclassifying most RA samples as healthy. This further illustrates the low sensitivity of the model for RA disease; Violin plot for IAA. The violin plot of IAA exhibits similar characteristics as RA, i.e., the predicted values of most healthy individuals are mainly concentrated in the lower region (high specificity), whereas the predicted values of IAA patients are distributed over a wider range (low sensitivity), which fails to efficiently identify a larger number of positive class samples, resulting in more false-negative results. The violin plots of the two, aptly corroborate, the lower sensitivity values achieved by AutoY in the prediction of both RA and IAA, suggesting that AutoY performs poorly in identifying the positive classes of both diseases, and further optimization is needed; Violin plot for T1D.The high density of predictive values for T1D patients, with a median close to 1, implies that the model successfully identifies the vast majority of T1D samples (high sensitivity). The median for healthy samples lies between 0 and 0.1 and most of them are located within the decision threshold line region, which is consistent with their high specificity scores; Violin plot for MS. AutoY demonstrated high sensitivity and specificity in categorizing MS patients and healthy populations, with the distribution of predictive values for both mainly concentrated within decision thresholds, with only a few samples misclassified.

ROC curves of AutoY on different autoimmune disease datasets.

The violin plot of AutoY.

Analysis of the results of LSTMY

The predictive performance of the LSTMY model for four different autoimmunities (i.e., RA, TID, MS, and IAA) is presented in Table 2 . Again, noticeable performance advantages are found on sample data for T1D and MS, the AUC value of T1D reaches 0.9932, and the AUC value of MS is even higher than 0.9963, and the standard deviation of both is no more than 0.0069. For other evaluation indexes of the model, except for the sensitivity value of T1D, which is low at 0.9237, the accuracy and specificity value of both is higher than 0.9641, and the standard deviation of both is less than 0.0287. This indicates that LSTMY performs stably in accurately identifying patients with T1D and MS, with low predictive volatility and equally high robustness. LSTMY performs well on IAA with an AUC of 0.9533. For the other evaluation metrics of the model, the accuracy and specificity values are high, with and 0.9149 and 0.9782, respectively. However, the sensitivity is relatively low, with a mere 0.7068 and a standard deviation of 0.0765. It is worth noting that LSTMY performs poorly on RA, with an AUC of 0.8780. Where, the accuracy and specificity values were 0.8228 and 0.9308, respectively. Sensitivity was even only 0.5540 with a standard deviation of 0.0669. This indicates that LSTMY is deficient in recognizing IAA and RA patients in correctly identifying the diseased individuals, with poor stability and high fluctuations. This further illustrates the huge impact that the imbalance of the disease sample dataset brings to the model performance and the necessity of exploring the two diseases, IAA and RA, in depth.

The ROC curves of the 5-fold cross-validation experiments of LSTMY on these four autoimmune diseases are included in Figure 3 .

ROC curves of LSTMY on different autoimmune disease datasets.

The experimental results show that LSTMY has excellent predictive performance on T1D and MS, with AUC values of 0.9868 and 0.9938, respectively. The performance on IAA is good, but with an AUC value of 0.9554. In comparison, the model performs poorly on RA, with an AUC value of only 0.8697. Overall, although the LSTMY model performed slightly inferior to the AutoY model, it still deserves recognition. Both models show the same trend in predicting the four diseases, i.e., notable performance on T1D and MS, good performance on IAA, but relatively poor performance on RA. It can be seen that the two models are not applicable to all autoimmune diseases, and in the subsequent discussion we will continue to explore the reasons for the differences in their performance for these diseases and actively search for possible optimizations to improve the overall performance of the models. This Fig. 4 shows the output of the LSTMYmodel, demonstrating how LSTMY categorizes healthy individuals and diseased patients: Violin plot for RA. Overall, LSTMY has a weak predictive performance for RA. In the figure, the model performs well in classifying non-RA patients as healthy, with most of them clustered within the threshold region. However, more than 44% of the RA patient samples have predicted values outside the threshold region, illustrating the model’s poor performance in identifying positive classes, consistent with its lower sensitivity; Violin plot of IAA. Although the sensitivity of the model has improved over RA, more than 29% of IAA patients are still misidentified. Notably, it performed extremely well in identifying healthy samples with high specificity; Violin plot for T1D. The LSTMY model is more accurate in classifying it, with more than 96% of all tested samples being correctly classified. This is reflected on the violin plot by the fact that more than 92% of the T1D patients and 98% of the healthy samples have predicted values that lie within the decision threshold; Violin plot for MS. LSTMY demonstrated high sensitivity and specificity, with more than 98% of samples correctly categorized overall.

The violin plot of LSTMY.

Discussion

In this study, after gaining an indepth understanding of the pathogenesis of the four autoimmune diseases and the development of related fields, we built two prediction models, one is a deep neural network model based on DeepCNN architecture—AutoY, and the other is a prediction model based on the bidirectional LSTM integrated attention mechanism—LSTMY, both models have the ability to predict autoimmune diseases. Of course, our study has some limitations. Due to limited experimental conditions, we only conducted prediction experiments for four autoimmune diseases, which somewhat limited the scope of our tests and the validation of the applicability of our model. In terms of data processing, it is difficult for us to exclude the existence of common partial characteristics between the disease samples and normal samples, as well as the possible effects of other etiologies in the samples. Therefore, we aim to provide methods as well as ideas and validate the feasibility of our proposed model. Ideally, we would like to collect more and diverse data on autoimmune diseases through self-sequencing to validate the reliability and applicability of the model.

The experimental results show that the two models, AutoY and LSTMY, have notable performance in the prediction on the four autoimmune diseases. However, specifically, the AutoY model outperforms the LSTMY model in terms of accuracy, sensitivity, specificity, and AUC scores. In order to fully assess the robustness of the two models, we conducted 100 rounds of 5-fold cross-validation experiments for each of the four different autoimmune diseases and obtained ROC curves of the two models for each disease superimposed on confidence intervals with a coefficient of 0.95 (see Figs. 5 & 6 ).

ROC curves for AutoY model, superimposed with confidence intervals with coefficient 0.95.

ROC curves for LSTMY model, superimposed with confidence intervals with coefficient 0.95.

As can be seen in Figs. 5 and 6 , although AutoY slightly outperforms LSMTY, LSMTY exhibits superior robustness. This is particularly evident in the prediction of MS: AutoY’s results are more volatile, with a wider CI ranging from 0.9752 to 1. In contrast, LSMTY is less volatile in its predictions and has the narrowest confidence interval, with a CI ranging from 0.9898 to 0.999, demonstrating higher robustness. On the whole, AutoY and LSMTY show similar performance trends in predicting the four autoimmune diseases (RA, IAA, T1D, and MS). In particular, both models performed well in the prediction of T1D and MS. For T1D, both models had narrow CI intervals, indicating that they showed high stability and robustness in multiple rounds of cross-validation. It performs well in IAA prediction but less well in RA prediction, and its wider CI range indicates larger performance fluctuations and lower robustness in different cross-validation rounds.The variability in the performance of the two models shown by these results may be due to the fact that AutoY focuses only on high frequency regions of size 2–4 and one low frequency region of size 5 in the CDR3 residue region, excluding other regions of lower frequency. This both extracts features associated with autoimmune diseases to a large extent and excludes the effects brought by other regions.

LSTMY, on the other hand, focuses on the entire region, which inevitably introduces many noise effects. However, it is worth noting that when we apply the two models in predicting RA and IAA, the performance of both of them decreases, which is not in line with expectations. We believe that there are several reasons for this. One of the notable reasons is the imbalance in sample size. In our experiments, the sample sizes of RA and IAA are relatively small, with only 86 and 65 observations, respectively. This leads to an imbalance between positive and negative cases in the training data. In deep learning tasks, data imbalance may cause the model to be biased towards the majority category, thus affecting the learning of the minority category. As a result, the model may not have adequate information to make accurate predictions for diseases with small sample sizes such as RA and IAA. It can also be seen from the Results section that both models have high specificity, i.e., they have a high rate of accurate identification for healthy samples. Whereas, it shows low sensitivity in identifying both RA and IAA, i.e., there are a number of patient cases that have not been identified as autoimmune by the model. This may be due to the fact that the biochemical features of RA and IAA in the data collected are not significant enough in relation to the healthy samples for the model to learn useful features to correctly distinguish between positive and negative samples. In order to further analyze the reasons for the differences in the performance of the models on different autoimmune disease datasets, we plotted the distribution heat maps of the TCRs of the four autoimmune groups and the healthy group according to the visualization method proposed by Yokota et al. 41 . These heat maps reflect the TCR distribution characteristics and their differences in different groups (see Fig. 7 ).

Sample characteristics difference plot: the number of TCRs used in each sample group and their mean number of repetitions were as follows: healthy group (20,612 TCRs, mean number of repetitions 1.0382), RA group (7944 TCRs, mean number of repetitions 1.0826), IAA group (4133 TCRs, mean number of repetitions 1.5727), T1D group (8721 TCRs, mean number of repetitions 1.1179), MS group (4670 TCRs, mean number of replicates 3.7259). This analysis was based on the top 100 TCR sequences with the highest abundance in each sample group.

Among them, the distribution of TCRs in the T1D sample is quite close to that of the healthy group, and the average number of replicates of the TCRs in the both samples was slightly larger than 1, that is, there are almost no repeated TCRs. At this point, our model can accurately distinguish T1D cases from healthy cases, indicating that the model can effectively learn the representations related to autoimmunity in T1D, thus it has very good generalization capabilities. We can see from the heat map that the TCR distribution of MS departs markedly from that of the healthy group. In fact, this is beneficial for the model’s classification task because it is easier for the model to learn the features that distinguish them. This is also the reason why our model performs best on this dataset. However, it is noted that the average number of TCR replicates in this sample is 3.7259, which is much higher than other data sets, suggesting that the cases in the MS sample are highly homogeneous with respect to TCR. Therefore, it is reasonable to suspect that this dataset deviates substantially from the full randomization design and does not balance out the effects other than autoimmunity. So, a new problem arises. Despite the remarkable performance of the model on the sample data of MS, the features extracted by the model may not be related to the autoimmunity in MS and may simply be features from this group of patients that are shared in other ways. If the sample data has significant bias, the performance of the model trained on it cannot be guaranteed. For RA, the heatmap shows that the distribution of RA samples in the feature space is similar to that of healthy samples, and the feature differences between the two are not significant. Unlike the group of T1D samples, the mean number of TCR replicates for RA samples is only 1.0826, but the model does not perform well in distinguishing between RA and healthy samples, and the model does not effectively learn the representations associated with RA autoimmunity. The possible reason for this is that the information of the TCR group pool of cases in the RA group, especially the top 100 TCRs in abundance, is not sufficient to distinguish them from healthy individuals, and although some distributional similarities can be observed, there is still a need for more indepth study of the biological background of RA and exploration of its more discriminatory TCR characteristics to enhance the diagnostic ability of the model. For IAA, the sample size is small and the average number of repetitions of TCR is 1.5727. The heatmap shows the sparseness of the distribution of its feature points and also shows some slight feature differences between it and the healthy group. Nevertheless, the models can still effectively extract the IAA-specific TCR features with good performance, making the model effective in extracting some key feature information even with a limited sample size. However, to improve the robustness of the model and ensure its diagnostic accuracy, it is still necessary to increase the sample size, which can help improve the model’s ability to recognize IAA-specific features and increase its stability over a wider range of samples. Of course, it is undeniable that RA and IAA are themselves very complex autoimmune diseases. Their pathogenesis involves the interaction of multiple genetic, immune and environmental factors, which makes their prediction even more challenging. The greater complexity of their etiology relative to the two diseases, T1D and MS, may lead to greater difficulties for models in processing the data. Additionally, feature engineering may also play a role in performance degradation. The prediction of RA and IAA may require more complex and fine-grained feature engineering to capture disease-specific features and signals. Insufficient feature engineering or poor feature selection may result in models that do not fully utilize the available information, thus affecting performance.

In future studies, we can improve the prediction performance for RA and IAA as well as other autoimmune diseases in several ways. First, collect more samples of RA and IAA and other autoimmune disease datasets to balance the dataset. We can do this through multicenter collaborations, data sharing, and clinical trials. Second, study the pathogenesis of these autoimmune diseases in greater depth and carefully design and select features that better reflect their characteristics. Finally, try to use more advanced machine learning algorithms, as well as exploring the introduction of more complex model architectures or integrated approaches combining multiple models to better deal with complexity and data imbalance to improve predictive performance. Also, parameter tuning and cross-validation can be performed to optimize model performance. By taking these steps, we will hopefully further improve the predictive accuracy and application scope of the model and increase the diagnostic accuracy for autoimmune diseases.

Methods

The autoimmune disease dataset used in this study is derived from publicly available TCR-seq data from Adaptive Biotechnologies immuneACCESS online database (IA). IA serves as a genomics online database that focuses on storing and sharing immune receptor and immunome library data. The database integrates data from multiple research laboratories and programs and contains information on the immune system of humans and other species. In addition, our dataset of healthy human samples is derived from a study by Ying Xu et al. 39 . These data can be downloaded via the following link: https://github.com/Bioinformatics7181/DeepLION/tree/master/Data/THCA/TraininData . The above dataset covers four autoimmune diseases and a sample set of healthy controls, as detailed in Table 3 .

Table 3 Information on autoimmune disease datasets. Autoimmune diseases Sample size Data type RA 42 – 44 86 TCR-seq T1D 45 104 TCR-seq MS 46 174 TCR-seq IAA 27 65 TCR-seq Health 214 TCR-seq

Information on autoimmune disease datasets.

Data preprocessing

With the aim of effectively extracting features from TCR sequence data and improving the overall performance of the model, this study refers to the methods used by DeepCAT and DeepLION and performs the following steps in the data preprocessing stage:

Length screening: Sequences that are too short ( ) or too long ( ) are eliminated to facilitate processing of model inputs, while ensuring that the TCRs used are representative;

Special character handling: Sequences containing special characters (e.g. X, +, *,etc.) may affect the data quality due to factors such as technical or sequencing errors, so such sequences were removed away to improve the cleanliness and consistency of the dataset;

Adherence to ImMunoGeneTics (IMGT) nomenclature: Incomplete sequences that do not begin with cysteine (C) or end with phenylalanine (F) are dropped according to the IMGT standard 47 to maintain sequence consistency and integrity;

Variable locus processing: Sequences with unresolved variable loci were eliminated to ensure data integrity and reliability;

Sample Screening: After completing the above steps, eliminate samples with less than 10 TCR sequences. Make sure that each sample has a sufficient number of sequences.

We paid special attention to high-frequency TCR sequences because they may have important biological activities in the immune response. By screening the high-frequency sequences that ranked in the top 100 cloning frequencies, we focused on T cell clones that may be important for immune system function and the development of autoimmune diseases.

Multi-instance learning

Deep convolutional neural networks (CNNs) have shown significant potential in functional genomics and protein structure analysis 48 – 50 , which can recognize complex patterns, solve classification problems, and outperform traditional methods 29 . On the other hand, as a variant of Recurrent Neural Network (RNN), LSTM is also widely used for classification prediction of sequence data. LSTM can efficiently handle and learn the time-dependence in sequence data. Based on this, we trained two types of models: one is the AutoY model based on deep CNN, and the other is the LSTMY model based on bidirectional LSTM. Multi-Instance Learning (MIL), first proposed by Dietterich in a drug activity prediction study 51 , is a supervised learning method. And the method has proven its effectiveness in the field of cancer prediction 39 , 40 . In this study, we treat the training samples as bags of instances (bags) rather than individual instances. An instance bag contains multiple instances that represent an individual collection of TCR sequences, where each sequence in the bag is treated as an instance. Traditionally, single-instance learning methods that classify each individual TCR sequence may not fully utilize the inter-sequence information and may fall short in capturing the complex relationships between these sequences, leading to inaccurate predictions of disease status. Predicting whether a TCR repertoire is associated with autoimmune diseases can be described as a MIL problem, where each TCR sequence is treated as an instance and the TCR repertoire as an instance bag. According to the standard MIL assumption, each instance within a bag can be labeled as either positive (1) or negative (0), and the bag itself is labeled as positive (1) if it contains one or more positive instances. However, in our study, we do not simply apply this assumption. The specifics are as follows:

The bag collection is denoted by , where each represents the set of all instances within a bag, and is the corresponding label. The determination of the label is based on the entire set , rather than relying solely on individual instances within the bag. The core objective of MIL is to learn a classifier f that can accurately predict the label based on the set of instances within the bag.

The classifier f is defined as 39 :

where  denotes the k-th TCR feature matrix, and denotes the probability that the library has autoimmune disease, the activation function is a sigmoid function, and and are the weight matrix and bias of , respectively. The probability that a TCR repertoire is associated with an autoimmune disease can finally be derived, and the TCR repertoire is predicted to be associated with an autoimmune disease when is greater than 0.5, otherwise it is predicted to be a non-autoimmune disease. is not only a simple maximum or mean aggregation of TCR scores, but it is also capable of learning the interrelationships among TCRs and assigning appropriate weights to each TCR after the model has been trained, which improves the model’s prediction accuracy.

TCR encoding

In order to accurately reflect the biochemical properties of biomolecules, Beshnova et al. 38 used Principal Component Analysis (PCA) to generate a 20 15 feature matrix from 531 AA indices 52 to characterize the biochemical profile of amino acids. Considering that the Beshnova matrix was obtained from the largest number of AA indices and contained the most biochemical information (explaining more than 95% of the variation in the data), this was experimentally validated in methods such as DeepCAT and DeepLION, and showed good performance. Therefore, in this study, we used the Beshnova matrix (d = 15) to encode the amino acids and achieve vectorization of the input sequences.

CNN-based model

As shown in Fig. 8 , AutoY consists of an embedding layer, four convolutional modules, and an MLP module:

Embedding Layer. 20 different amino acids were encoded as 15-dimensional embedding vectors, and the amino acid sequences were encoded and transformed into tensors as input. The encoding was randomly initialized using a uniform distribution and padding was used to keep the length dimensionally constant. The weights of the convolutional and fully connected layers were initialized using Xavier 53 : 2 where is the weight parameter of the 1th layer and is the number of neurons in the 1th layer.

Convolutional Modules. 4 Convolutional modules consist of four different types of convolutional layers stacked twice. The AutoY model focuses on the high-frequency regions of lengths 2–4 and low-frequency regions of length 5 in the region of consecutive CDR3 residues in the CDR3 sequence, and uses different sizes and numbers of convolutional filters to process these regions. Specifically, these four convolutional layers have convolutional kernel sizes of 2, 3, 4, and 5, and the corresponding numbers of convolutional kernels or filters are 12, 12, 12, and 6. The step size is 1, and each convolutional layer is followed by a linear rectifier function with leakage (Leaky ReLU) 54 .

Convolutional layers. A Dropout layer 55 is added between each of the stacked convolutional layers for regularization to avoid overfitting by randomly discarding a portion of neurons during training, and the dropout rate of the Dropout layer is taken to be 0.5. A Batch Normalization layer is added between each of the convolutional layers to reduce the problem of vanishing gradients, which makes it easier to train the model parameters and allows for a larger learning rate to be used, helping the model to converge faster. It makes the parameters of the model easier to train, allows the use of a larger learning rate, helps the model converge faster, and improves the stability of the model. Global Maximum Pooling was applied to each of the four convolutional modules, aiming to select the maximum values from each feature mapping, which are considered to be the most likely features associated with autoimmune diseases.

MLP module. The outputs of the four convolutional modules are assembled into a vector by splicing, and then fed into the MLP module consisting of two fully connected layers (FC1, FC2) for processing. And Dropout layers with a dropout rate of 0.5 are added between FC1 and FC2 and after FC2. By applying sigmoid activation to the logits output of FC2, the probability of autoimmune disease is obtained.

AutoY framework diagram.

Model based on BiLSTM and attention

As shown in Fig. 9 , the LSTMY structure consists of an embedding layer, a BiLSTM module, an attention layer, and an MLP module consisting of two linear layers (FC1, FC2):

Embedding layer. 20 different amino acids were encoded as 15-dimensional embedding vectors, and the amino acid sequences were encoded and converted to tensor as input. The encoding was randomly initialised using a uniform distribution and padding was used to keep the length dimension constant. The length of each TCR sequence was recorded in order to populate the region with masks.

BiLSTM module. It is the core component of LSTMY and consists of three BiLSTM layers stacked together. Each of these BiLSTM layers consists of a combination of two LSTMs, one processing the input sequence in the forward direction and the other in the reverse direction. This structure enables the model to effectively capture the longterm dependencies in the sequences, and the internal state and gating mechanism introduced effectively mitigates the gradient vanishing or gradient explosion problem. Its hidden layer size is set to 30, the number of layers is 3, and dropout is added to prevent overfitting, with the dropout parameter set to 0.6.

Attention Layer. A Self-Attention mechanism is used to capture key features in the bidirectional LSTM layer more efficiently. Specifically, the core idea is to take into account the degree of association of each position in the sequence with all other positions when computing the attention weights.

MLP module. The above output is batch normalized by the Batch Normalization layer and fed to the MLP module consisting of two linear layers (FC1, FC2) and a Dropout layer is added between FC1 and FC2 and after FC2 with the Dropout parameter taken as 0.6. By applying sigmoid activation to the logits output of FC2, the probability of autoimmune disease is obtained.

LSTMY framework diagram.

Other settings

Using the Adam optimizer combining adaptive learning rate and momentum approach 56 :

where is the gradient at moment ; and are the exponential smoothing of the first- and second-order moments of the gradient , while and are the corresponding smoothing coefficients, with the default values used ( ); and is the model parameter, with being the learning rate ( ), and , which prevents the divisor from being 0.

The total number of training rounds for AutoY and LSTMY are set to 3000 and 2000, respectively.

An early-stopping mechanism is introduced, and in the experiments, the patience parameter for early-stopping is taken to be 300.

Metrics

In order to comprehensively evaluate the performance of the proposed model, we adopt the following key These evaluation metrics aim to comprehensively consider the performance of the model in different aspects.

Accuracy:

where denotes true positive cases, denotes true negative cases, denotes false positive cases, and denotes false negative cases.

Sensitivity, also known as true positive rate or recall, is defined as:

Specificity, which measures the proportion of true negative cases correctly identified, is given by:

Classification model performance is comprehensively evaluated using the AUC value under the ROC (Receiver Operating Characteristic) curve. This value reflects the model’s ability to distinguish between positive and negative cases under various thresholds. The closer the value of AUC is to 1, the better the model performs in categorizing positive and negative cases.

For the purpose of accurately evaluating the model, we choose iterated K-fold validation with shuffling of the data. This is done as follows: repeat the K-fold cross-validation n times, before dividing the data in each K-fold validation, the data are randomly disrupted and randomly divided into five equal parts, each time, four of them are selected as the training set, and the remaining one is used as the test set. Among them, this training set is then divided into training set and validation set in the ratio of 8:2, and they are input into the model as the final training data for training. And the average metric value of validations is taken as the final score:

here, and . The confidence level of the resulting model assessment metrics is reliable.

Experimental environment

For the experimental environment, we used Windows 10 operating system, Python version 3.10, CUDA version 12.2, and PyTorch version 2.0.1. For the hardware, we used an RTX 2080 for the GPU and an AMD R9 3900X for the CPU.

Conclusions

In this paper, in the context of how to effectively predict autoimmune diseases and contemporary advanced deep learning algorithms, after in-depth study of their pathogenesis, two deep neural network models are proposed, one is based on the CNN architecture named AutoY, and the other is based on the bidirectional LSTM integrated attention mechanism named LSTMY. And the sources of data, data processing, neural network architecture of the model, selection of key parameters of the model, and performance evaluation of the model in the thesis are described and analyzed in detail. According to the experimental results, the two models, AutoY and LSTMY, have notable performance in the prediction of the four autoimmune diseases, but it can also be seen that, due to the small sample size of RA and IAA in the dataset and their features are not significant enough, the performance of the models is not significant enough for prediction, which in fact reflects the efficient performance of our models in another way. In addition, through the analysis of the heatmap of the differences in sample characteristics, we explored the dataset in depth. Finally, we also actively discussed the other advantages and identified some causes as well as proposed some solutions, and these efforts are expected to provide effective tools and methods for the development of the fields of diagnosis and treatment of autoimmune diseases as well as the design of vaccines.

Author contributions

D.Y. contributed to Methodology, Software, and Writing—Original Draft. X.P. was involved in Investigation, Resources, and Visualization. S.Z. focused on Conceptualization. S.P. was responsible for Validation, Writing—Review & Editing, and Supervision.

Funding

This work was supported by the National Natural Science Foundation (Grant number 12461029).

Data availability

The data supporting the findings of this study are derived from previously reported studies and datasets, which have been cited. The processed data are available within the article. Raw data were accessed through the public domain and are available from the corresponding publications. Further inquiries about the datasets can be directed to the corresponding author. The code for this article can be downloaded at: https://github.com/solfix/AutoY.git.

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

Associated Data

Data Availability Statement

The data supporting the findings of this study are derived from previously reported studies and datasets, which have been cited. The processed data are available within the article. Raw data were accessed through the public domain and are available from the corresponding publications. Further inquiries about the datasets can be directed to the corresponding author. The code for this article can be downloaded at: https://github.com/solfix/AutoY.git.