# Accurate identification of broadly neutralizing antibodies against dengue virus based on deep stacking strategy with multi-perspective features

## TL;DR

This paper introduces Deepstack-NAb, a computational method that accurately identifies antibodies effective against all four dengue virus types using machine learning and deep learning techniques.

## Contribution

The novel contribution is the development of Deepstack-NAb, a high-accuracy stacking ensemble method for in silico identification of broadly neutralizing dengue antibodies.

## Key findings

- Deepstack-NAb outperformed existing methods with an accuracy of 0.905 on independent tests.
- The method achieved a 20.65% improvement in MCC over the PredNAb method.
- Multi-perspective features and stacking ensembles enhanced the model's predictive and generalization abilities.

## Abstract

Rapid and precise screening for broadly neutralizing antibodies (bNAbs) against all four dengue virus (DENV) serotypes (DENV-1 to DENV-4) is essential to accelerate the development of effective therapeutic antibodies and combat the global burden of dengue fever. However, characterizing bNAbs against DENV through experimental methods remains a significant challenge, as it requires costly laboratory conditions and time-consuming procedures. Therefore, computational methods for in silico identification of bNAbs against DENV based on sequence information are highly desirable and can effectively complement established experimental approaches. Thus, we propose a novel and high-accuracy computational approach, called Deepstack-NAb, for the precise identification of bNAbs against DENV. This method utilizes a stacking ensemble of multiple machine learning (ML) and deep learning (DL) algorithms. First, in Deepstack-NAb, multi-source feature encoding schemes were employed to capture key information from CDR-H3 and epitope data, including conventional feature encodings, natural language processing (NLP)-based feature encodings, and pretrained protein language models (PLMs). Second, all feature descriptors were then fused and optimized using a feature selection technique. Subsequently, well-trained ML and DL methods were constructed to generate multi-perspective features, covering both probabilistic and class information. Finally, the most informative features were selected and employed for optimizing the stacking ensemble model. We conducted comparative experiments to evaluate the performance of Deepstack-NAb against its baseline models and the existing method (i.e., PredNAb). Both cross-validation and independent test results showed that Deepstack-NAb provided a superior performance compared with both its baseline models and PredNAb. On the independent test, Deepstack-NAb achieved an accuracy of 0.905, sensitivity of 0.922, and an MCC of 0.810, representing improvements of 10.34, 13.44, and 20.65%, respectively, over PredNAb. Altogether, the excellent predictive and generalization ability of this new method Deepstack-Nab, highlights its potential as a valuable tool for accurately identifying bNAbs against DENV, significantly reducing false negatives in the process.

The online version contains supplementary material available at 10.1038/s41598-025-31332-3.