Enhancing tumor T cell antigen prediction by integrating deep protein representations

Abstract

Background

Cancer is a major threat to human health, and cancer immunotherapy is considered the most promising treatment option. It offers high effectiveness and precision, with fewer side effects than traditional treatments. Tumor T cell antigens (TTCA) are the proteins or protein fragments that stand on the surface of cancer cells and that are recognized by the immune system. However, physical experimental approaches to infer them might be costly and time-consuming, even though they are important for cancer immunotherapy.

Results

We propose DeepTHybrid to predict TTCAs by combining chemically hand-designed features with deep protein representations from ProtBERT, an adapted version of the BERT large Language Model, for finding complex representations from the protein sequences. It is the first framework to incorporate these protein embeddings with complementary chemical features for this problem. After the feature extraction step, among multiple classifiers tested, we find Inception-based neural network to outperform the rest, so DeepTHybrid includes it. We evaluate the prediction performance on multiple datasets, where DeepTHybrid consistently outperforms the competing approaches across all datasets. For instance, on the first dataset, DeepTHybrid achieves an accuracy of 0.76, an F1-score of 0.78, outperforming existing machine learning and state-of-the-art TTCA predictors. Across both datasets, combining deep protein embeddings with chemical attributes yields superior performance compared to using either feature type alone. We also find the mixture of deep features and chemical features more informative in terms of Shapley-based explanation values.

Conclusions

Overall, our method is promising with its high accuracy rates and strong predictive skills. In addition to its predictive performance, DeepTHybrid provides a fast and scalable computational screening framework that can substantially reduce the cost and time associated with biological experiments by prioritizing high-confidence tumor T cell antigen candidates for downstream experimental validation.

Introduction

A report from the International Agency for Research on Cancer (IARC) indicates that in 2020, around 10 million people died from cancer, and 19.3 million new cases were diagnosed worldwide [1, 2]. Cancer has become the second leading cause of death. The main approaches in modern cancer drug therapy include tumor molecular targeted therapy, radiotherapy, and chemotherapy [3–5]. While traditional broad-spectrum anticancer drugs and radiotherapy are effective at killing tumor cells, they also harm normal cells, leading to significant adverse reactions and a higher risk of drug resistance [6–8]. However, advancements in equipment and immuno-oncology are driving a transformative shift in cancer treatment.

New cancer treatments are emerging, with targeted immunotherapy standing out as one of the most promising options. Unlike chemotherapy and radiotherapy, which often cause harmful side effects, immunotherapy has proven to be highly selective and effective, while minimizing side effects [9, 10]. This approach opens new avenues for developing potential cancer treatments. T cells are crucial in recognizing and eliminating tumor antigens presented on the surface by major histocompatibility complex (MHC) class I and class II molecules [11–13]. As a result, T cells play a significant role in tumor rejection and cancer immunotherapy. Accurately identifying T cell antigens not only enhances our understanding of their protective mechanisms but also aids in the development of highly effective cancer peptide vaccines [14]. TTCAs are key components of peptide-based cancer vaccines, where selected antigenic peptides are administered to stimulate tumor-specific T cell responses and enhance immune-mediated tumor clearance. In addition, the presence, diversity, and immunogenicity of tumor-associated antigens strongly influence patient responses to immune checkpoint inhibitors [15, 16], such as anti-PD-1/PD-L1 and anti-CTLA-4 therapies, by shaping T cell activation, clonal expansion, and immune surveillance within the tumor microenvironment. Consequently, accurate identification of high-confidence TTCAs is critical not only for understanding tumor-immune interactions, but also for prioritizing candidate epitopes for vaccine development, immunotherapy stratification, and personalized cancer treatment pipelines.

The emergence of deep learning has created new opportunities for studying and treating complicated diseases such as cancer in the quickly developing field of biomedical research. It has been applied to many bioinformatics problems, leading to state-of-the-art performance in many of them. Among these deep learning techniques, Large Language Models (LLMs) have recently shown superior performance in many natural language processing tasks [17, 18]. Some studies have focused on adapting those LLMs and machine learbing to protein and RNA sequences [19–24], leading to promising results on a number of benchmark datasets. On the other hand, many biological experiments are time-consuming and costly. So, replacing or speeding up biological experiments with faster and more accurate computational approaches has gained importance. Beyond tumor T-cell antigen prediction, related computational biology studies have shown that integrating heterogeneous molecular modalities can improve the modeling of complex biological mechanisms. For example, [25] demonstrates how epigenetic signals can be used to infer higher-order genome organization, while [26] further highlights the value of jointly modeling epigenomic and transcriptomic effects in spatial genome regulation. Although DeepTHybrid focuses on antigenicity prediction rather than 3D genome architecture, these studies support the broader motivation of our work: combining complementary biological feature sources, in our case, ProtBERT-derived contextual protein representations and chemically interpretable descriptors, can capture richer biological signals than single-modality models.

Despite recent advances in deep learning for protein sequence analysis, relying on a single feature paradigm, either deep representations or handcrafted chemical descriptors, poses inherent limitations for tumor T cell antigen (TTCA) prediction. Deep protein language models such as ProtBERT excel at capturing long-range dependencies, contextual semantics, and implicit sequence patterns learned from large-scale unlabeled protein corpora. However, these embeddings are inherently high-dimensional and abstract, often lacking direct correspondence to well-understood physicochemical properties that are known to influence antigen processing, MHC binding, and T cell recognition. As a result, models based solely on deep embeddings may obscure biologically interpretable signals and may be sensitive to dataset-specific noise or annotation bias.

Conversely, traditional chemical and physicochemical descriptors, such as amino acid composition, charge distribution, hydrophobicity, and sequence-order statistics, offer strong biological interpretability and have been widely used in TTCA prediction. These handcrafted features encode explicit biochemical constraints relevant to antigenicity, but they are limited in their ability to capture higher-order dependencies, long-range interactions, and complex contextual patterns across protein sequences. Consequently, models based only on chemical features may fail to fully represent the rich semantic structure underlying tumor antigens.

To address these complementary limitations, we adopt a hybrid representation strategy that integrates deep contextual embeddings from ProtBERT with multiple families of chemically interpretable features. We propose DeepTHybrid to predict Tumor T Cell antigens by combining deep feature embeddings with efficient Inception-based neural network. Our method is a hybrid approach in the sense that it combines deep feature representations with hand-designed chemical features. Tumor T cell antigens are proteins or protein fragments that are expressed on cancer cells and that the immune system recognizes. Identifying tumor T cell antigens is important to understand the cancer dynamics, especially when developing treatment approaches [27]. In-depth, when obtaining deep feature representations, our method focuses on ProtBERT [28], a transformer model inspired by BERT, which is designed to extract important properties from protein sequences. Very recent advancements in natural language processing have shown that transformers are very good at capturing complex patterns and relations in sequential data. ProtBERT, customized from transformers to explore protein sequences, deciphers the complex biological data and provides insights into how these sequences affect the way tumors interact with the immune system. In addition, integrating Inception Module-inspired [29] CNN improves the inference accuracy, while utilizing the extracted features better.

Although deep learning has so far been successful in several bioinformatics tasks, its application in predicting tumor T cell antigens has been slower. In our case, when we encode for chemical attributes, we use 4 types of chemical attribute encoding approaches, which are Pseudo Amino Acid Composition (PAAC), Adaptive Skip Dipeptide Composition (ASDC), Grouped Amino Acid and Peptide Composition (GAAPC), and Global Protein Sequence Descriptors (GPSD). Our datasets are imbalanced, so we apply an oversampling procedure, SMOTE. Afterwards, we use the Maximum Relevance Maximum Distance (MRMD) method to select a subset of important attributes. For the two considered datasets, we obtain 0.76, 0.81 accuracies and 0.78, 0.83 F1 scores respectively. Our methods outperform classical machine learning models such as Random Forests and Gradient Boosting Machines [30] as well as competing approaches. As a result, deep learning-based methods and deep learning-based feature representations to identify Tumor T Cell antigens are quite promising and motivating. An advancement in the field of cancer immunotherapy was made possible by the implementation of highly accurate methodologies, which produced encouraging findings in the identification and classification of tumor T-cell antigens. Our methodology not only exhibits strong predictive powers and great accuracy rates, but also opens up new avenues for fighting tumors. Furthermore, we systematically demonstrate—through extensive ablation studies and SHAP-based interpretability analysis—that deep contextual embeddings and chemical attributes provide complementary and non-redundant biological signals, leading to consistent performance gains across datasets. Together, these contributions establish DeepTHybrid as a principled and explainable deep-learning framework that extends the state of the art in tumor T-cell antigen prediction.

While prior studies combine handcrafted physicochemical features with conventional machine-learning classifiers (e.g., RF, SVM, ensemble stacking), DeepTHybrid differs methodologically and conceptually in several key aspects:

• DeepTHybrid is the first framework to incorporate large-scale protein language model embeddings (ProtBERT) for tumor T-cell antigen prediction. These embeddings capture long-range contextual and semantic dependencies in protein sequences that cannot be modeled by classical handcrafted features alone.

• Hybrid fusion of deep contextual embeddings with chemically interpretable features. Unlike prior hybrid methods that fuse multiple handcrafted descriptors, DeepTHybrid explicitly fuses deep contextual representations with four complementary chemical feature families.

• Inception-based multi-scale learning over protein embeddings. DeepTHybrid introduces an Inception-inspired CNN architecture that processes ProtBERT embeddings at multiple receptive field scales, enabling the model to capture both local motif-level and broader contextual patterns in protein sequences.

• DeepTHybrid uniquely provides SHAP-based interpretability, demonstrating that deep protein embeddings and chemical attributes jointly contribute to predictions.

The paper is organized as follows: The related methods brief survey is provided in Section Related work. Section Materials and methods will give details about DeepTHybrid, chemical features component, ProtBERT component for obtaining deep representations, and the rest of the components. The details about the experimental evaluation are described in Section Experimental setup. Finally, our results are discussed in Section Results and discussion, and the conclusion is discussed in Section Conclusion and future work.

Related work

A comprehensive and up-to-date overview of this rapidly growing field is provided by Shoombuatong et al. [31], who systematically reviewed and benchmarked machine-learning-based TTCA identification methods. Their study categorizes existing predictors into classical ML, ensemble learning, and deep learning paradigms, highlighting common design choices, datasets, and evaluation protocols. Importantly, the review underscores two key challenges that remain insufficiently addressed: 1- the limited use of large-scale protein language models for antigen representation, and 2- the lack of transparent interpretability mechanisms that connect model predictions to biologically meaningful sequence features.

There are a number of studies focusing on specifically tumor T cell antigen prediction. Among them, iTTCA-RF [32] utilizes random forest for predicting tumor T cell antigens. In their study, 4 feature extraction methods were used. After feature extraction, a two-step feature selection technique is employed to find an optimal feature subset. Finally, Random Forest algorithm is used to build the final prediction model. In another paper, [33] extends this by providing a new computational model known as iTTCA-MFF, which incorporates the physicochemical (PC) properties of amino acids and the residue pairwise energy content matrix to help in providing a more effective representation of the peptide sequences. Discriminative data are extracted using covariance and Pearson’s correlation coefficient (PCC), and the best features are selected using the Least Absolute Shrinkage and Selection Operator (LASSO). The best features are then input into a support vector machine for the detection of Tumor T Cell antigens, thus providing a superior and more accurate model than its predecessor, the iTTCA-RF.

PSRTTCA [34] is a more recently developed method where they introduce a new learning method to improve the detection and characterization of Tumor T Cell antigens (TTCAs) through propensity score representation learning. PSRTTCA can produce various sets of propensity scores for TTCAs in amino acids, dipeptides, and g-gap dipeptides. What brings the best sets of such variant propensity scores to the final meta-predictor is a big advantage and helps PSRTTCA significantly to have a much enhanced prediction performance. This method not only brings a higher level of accuracy, but it also helps in a better understanding of the TTCA properties using PSR-derived amino acid propensities.

Next, StackTTCA method [35] is proposed as an advanced stacking ensemble learning-based architecture for accurate and high-throughput identification of TTCAs. StackTTCA starts by using 13 major machine learning methods, along with 12 different feature encoding schemes to make a total of 156 individual baseline models. These individual baseline models are then used to make a new feature vector. However this new feature vector is a probabilistic vector, so it is further refined by applying a feature selection procedure to obtain the best probabilistic feature vector. The final stacking model formed using the refined features has shown substantially better performance in independent test sets. In summary, Table 1 compares DeepTHybrid with the competing approaches in terms of feature types, model details, interpretability.

Table 1.

Comparison of DeepTHybrid with representative state-of-the-art TTCA prediction methods

Method	Feature	Model	Deep protein	Interpretability
	Types		Representations
iTTCA-RF	Physicochemical descriptors	Random Forest	×	Feature importance
PSRTTCA	Propensity-score features	SVM-based ensemble	×	Limited (score analysis)
StackTTCA	Handcrafted features (multiple encodings)	Stacking ensemble	×	×
DeepTHybrid (Ours)	Chemical + contextual embeddings	Inception-based CNN	\checkmark (ProtBERT)	\checkmark (SHAP)

Among other methods, iTTCA-MVL [36] (improved Tumor T-Cell Antigen prediction via Multi-View Learning) is a computational framework designed to identify tumor-associated T-cell antigens by integrating heterogeneous biological feature representations. It encodes peptide or protein sequences from multiple complementary perspectives—such as physicochemical descriptors, sequence-order information, and evolutionary or compositional features—and fuses them within a multi-view learning architecture to capture both local and global antigenic patterns. iTTCA-MVL jointly learns from these diverse views rather than relying on a single feature space. Similarly, TTCA-IF [37] (Tumor T-Cell Antigen prediction via Information Fusion) is a machine-learning framework that identifies tumor T-cell antigens by integrating multiple complementary feature representations extracted from protein or peptide sequences. Instead of relying on a single descriptor, TTCA-IF combines diverse information sources, such as amino-acid composition, physicochemical properties, sequence-order features, and evolutionary signals through an explicit information-fusion strategy to construct a unified representation. Another method is Sa-TTCA [38] (Self-attention–based Tumor T-Cell Antigen prediction), which explicitly models long-range dependencies and contextual interactions within antigen sequences. It employs a self-attention mechanism to automatically learn which amino-acid positions and sequence patterns contribute most strongly to antigenicity, enabling the model to capture global sequence semantics beyond local motifs. By adaptively weighting informative residues and suppressing irrelevant noise, Sa-TTCA improves feature representation quality and interpretability.

Materials and methods

Dataset collection

To comprehensively evaluate the robustness and generalization ability of the proposed DeepTHybrid, we conducted experiments on two independently constructed tumor T cell antigen (TTCA) datasets that differ in data curation strategy, class balance, and sequence-length distribution.

The first dataset (Dataset 1) was obtained from Charoenkwan et al. [39], which has been widely adopted as a benchmark in prior TTCA prediction studies. This dataset contains a total of 985 protein sequences, comprising 592 experimentally validated TTCAs (positive samples) and 393 non-TTCAs (negative samples). As noted in subsequent studies [40], this dataset exhibits class imbalance and may include a limited number of ambiguous or weakly characterized negative samples due to historical annotation practices. The peptide lengths in Dataset 1 range from 8 to 50 amino acids. Owing to its imbalanced nature, Dataset 1 represents a more challenging and noisy prediction scenario; therefore, oversampling techniques are applied during training to mitigate class imbalance effects.

The second dataset (Dataset 2) was taken from Herrera-Bravo et al. [40], who re-curated the original TTCA benchmark by removing potentially incorrect negative samples and enforcing stricter annotation criteria. As a result, Dataset 2 is fully balanced, consisting of 592 TTCAs and 592 non-TTCAs. In addition, this dataset is restricted to shorter peptide sequences, with lengths ranging from 8 to 20 amino acids, yielding a more homogeneous input distribution. Because Dataset  is already balanced and less noisy, no data resampling techniques are applied during training.

For both datasets, we follow established evaluation protocols to ensure fair comparison with prior work. In Dataset 1, the training set contains 470 TTCAs and 318 non-TTCAs, while the test set includes 122 TTCAs and 75 non-TTCAs. For Dataset 2, the training set is composed of 474 TTCAs and 474 non-TTCAs, and the remaining 118 TTCAs and 118 non-TTCAs are used as an independent test set. Test samples are strictly held out and are not used during feature selection, model tuning, or validation.

Although both datasets share the same set of positive TTCA samples, they differ substantially in negative sample construction, class balance, and sequence-length characteristics. Dataset 1 reflects a more realistic and noisy setting with imbalanced classes, whereas Dataset 2 provides a corrected and balanced benchmark with cleaner annotations. Evaluating DeepTHybrid on both datasets allows us to assess its robustness across heterogeneous data conditions and to demonstrate consistent performance improvements under both challenging and idealized scenarios (Table 2).

Table 2.

The summary of datasets

Species	Positive	Negative	Min seq	Max seq
Dataset 1: Charoenkwan et. al. [39]	592	393	8	50
Dataset 2: Herrera-Bravo et. al. [40]	592	592	8	20

Non-deep feature representations

The extracted set of attributes quality will have a huge impact on the method’s performance. A number of strategies which convert biological sequences into quantitative vector representations have been suggested [41–44]. Here, we will be numerically encoding the protein sequence’s attributes via 4 approaches which are discussed next. We generated all these four types of numerical embeddings via iLearn tool package [42].

Global protein sequence descriptor (GPSD)

This approach details out the given protein’s amino acids properties globally, and creates 188 numerical attributes which combine sequence knowledge with amino acid qualities [45, 46]. Generally, Global Protein Sequence Descriptor (GPSD) will present its representation via two components. The first component decomposes the amino acids into numerical representations. The initial twenty attributes are each natural amino acid’s frequency in the protein. The second component has 168 attributes which are associated with amino acids 8 physicochemical attributes. The detailed knowledge about these physicochemical properties can be found in [47–49]. For each of these eight physicochemical properties, the 20 amino acids are partitioned into three groups according to predefined thresholds. Given a protein sequence of length L, the CTD descriptor consists of the following three components [50]:

Composition (C): The composition describes the fraction of residues belonging to each group:

1

where denotes the number of residues in group .

Transition (T): Transition measures the frequency of transitions between different groups along the sequence:

2

where is the number of times a residue from is immediately followed by a residue from .

Distribution (D): Distribution captures the relative sequence positions where group-specific residues appear. For each group , five positional descriptors are computed, corresponding to the first occurrence and the positions where 25%, 50%, 75%, and 100% of residues of that group are observed:

3

where denotes the sequence index at which the k-th percentile of group is reached.

In CTD decomposition, the C part returns 3-dimensional features where it counts each group’s occurrence frequencies. T part returns 3-dimensional features which represent frequencies of transition among 3 groups. Lastly, the D part models the first, 25%, 50%, 75%, and the final occurrences of a group in the protein sequence, which in total makes up 15 dimensions. Overall, CTD model generates (3 + 3 + 15) * 8 = 168 attributes.

Grouped amino acid and peptide composition (GAAPC)

In grouped amino acid and peptide composition, we divide each amino acid into 5 groups in terms of their physicochemical qualities, such as charge, hydrophobicity, and molecular size. We name these 5 groups from c1 to c5. Aliphatic group c1 is composed of amino acids with the letters GAVLMI. Aromatic group c2 is composed of amino acids with the letters FYW. Positively charged group c3 is composed of amino acids with letters KRH. Negatively charged group c4 is composed of amino acids with the letters DE. Lastly, uncharged group c5 is composed of amino acids with the letters STCPNQ. Let L be the protein sequence length, we can encode a protein sequence as follows:

Let N(c) represent the number of amino acids in group c, and N(i) define the number of amino acids of type i. The grouped amino acid composition (GAAC) [42, 51] descriptor can be defined as:

4

5

Let be the number of dipeptide proteins defined by amino acid type groups x and y. The grouped dipeptide composition (GDPC) [42, 51] encoding is a variant of the dipeptide composition descriptor. GDPC is defined as:

6

Let be the number of tripeptide proteins defined by amino acid type groups x, y, and z. The grouped tripeptide composition (GTPC) [42, 51] encoding is a variant of the tripeptide composition descriptor. GTPC is defined as:

7

As a result, GAAPC is a union of the three descriptors GAAC, GDPC, and GTPC discussed above. Overall, GAAPC will generate 155 dimensional attribute vector.

Adaptive skip dipeptide composition (ASDC)

Adaptive skip dipeptide composition has first been introduced in [52]. This composition approach can be seen as another variation of the dipeptide composition above, where it not only takes into account the appropriate knowledge among neighbouring residues, but also considers the knowledge of the interfering residues [53]. Let define all potential dipeptide proteins occurrence frequency where the number of interfering amino acids is less than N-1. ASDC converts the sequence into a 400-dimensional vector, and it is defined as:

8

9

Pseudo amino acid composition (PAAC)

Pseudo-amino acid composition descriptor is frequently utilized in drug target prediction, drug development, and protein attribute prediction [54]. Unlike conventional amino acid composition, PAAC incorporates sequence-order information by introducing correlation factors between residues separated by a certain distance along the sequence [55]. We used iLearn’s default parameters to obtain a 22-dimensional attribute vector.

Let denote the normalized occurrence frequency of the i-th amino acid (i = 1,… ,20). The PAAC vector is defined as:

10

where λ denotes the maximum sequence-order correlation tier.

The first 20 components represent the normalized amino acid composition:

11

where w is a weighting factor controlling the influence of sequence-order effects.

The remaining λ components encode sequence-order correlation factors:

12

where is defined as:

13

with d(· ) representing the squared Euclidean distance between physicochemical property vectors of residues at positions i and i+k.

Attribute selection

The attributes composed from a sequence include noise when they are used in predictive modeling. We enhance the quality of representing the attributes by adopting an attribute selection strategy. These attribute selection strategies help remove redundant knowledge and remove the possibility of overfitting. A number of methods exist in analyzing the attributes where some of them are Maximum Relevance Maximum Distance (MRMD) [56, 57], Minimal Redundancy Maximal Relevance (MRMR) [58, 59], and Analysis of Variance (ANOVA) [60]. Those approaches have been frequently utilized in prediction problems covering proteins, RNAs, and DNAs.

In this study, we used MRMD to choose an optimal subset of attributes for our deep learning model. MRMD attribute selection methodology is composed of two parts. The beginning part uses the Pearson correlation coefficient to find the correlation between the attribute and the target class vector. The next part calculates the redundancy between attributes according to distance formulas such as the Tanimoto coefficient, Cosine distance, or Euclidean distance. The higher correlation coefficient corresponds to a stronger relationship between the attribute and the class label, as well as higher distance corresponds to a smaller redundancy between the attributes. Overall, the MRMD selection methodology chooses a subset of attributes which are both correlated highly with the class label as well as exhibit lower redundancy among attributes. In our case, we sorted the original set of attributes in terms of MRMD attribute ranking algorithm. Afterwards, we run IFS strategy in selecting an optimal subset of attributes.

Strategy for handling imbalanced datasets

Imbalanced dataset is a common issue in bioinformatics problems. A number of techniques have been introduced to deal with imbalanced datasets. The strategy mainly either increases the number of samples in the minority class or deletes the number of samples in the majority class. The strategies can be partitioned into 3 groups: 1- Oversampling, 2- Undersampling, and 3- Hybrid Sampling [61–63]. In our study, for dataset 1, we use the oversampling approach Synthetic Minority Over-sampling TEchnique (SMOTE) [64] to deal with imbalanced datasets by preventing overfitting. Dataset 2 is already balanced, so we do not apply SMOTE when we run DeepTHybrid for dataset 2.

Deep feature representations via ProtBERT

Word embedding techniques are valuable tools for examining the strength of associations between words and their surrounding context within a fixed window size. Unlike conventional static embedding models like Word2vec [65] and GloVe [66], modern models such as BERT [67] overcome limitations related to one-directional training. BERT leverages unlabeled text data by learning bidirectionally across all its layers during pretraining. This is accomplished by incorporating both forward and backward contextual information. As a result, the embeddings produced by BERT are context-sensitive (dynamic), meaning that the same word can yield different vector representations depending on its usage within a sentence. BERT has demonstrated strong performance on various natural language processing tasks, including textual entailment and question answering [68], often without requiring extensive task-specific adjustments.

Several versions of BERT have been pretrained with different architectural setups, varying in terms of the number of parameters, layers, attention heads, and hidden units. One notable variant is ProtBERT [69], a protein-specific adaptation of BERT trained on the UniRef100 dataset. ProtBERT extends the original BERT architecture by increasing the number of transformer layers from 12 to 30. Initially, it was trained for 300,000 steps using sequences up to 512 tokens long. Following this, it underwent an additional 100,000 training steps on sequences with lengths of up to 2,000 tokens. This two-phase training strategy enables ProtBERT to efficiently handle shorter sequences with larger batch sizes, which in turn supports more effective training on longer sequences. For our experiments, we utilize only the final layer output of ProtBERT, as empirical evidence suggests minimal performance variation when comparing the final layer to the top four layers combined. After the model is trained, each amino acid in a sequence is represented as a contextual embedding vector with 1024 dimensions. Specifically, in dataset 1, each protein sequence of length 50 is transformed into a matrix containing 50 vectors, each of size 1024. If a sequence contains fewer than 50 amino acids, the matrix is padded with zeros to maintain uniformity. Likewise, dataset 2 is represented as a 20 × 1024 matrix.

Inception-based 2D CNN

We process ProtBERT embeddings as well as chemical features via Inception-based module as shown in Fig. 1. The Inception-based neural network employed in DeepTHybrid is designed to capture multi-scale spatial patterns from ProtBERT-derived protein embeddings. Each protein sequence is represented as a two-dimensional matrix of size L × 1024, where L denotes the padded sequence length 11. This representation is treated as a single-channel input and processed by an Inception-inspired module composed of multiple parallel convolutional branches.

Fig. 1.

DeepTHybrid architecture

Specifically, the Inception module consists of parallel 2D convolutional layers with kernel sizes 1 × 1, 3 × 3, and 5 × 5, each followed by ReLU activation and batch normalization. These parallel branches enable the model to simultaneously capture local motif-level patterns and broader contextual dependencies within protein sequences [70, 71]. The resulting feature maps from each branch are concatenated along the channel dimension to form a unified multi-scale representation.

Following the Inception module, the concatenated feature maps are flattened and combined with the selected chemical feature vector obtained after MRMD feature selection. This late-fusion strategy preserves the interpretability of physicochemical descriptors while enriching them with deep contextual embeddings. The fused feature vector is then passed through two fully connected layers with 64 and 2 neurons, respectively. Dropout regularization is applied between dense layers to reduce the risk of overfitting. Our overall architecture is summarized in Fig. 2.

Fig. 2.

The whole framework of the proposed DeepTHybrid to identify tumor T Cell Antigen

Experimental setup

Baseline machine learning and deep learning classifiers

We have experimented with 2 well-known machine learning (ML) classifiers as our baseline: Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The main reasons for choosing Random Forest [72] and XGBoost [73] algorithms as baseline models are the high performance of these models, their ability to determine the importance of features, and their ability to efficiently model complex biological interactions. Both Random Forest and XGBoost are ensemble-based techniques. We have also experimented with a deep learning baseline, 2D Convolutional Neural Networks. We evaluate the performance via 5-fold cross-validation (CV), where grid search is used to tune model hyperparameters.

To achieve optimal performance, both deep learning and machine learning models require tuning of hyperparameters. In our experiments, we performed a grid search for the Random Forest model by varying the maximum depth over the values [3, 5, 10, 20], the minimum number of samples required at a leaf node over [1, 2, 4], the minimum number of samples required to split a node over [2, 5, 10], and the number of estimators over [100, 200, 400]. For the XGBoost model, a grid search was also conducted to identify the best combination of hyperparameters. Specifically, we explored values for the maximum depth as [1, 2, 4, 6], the minimum child weight as [1, 3, 5], and the gamma parameter as [0, 0.1, 0.2, 0.3, 0.4, 0.5]. In the case of Convolutional Neural Network (CNN) models, we optimized several hyperparameters, including the learning rate (chosen from , , , ), the number of training epochs (ranging from 30 to 100 in increments of 10), the kernel size (2 × 2 or 3 × 3), and the number of convolutional filters (16, 32, or 64).

For training baseline machine learning models, when ProtBERT embeddings were not used, each nucleotide was numerically encoded through one-hot encoding. In this scheme, each of the 20 amino acids in a protein sequence is represented by a unique 20 × 1 binary vector, ranging from [1, 0, … , 0] to [0, 0, … , 1]. Following this encoding, the presence or absence of tumor T cell antigens was denoted using binary labels: sequences with antigens were labeled as 1, and those without as 0.

In terms of competing approaches, we use 3 state-of-the-art classifiers for tumor T cell antigen prediction which are iTTCA-RF [32], PSRTTCA [34], StackTTCA [35] by using their default parameters. We use the code provided for iTTCA-RF. We used the provided web servers for PSRTTCA and StackTTCA where only binary classification results are returned without any probability values. So, we could not calculate some metrics requiring probabilities for performance evaluation, such as AUC.

Performance validation and overfitting control

For each dataset, we first construct a fixed train-test split consistent with prior TTCA benchmarks. The test set is held out entirely and is never used during feature selection, hyperparameter tuning, or model selection. All reported test-set results, therefore, reflect genuine out-of-sample generalization [74–76].

Within the training set only, we perform 5-fold cross-validation to optimize all model hyperparameters. For classical machine learning baselines (Random Forest and XGBoost), this includes tree depth, number of estimators, and regularization parameters. For deep learning models, including the proposed DeepTHybrid, we tune learning rates, convolutional filter sizes, number of filters, and training epochs. This nested validation protocol prevents information leakage between training and testing stages [77, 78].

Several complementary mechanisms are employed to mitigate overfitting. At the model level, dropout regularization is applied in the fully connected layers of the Inception-based network, and early stopping based on validation loss is used to halt training once convergence is reached. At the feature level, Maximum Relevance Maximum Distance (MRMD) feature selection is applied exclusively within training folds to reduce redundancy and noise in the chemical feature space.

Performance evaluation

We assess and compare the performance of the methods by using the following well-known classification metrics:

14

15

16

17

where TP, TN, FP, and FN mean true positive, true negative, false positive, and false negative sample counts, respectively. Let be the true positive rate and be the false positive rate, ROC curve plots the true positive rate against the false positive rate at various thresholds. We report Area Under the Receiver Operating Characteristic Curve (AUC) which summarizes the relationship between these two metrics in terms of performance in a threshold-independent manner. A method’s capabilities in separating the positive classes from negative classes are revealed by AUC, where a more effective classification method has a greater AUC score.

Results and discussion

DeepTHybrid outperforms the baseline methods

First, we evaluated the baseline methods with either only ProtBERT protein representations or combined with chemical attributes (hybrid) in order to analyze the importance of a subset of feature types. DeepTHybrid is compared with 2D CNN, Random Forest, and XGBoost methods as seen in Tables 3 and 4 for Charoenkwan et al. [39] (Dataset 1) and Herrera-Bravo et al. [40] (Dataset 2) respectively. According to Table 3, hybrid DeepTHybrid which combines ProtBERT embeddings, Chemical attributes, and Inception neural network outperforms the rest of the combinations in all metrics. For instance, our hybrid DeepTHybrid can obtain 0.76 accuracy, 0.78 F1 score, and 0.78 AUC outperforming the rest. Among the remaining scenarios, hybrid methods also perform better. For instance, in terms of accuracy and F1 score, the second best-performing method is ProtBERT + Chemical + XGBoost. In this case, even though ProtBERT + Chemical + 2D CNN appears to perform lower than the one with the XGBoost, the performance difference is small. In general, hybrid methods outperform their corresponding non-hybrid ones that use only ProtBERT embeddings. We obtain similar results for Dataset 2 as seen in Table 4. According to Table 4, hybrid DeepTHybrid combining ProtBERT embeddings, Chemical attributes, and the Inception neural network again outperforms the rest of the combinations in all metrics. However, in this case, the second best-performing method is the hybrid one using 2D CNN instead of ProtBERT + Chemical + XGBoost. Similar to Dataset 1, hybrid methods outperform their corresponding non-hybrid ones that use only ProtBERT embeddings in Dataset 2.

Table 3.

The performance of our method compared with baselines over the dataset taken from Charoenkwan et al. [39] (Dataset 1)

Method	Accuracy	Precision	Recall	F1	AUC
ProtBERT + RF	0.68	0.70	0.69	0.69	0.69
ProtBERT + XGBoost	0.72	0.71	0.73	0.72	0.74
ProtBERT + 2D CNN	0.73	0.73	0.75	0.74	0.76
ProtBERT + Inception	0.74	0.78	0.76	0.77	0.76
ProtBERT + Chemical + RF	0.70	0.72	0.69	0.70	0.71
ProtBERT + Chemical + XGBoost	0.74	0.75	0.75	0.75	0.76
ProtBERT + Chemical + 2D CNN	0.74	0.74	0.75	0.74	0.75
DeepTHybrid: ProtBERT + Chemical + Inception	0.76	0.80	0.76	0.78	0.78

Bold values represent the methods with the best performance in each metric

Table 4.

The performance of our method compared with baselines over the dataset taken from Herrera-Bravo et al. [40] (Dataset 2)

Method	Accuracy	Precision	Recall	F1	AUC
ProtBERT + RF	0.73	0.74	0.73	0.73	0.72
ProtBERT + XGBoost	0.77	0.76	0.77	0.76	0.78
ProtBERT + 2D CNN	0.78	0.78	0.78	0.78	0.80
ProtBERT + Inception	0.79	0.83	0.80	0.81	0.80
ProtBERT + Chemical + RF	0.75	0.77	0.73	0.75	0.75
ProtBERT + Chemical + XGBoost	0.79	0.79	0.80	0.79	0.80
ProtBERT + Chemical + 2D CNN	0.79	0.79	0.79	0.79	0.80
DeepTHybrid: ProtBERT + Chemical + Inception	0.81	0.85	0.81	0.83	0.83

Bold values represent the methods with the best performance in each metric

Comparison with the competing approaches

We evaluate the performance of our DeepTHybrid with 3 existing competing approaches. These methods are: iTTCA-RF [32], PSRTTCA [34], StackTTCA [35]. As seen in Table 5 for Dataset 1, DeepTHybrid outperforms the competing approaches in almost all metrics except one of them. We do not report AUC scores for PSRTTCA and StackTTCA since only binary classification results are returned by them. We obtain similar outperformance of DeepTHybrid in Dataset 2 which can be seen in Table 6. Our method’s outperformance in both tables could be mainly explained by its ability to integrate deep protein representations together with hand-designed chemical attributes. Using Inception-based neural network to process embeddings was also another factor for our method’s outperformance. None of the competing approaches use deep protein representations, but instead, they either use chemical attributes or one-hot encoding to process protein sequence knowledge. Competing approaches also use traditional machine learning methods instead of deep learning methods such as Inception neural network to make classifications from the input. With the help of deep learning, we have observed that, CNN has a better predictive capability compared to other machine learning algorithms in the area of predicting tumor T cell antigens.

Table 5.

Our method with the competing methods over dataset taken from Charoenkwan et al. [39] (Dataset 1)

Method	Accuracy	Precision	Recall	F1	AUC
iTTCA-RF	0.72	0.70	0.68	0.69	0.74
PSRTTCA	0.70	0.70	0.72	0.71	–
StackTTCA	0.74	0.85	0.69	0.77	–
DeepTHybrid	0.76	0.80	0.76	0.78	0.78

Bold values represent the methods with the best performance in each metric

Table 6.

Our method with the competing methods over dataset taken from Herrera-Bravo et al. [40] (Dataset 2)

Method	Accuracy	Precision	Recall	F1	AUC
iTTCA-RF	0.72	0.75	0.77	0.76	0.79
PSRTTCA	0.75	0.75	0.77	0.76	–
StackTTCA	0.78	0.9	0.7	0.8	–
DeepTHybrid	0.81	0.85	0.81	0.83	0.83

Bold values represent the methods with the best performance in each metric

The impact of protein embedding approach

While DeepTHybrid employs ProtBERT as its default deep protein representation, recent advances in protein language modeling have produced several alternative pretrained architectures that differ in model size, training objective, and representational bias. To assess whether the performance gains of DeepTHybrid are specific to ProtBERT or generalize across different protein embedding models, we conduct an ablation study by replacing ProtBERT with four widely used pretrained protein language models: ESM-2 [79], ProtT5 [69], ProtElectra [69], and ProtAlbert [69]. In this analysis, all experimental conditions are kept fixed except for the protein embedding backbone. Specifically, the same chemical feature set selected via MRMD, the same Inception-based CNN architecture, identical training–test splits, and identical hyperparameter tuning procedures are used across all embedding variants. This controlled setup allows us to isolate the contribution of the protein embedding approach to the overall predictive performance.

Table 7 reports the comparative results on Dataset 1. We observe that DeepTHybrid exhibits consistently strong performance across all embedding backbones, indicating that the hybrid framework is robust to the choice of pretrained protein language model. Among the tested embeddings, ProtBERT and ESM-2 achieve the best overall performance, with ProtBERT yielding the highest F1 and AUC scores. This result can be attributed to ProtBERT’s deep bidirectional transformer architecture and its large-scale pretraining on UniRef100, which enables it to capture rich contextual and long-range dependencies relevant to antigenicity. ESM-2 and ProtT5 perform competitively, demonstrating that encoder-only (ESM-2) and encoder–decoder (ProtT5) architectures both provide informative representations for TTCA prediction. However, models with lighter parameterization and alternative pretraining objectives, such as ProtElectra and ProtAlbert, show a modest drop in performance. This suggests that while these models capture local sequence patterns effectively, they may encode less global semantic information compared to larger transformer models.

Table 7.

Performance comparison of DeepTHybrid using different protein embedding models on Dataset 1

Protein embedding	Accuracy	Precision	Recall	F1	AUC
ProtAlbert + Chemical + Inception	0.73	0.74	0.73	0.72	0.74
ProtElectra + Chemical + Inception	0.73	0.75	0.72	0.73	0.74
ProtT5 + Chemical + Inception	0.75	0.79	0.76	0.77	0.78
ESM-2 + Chemical + Inception	0.75	0.79	0.76	0.77	0.77
D eep T Hybrid	0.76	0.80	0.76	0.78	0.78

All models use identical chemical features, MRMD selection, and Inception-based CNN classifier

Overall, these findings indicate that the performance gains of DeepTHybrid do not rely on a single embedding model, but are instead driven by the principled integration of deep contextual protein representations with chemically interpretable features. Nevertheless, larger and more expressive protein language models tend to yield superior results, highlighting the importance of embedding capacity and pretraining strategy in antigen prediction tasks Table 8.

Table 8.

The impact of distinct categories of chemical features on the classification performance over the dataset taken from Charoenkwan et al. [39] (Dataset 1)

Method	Accuracy	Precision	Recall	F1	AUC
1 + 2	0.69	0.73	0.74	0.73	0.75
1 + 3	0.70	0.70	0.74	0.72	0.72
1 + 4	0.70	0.70	0.72	0.71	0.72
2 + 3	0.72	0.71	0.73	0.72	0.74
2 + 4	0.72	0.72	0.74	0.73	0.75
3 + 4	0.73	0.73	0.75	0.74	0.75
1 + 2 + 3	0.72	0.72	0.74	0.73	0.75
1 + 2 + 4	0.74	0.77	0.75	0.76	0.76
1 + 3 + 4	0.75	0.78	0.78	0.78	0.77
2 + 3 + 4	0.74	0.82	0.74	0.78	0.78
1 + 2 + 3 + 4 (DeepTHybrid)	0.76	0.80	0.76	0.78	0.78

1: GPSD, 2: ASDC, 3: GAAPC, 4: PAAC. Bold values represent the methods with the best performance in each metric

The impact of subset of chemical feature categories

When we combine various groups of chemical features together, each group can compensate for other groups deficiencies by integrating the content and knowledge in sequences detailly which will help predict tumor T cell antigens. By considering these assumptions, we evaluate the subset of chemical attributes performance comprehensively. Here, by considering all 4 groups of chemical attributes, we tried and evaluated all their combinations where SMOTE-Tomek handles the imbalanced dataset. By using these 11 combinations and four classifiers above with chemical features, we constructed 44 models in total and measured the predictive performance by 5-fold cross-validation. We found almost all combinations (10 out of 11) have performed better on combining ProtBERT and chemical attributes over the Inception neural network. Such a result also shows Inception neural network is the best classification method to separate tumor T cell antigens from non-antigens, once used with a proper subset of attributes. To find the best subset of chemical feature groups that leads to the best method, we show the results for ProtBERT + Chemical Features + Inception in Table 10 for Dataset 1. According to the table, a combination of at least 2 groups of chemical attributes outperforms the individual groups of chemical attributes. Among the remaining combinations with 2 or 3 groups of chemical attributes, there does not seem to be a regular pattern showing the outperformance of one combination over another combination. We obtain the best performance when we use all 4 groups of chemical attributes in terms of all performance metrics which corresponds to DeepTHybrid. By using all groups, we can obtain 0.76 accuracy and 0.78 F1 score. We obtain similar results when we repeat the same analysis for the Dataset 2 as shown in Table 11. According to the table, among the remaining combinations with 2 or 3 groups of chemical attributes, there again does not seem to be a regular pattern showing the outperformance of one combination over another combination. We again obtain the best performance when we use all 4 groups of chemical attributes in terms of all performance metrics which corresponds to DeepTHybrid Table 9.

Table 10.

The impact of the combination of attributes subset selection approach and the approach to handle imbalanced dataset on the classification performance over the dataset taken from Charoenkwan et al. [39] (Dataset 1)

Method	Accuracy	Precision	Recall	F1	AUC
ANOVA + SMOTE	0.70	0.72	0.74	0.73	0.72
ANOVA + Tomek	0.69	0.72	0.75	0.73	0.72
ANOVA + (SMOTE + Tomek)	0.71	0.73	0.74	0.73	0.73
ANOVA + (No Balancing)	0.72	0.74	0.73	0.74	0.74
MRMR + SMOTE	0.75	0.77	0.75	0.76	0.78
MRMR + Tomek	0.72	0.73	0.74	0.73	0.74
MRMR + (SMOTE + Tomek)	0.74	0.74	0.74	0.74	0.75
MRMR + (No Balancing)	0.74	0.75	0.74	0.75	0.76
D eep T Hybrid : MRMD + SMOTE	0.76	0.80	0.76	0.78	0.78
MRMD + Tomek	0.75	0.78	0.70	0.74	0.77
MRMD + (SMOTE + Tomek)	0.75	0.82	0.73	0.77	0.77
MRMD + (no balancing)	0.75	0.76	0.75	0.76	0.77

Bold values represent the methods with the best performance in each metric

Table 11.

The impact of the feature subset selection approach on the classification performance over the dataset taken from Herrera-Bravo et al. [40] (Dataset 2)

Method	Accuracy	Precision	Recall	F1	AUC
ANOVA	0.75	0.77	0.79	0.78	0.77
MRMR	0.80	0.80	0.80	0.80	0.81
D eep T Hybrid : MRMD	0.81	0.85	0.81	0.83	0.83

Bold values represent the methods with the best performance in each metric

Table 9.

The impact of distinct categories of chemical features on the classification performance over the dataset taken from Herrera-Bravo et al. [40] (Dataset 2)

Method	Accuracy	Precision	Recall	F1	AUC
1 + 2	0.70	0.71	0.72	0.71	0.71
1 + 3	0.70	0.70	0.70	0.70	0.72
1 + 4	0.72	0.71	0.73	0.72	0.72
2 + 3	0.73	0.73	0.74	0.73	0.74
2 + 4	0.74	0.74	0.76	0.75	0.75
3 + 4	0.76	0.76	0.79	0.77	0.77
1 + 2 + 3	0.79	0.82	0.89	0.81	0.81
1 + 2 + 4	0.77	0.77	0.81	0.79	0.80
1 + 3 + 4	0.79	0.87	0.81	0.84	0.81
2 + 3 + 4	0.80	0.83	0.83	0.83	0.82
1 + 2 + 3 + 4 (DeepTHybrid)	0.81	0.85	0.81	0.83	0.83

1: GPSD, 2: ASDC, 3: GAAPC, 4: PAAC. Bold values represent the methods with the best performance in each metric

The impact of attribute subset selection approach and imbalanced dataset

We also analyze the impact of feature selection methodology such as ANOVA, MRMR, MRMD as well as the impact of the method to handle our imbalanced datasets as discussed in Section Attribute selection. Our tested methods to handle imbalanced datasets are SMOTE, Tomek, and (SMOTE + Tomek) as discussed in Section Strategy for handling imbalanced datasets. Our proposed DeepTHybrid combines MRMD with SMOTE outperforms the rest of the combinations in terms of many metrics as seen in Table 10 for Dataset 1. In Dataset 1, apart from MRMD combined with SMOTE + Tomek performs the best in some metrics. Other than these two combinations, MRMR + SMOTE also appears to perform reasonably well showing that SMOTE can prevent overfitting in this dataset. Our results are similar in Dataset 2 as seen in Table 11 where the MRMD that is part of DeepTHybrid outperforms the rest of the feature selection methods in all metrics. In this case, dataset 2 is already balanced so we do not apply data balancing approaches such as SMOTE, Tomek, and SMOTE + Tomek.

Comparison of the performance with respect to protein sequence lengths

We also compare the effect of amino acid sequence lengths in DeepTHybrid. As seen in Table 2, for both datasets, protein lengths differ. For Dataset 1, the minimum protein length is 8 whereas the maximum length is 50. On the other hand, for Dataset 2, the minimum protein length is 8 whereas the maximum length is 20. By default, we use the maximum protein length in DeepTHybrid and padding was applied using 0 to equalize the length of the other amino acid sequence to the longest amino acid sequence. Table 12 shows the results for Dataset 1 where we compare the original DeepTHybrid with length 50 again with DeepTHybrid where it now trims proteins to 8 and 11 lengths. Here 11 is selected since it is the most repetitive amino acid sequence length in the dataset. Similarly, we show the results for Dataset 2 in Table 13 where the maximum protein sequence length is now 20. We compare the original DeepTHybrid using the whole proteins (20 length) with the minimum length 8, and 11 which is again selected since it is the most repetitive amino acid sequence length in the dataset. According to both tables, the best results are obtained when we use the whole sequence length and padding has been applied using 0 to equalize the length of the other amino acid sequence to the longest amino acid sequences.

Table 12.

Comparison of DeepTHybrid trained with a given amino acid sequence lengths on Dataset 1

Method	Accuracy	Precision	Recall	F1 Score
DeepTHybrid (length = 8)	0.70	0.75	0.75	0.75
DeepTHybrid (length = 11)	0.74	0.78	0.76	0.77
DeepTHybrid (length = 50)	0.76	0.80	0.76	0.78

The DeepTHybrid is originally trained with 50 amino acid lengths which is the maximum length of a potential antigen in the dataset. Bold values represent the methods with the best performance in each metric

Table 13.

Comparison of DeepTHybrid trained with a given amino acid sequence lengths on Dataset 2

Method	Accuracy	Precision	Recall	F1 Score
DeepTHybrid (length = 8)	0.73	0.75	0.76	0.76
DeepTHybrid (length = 11)	0.79	0.83	0.80	0.81
DeepTHybrid (length = 20)	0.81	0.85	0.81	0.83

The DeepTHybrid is originally trained with 20 amino acid lengths, which is the maximum length of a potential antigen in the dataset. Bold values represent the methods with the best performance in each metric

Feature importance analysis

In order to understand the contribution of both chemical and deep feature representations in the classification result, we run Shapley analysis which uses additive models for Shapley-based explanations of machine learning models [80]. In detail, we use SHAP (SHapley Additive exPlanations) [81] which is a game theoretic approach to explain the output of any machine learning model. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions. As seen in Fig. 3a–b for both datasets, SHAP analysis reveals that the features extracted by ProtBERT are as effective as the chemical features, and even they are more effective than the chemical features in Dataset 1. Although for Dataset 2, the second most effective feature is the GAAC Positive Charger with the second-highest SHAP value, according to SHAP analysis, ProtBERT features are also as influential since they are ranked at the topmost positions. In Fig. 3a consistency than GAAC Positive Charger. On the other hand, the chemical property groups CTDD and CTDC’s features importance are less than GAAC Positive Charger, but they are in between deep representation attributes from ProtBERT.

Fig. 3.

Attributes ranked by their importance of SHAP (Shapley Additive exPlanations) values for both datasets

Our analysis reveals several important observations. First, deep contextual embeddings derived from ProtBERT consistently appear among the highest-ranked features in terms of absolute SHAP values, confirming that the model effectively leverages semantic and long-range dependencies captured from protein sequences. This effect is particularly pronounced in Dataset 1, where multiple ProtBERT-derived dimensions dominate the top positions, indicating that deep representations play a central role when the data distribution is noisier or more heterogeneous. Second, chemical descriptors remain highly influential and complementary rather than redundant. In Dataset 2, grouped amino acid composition features, most notably positively charged residue-related attributes, exhibit strong SHAP contributions, suggesting that physicochemical properties become more discriminative in balanced and cleaner datasets. Importantly, neither deep nor chemical features uniformly dominate across both datasets; instead, SHAP analysis demonstrates that DeepTHybrid adaptively balances these information sources depending on dataset characteristics.

Finally, the dispersion patterns in SHAP summary plots indicate that ProtBERT features tend to show higher variance in their contributions across samples, reflecting sensitivity to sequence-specific context, whereas chemical features provide more stable and globally consistent signals. This complementary behavior explains the superior performance of the hybrid architecture and provides interpretability evidence that the model’s predictions are driven by biologically meaningful and diverse feature cues, rather than spurious correlations.

Specifically, ProtBERT-derived embeddings consistently rank among the most influential features, indicating that long-range contextual dependencies and higher-order sequence semantics play a central role in antigenicity prediction, in agreement with immunological evidence that antigen presentation and T-cell recognition depend on both local motifs and broader sequence context rather than isolated residues [82, 83]. From an immunological perspective, these findings suggest that DeepTHybrid captures complementary determinants of antigenicity. The high contribution of ProtBERT-derived features indicates that tumor T cell antigen recognition is not governed only by isolated amino-acid residues, but also by broader sequence context, motif organization, and higher-order dependencies. This is consistent with the biological mechanism of antigen presentation, where peptide processing, MHC binding, and T-cell receptor recognition depend on both local residue patterns and the surrounding sequence environment. In parallel, the strong contribution of chemically interpretable descriptors, such as grouped amino-acid composition and charge-related attributes, suggests that physicochemical properties remain important for distinguishing immunogenic from non-immunogenic peptides. Positively charged, hydrophobic, and composition-based features may influence peptide stability, MHC-binding compatibility, and recognition by immune receptors. Therefore, the SHAP results support the biological rationale of the hybrid design: ProtBERT embeddings provide contextual sequence-level information, while chemical descriptors provide explicit biochemical signals relevant to antigen processing and immune recognition.

Limitations and practical considerations

Despite the strong empirical performance of DeepTHybrid, several limitations should be acknowledged. First, regarding dataset bias, the benchmark datasets used in this study are derived from curated repositories of experimentally reported tumor T cell antigens. As with many bioinformatics benchmarks, these datasets may reflect biases introduced by experimental discovery pipelines, annotation practices, and class definitions. In particular, negative samples may include latent or weakly immunogenic peptides that have not yet been experimentally validated, potentially introducing label noise. In addition, the datasets are restricted to specific peptide length ranges and species distributions, which may limit the generalizability of the model to underrepresented cancer types, organisms, or experimental conditions. Although we partially mitigate these issues by evaluating DeepTHybrid on two independently constructed datasets, including a corrected and balanced version, further validation on larger and more diverse TTCA datasets remains an important direction for future work.

These dataset constraints also limit the extent to which the current results can be interpreted as evidence of broad clinical or pan-cancer generalizability. Although the two datasets used in this study provide complementary evaluation settings, they are still relatively small compared with the diversity of tumor antigens observed across cancer types, HLA/MHC backgrounds, patient populations, and experimental platforms. In particular, the model has not yet been evaluated across distinct tumor types, different ethnic or HLA distributions, independent laboratory-generated cohorts, or experimentally validated neoantigen datasets. Moreover, the negative class in TTCA prediction is inherently difficult to define, because peptides labeled as non-antigens may include untested, context-dependent, or weakly immunogenic candidates. Consequently, the reported performance should be interpreted as benchmark-level predictive performance rather than definitive evidence of clinical validity.

Second, in terms of computational complexity, the primary computational cost of DeepTHybrid arises from the use of ProtBERT embeddings. ProtBERT is a large transformer-based protein language model with substantial memory and inference requirements compared to traditional handcrafted feature extraction pipelines. While ProtBERT embeddings are computed offline and reused during training and inference, rendering the classification stage itself efficient, the overall pipeline is still more computationally demanding than classical machine-learning approaches. This may limit applicability in resource-constrained environments or large-scale high-throughput screening scenarios. Future work may explore lighter protein language models, embedding compression or distillation strategies, and parameter-efficient fine-tuning methods to reduce computational overhead while preserving predictive performance.

Future validation should therefore include several complementary strategies. First, DeepTHybrid should be tested on external TTCA and neoantigen datasets that were not used during model development or hyperparameter tuning. Second, cross-dataset validation should be performed, where the model is trained on one curated dataset and evaluated on another dataset with different annotation criteria, peptide-length distributions, or cancer-type composition. Third, prospective biological validation should be conducted by experimentally testing high-confidence predicted antigens using MHC-binding assays, T-cell activation assays, or immunogenicity assays. Finally, integrating additional biological context, such as HLA allele information, MHC-binding affinity, antigen processing scores, tumor expression levels, mutation status, and immune microenvironment features, may improve both the biological specificity and translational relevance of the framework.

Conclusion and future work

In this research, we have come up with a novel approach, DeepTHybrid, which is a Hybrid Inception-based neural network model that processes deep features from ProtBERT, which has a structure similar to BERT, and combines them with hand-designed chemical features. Overall, DeepTHybrid runs these features in parallel with convolutional layers inspired by the Inception module. Our approach utilized the feature selection technique MRMD, and the imbalance problem was addressed using the resampling method SMOTE. Via our approach, we have analyzed the impact of deep features and chemical features when they are used together in predicting tumor T cell antigens. To our best knowledge, our approach is the first attempt to use deep learning, especially deep protein representations, in tumor T cell antigen prediction.

We expect DeepTHybrid to be a robust, reliable, and useful computational tool for tumor T cell antigen identification. Overall, we outperform a number of existing methods while predicting tumor T cell antigens. In general, our method shows consistent performance across multiple datasets, whereas the existing approaches may have a lower performance in some of them. Our method still requires further development.

As future work, DeepTHybrid should be extended and validated in several directions. First, larger and more diverse TTCA and neoantigen datasets should be incorporated to better assess generalizability across cancer types, HLA/MHC backgrounds, species, peptide-length distributions, and experimental platforms. Second, external and cross-dataset validation should be performed to evaluate whether the learned representations transfer beyond the current benchmark datasets. Third, high-confidence predicted antigens should be prioritized for downstream experimental validation, including MHC-binding, T-cell activation, and immunogenicity assays. Fourth, additional biologically informative variables, such as HLA allele type, antigen processing likelihood, tumor expression, mutation status, and immune microenvironment features, could be integrated into the model to improve biological interpretability and clinical relevance. Finally, alternative deep architectures, including Vision Transformers and lighter protein language models, may be explored to improve accuracy, scalability, and deployment in large-scale antigen screening pipelines.

Author contributions

UO: Software, Methodology, Writing. BT: Software. ES: Conceptualization, Software, Formal Analysis, Methodology, Writing. All authors read and approved the final manuscript.

Data availability

Two different datasets were used in this paper. The first dataset is taken from Charoenkwan et al. The second dataset appears in Herrera-Bravo et. al. The datasets are also available on https://github.com/seferlab/deepthybrid.

Code availability

The proposed method and analysis code is available on https://github.com/seferlab/deepthybrid.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no conflict of interest.