Discussion

There has been significant progress in understanding bacterial and viral VFs, but fungal VFs remain comparatively less characterized, particularly at the molecular and protein levels. To help address this gap, we developed and validated ML models specifically designed to predict VFs in human fungal pathogens. This approach aims not only to improve predictive performance but also to provide biologically meaningful insights that can guide future studies of fungal pathogenicity and antifungal strategies.

We evaluated four ML models: RF, ANN, MLP, and DNN, using a diverse set of protein features, including two conventional sequence-based features, namely, AAC and DDE, deep-learning-based UniRep embeddings, and combinations of these features through feature fusion. Our analysis showed that UniRep embeddings, both independently and when combined with traditional sequence-based features, improved model performance across validation and independent datasets. The improvement associated with UniRep suggests that learned protein sequence representations may capture information related to protein structure and function that is not fully represented by conventional sequence-derived features alone.

We employed a range of commonly used metrics, including accuracy (ACC), sensitivity (SN), specificity (SP), MCC, and area under the receiver operating characteristic curve (AUC-ROC), to evaluate classification performance. The binary classification framework provided a basis for applying ML algorithms to distinguish VFs from non-virulent proteins. Our analysis revealed that incorporating UniRep embeddings enhanced prediction performance. This improvement was observed when UniRep was used alone or in combination with traditional sequence-based features such as AAC and DDE, as demonstrated in Table 2. These results highlight RF as the most effective approach among the evaluated models, with the integration of pre-trained protein embeddings such as UniRep contributing to improved prediction performance. To further evaluate the RF classifier, its performance was assessed on an independent dataset.

Performance evaluation of features used in the ML model

Traditional features, including dipeptide composition (DPC) and AAC, have been extensively applied in bacterial virulence prediction approaches such as PathoFact33 and VirulentPred.16 For example, PathoFact incorporated features such as AAC, DPC, and composition-, transition-, and distribution-based descriptors (CTDC, CTDT, and CTDD).33 Based on these established approaches, we incorporated commonly used sequence-derived features into our pipeline for predicting fungal VFs. However, our results showed that models using only conventional sequence-based features achieved moderate performance, with accuracies ranging from 62% to 68%. These results indicated that additional feature representations were required to improve prediction performance. We therefore explored UniRep embeddings, a pre-trained deep representation that captures 1900-dimensional protein sequence features derived from large-scale protein sequence datasets.

Our analysis identified UniRep, a unified representation for protein sequence embeddings, as a promising feature representation for protein classification. By integrating UniRep with conventional features into a single dataset, we improved the classification of virulent and non-virulent proteins using RF, which achieved an accuracy of 77.4% and an MCC of 0.5509 on the validation set. Testing on independent data demonstrated the ability of the model to generalize to previously unseen fungal sequences, achieving an accuracy of 68.7% and an MCC of 0.4803. UniRep is a sequence-based embedding approach trained on approximately 24 million UniRef50 protein sequences and generates a flexible representation of protein sequence properties that has been applied in protein informatics and protein engineering.32 Previous studies have also demonstrated the utility of UniRep representations. For example, Zhong et al. reported that a LightGBM model incorporating UniRep sequence embeddings together with secondary structure annotation (SSA) achieved the highest performance among the evaluated individual predictors.34

These observations prompted us to incorporate UniRep embeddings into our prediction pipeline, both independently and in combination with traditional features such as AAC and DDE. In 5-fold cross-validation, incorporation of UniRep substantially improved model performance, with accuracies exceeding 75% in the evaluated datasets. Furthermore, the use of UniRep embeddings resulted in consistently improved predictive performance across the validation and independent datasets, supporting their robustness as a protein representation for fungal VF prediction.

These findings indicate the potential of UniRep as a feature representation for virulence prediction. By leveraging deep representation learning, UniRep can encode complex sequence information that may not be adequately captured by conventional features alone. This may contribute to the ability of the model to distinguish between virulent and non-virulent proteins and supports the application of protein language model-derived representations in computational studies of fungal pathogenicity.

FunVFPred addresses an important computational need in fungal pathogenicity research by providing an approach specifically designed to predict fungal VFs. The predictive performance of the model was enhanced through the incorporation of UniRep embeddings, either independently or in combination with traditional sequence-based features such as AAC and DDE. Among the evaluated ML and DL approaches, RF demonstrated the strongest overall performance, achieving an accuracy of 77.4% and an MCC of 0.5509 on the validation set. Further evaluation on an independent dataset containing fungal sequences not included in the training data demonstrated the model’s ability to generalize, with an accuracy of 68.7% and an MCC of 0.4803. The robustness of the RF model was further assessed using 5-fold cross-validation, yielding an accuracy of 85.7% (MCC 0.7142) for the independent dataset and 73.4% (MCC 0.4681) for the validation dataset.

Beyond predictive performance, FunVFPred also provides opportunities for biological interpretation of predicted virulence-associated proteins. Several predicted proteins were associated with functional categories relevant to fungal pathogenicity, including adhesion, hydrolysis, and regulatory processes, which contribute to host colonization, tissue invasion, and evasion of host defense mechanisms. For example, in Candida albicans, proteins such as ALS3 and HWP2 play important roles in adhesion and biofilm formation, while BUD4 and VPS11 contribute to morphogenesis and pathogenicity. Comparison of predicted proteins with established classes of VFs provides additional support for the potential utility of FunVFPred in prioritizing candidates for experimental investigation and identifying potential targets for antifungal research.22

Feature importance and sequence pattern analyses further indicated that the predictions generated by FunVFPred can be examined in a biologically interpretable context. Adhesins and hydrolytic enzymes were predicted with relatively high accuracy, whereas predictions for signaling and metabolic proteins were comparatively less accurate. This difference may reflect the greater sequence heterogeneity of signaling and metabolic proteins compared with proteins involved in adhesion and hydrolytic functions. These observations suggest that future improvements in dataset size, annotation quality, and feature representation may further enhance VF prediction across diverse functional categories.

There are several limitations to our approach. First, publicly available databases may contain incomplete or unevenly distributed VF annotations, which can introduce biases during model training. Second, the predominantly Candida-based training dataset may limit model performance when predicting VFs from more distantly related fungal species. Third, as with many ML approaches involving high-dimensional learned representations, the use of deep protein embeddings may limit the biological interpretability of some model predictions. Despite these limitations, the developed framework provides a useful starting point for large-scale computational prioritization of candidate VFs in human-pathogenic fungi and for guiding subsequent experimental investigations.

FunVFPred integrates traditional sequence-based features with deep protein embeddings to provide a computational approach for predicting fungal virulence-associated proteins. The combination of biological interpretation and model-based prediction may facilitate the generation of hypotheses for experimental validation, prioritization of candidate antifungal targets, and further investigation of fungal virulence mechanisms. FunVFPred is freely accessible through its GitHub repository, GUI web server, and Docker container. Overall, FunVFPred provides a computational resource for prioritizing candidate virulence-associated proteins across fungal proteomes and supporting future studies of fungal pathogenicity.

Limitations of the study

A key limitation of this study is the scarcity of experimentally validated fungal VF datasets. Unlike bacterial VFs, which benefit from large, well-curated repositories, fungal VF data in the public domain remain relatively scarce, limiting the diversity of virulence-associated proteins available for model development. In addition, although we compared our RF-based approach with multiple ML models, including ANN, DNN, and MLP, no dedicated computational tool is currently available for predicting fungal VFs in human-pathogenic fungi, precluding direct, tool-to-tool benchmarking. As additional experimentally validated fungal VF datasets become available, future studies may further improve model performance, enable more comprehensive benchmarking, and facilitate the development of increasingly robust computational approaches for fungal virulence prediction.