RLEAAI: improving antibody–antigen interaction prediction using protein language model and sequence order information

Antibody-antigen interactions (AAIs) are a pervasive phenomenon in the natural and are instrumental in the design of antibody-based drugs. Despite the emergence of various deep learning-based methods aimed at enhancing the accuracy of AAIs predictions, most of these approaches overlook the significance of sequence order information. In this study, we propose a new deep learning-based method RLEAAI, to improve the prediction performance of AAIs. In RLEAAI, a sequence order extraction strategy, called Composition of K-Spaced Amino Acid Pairs, is employed to generate feature representation from the feature embedding outputted by a pre-trained protein language model. In order to fully dig out the discrimination information from features, three neural network modules, i.e. convolutional neural network, bidirectional long short-term memory network and recurrent criss-cross attention mechanism, are integrated. Benchmarked results on two independent test sets demonstrate that RLEAAI is capable of achieving an average accuracy of 0.7787 and an average Matthews's correlation coefficient (MCC) value of 0.5552, representing a 5.2% and 15.8% improvement over the start-of-the-art method DeepAAI. Furthermore, the complementary determining regions-sensitivity value calculated on MCC of RLEAAI is 216.4% higher than that of the state-of-the-art method DeepAAI. The standalone package of RLEAAI is freely available at https://github.com/zhouyu9931/RLEAAI.git.