Integrating Graph Convolutional Network to BiLSTM for Enhanced Polymer Knot Identification

Classifying complex knots in polymer science is a time-consuming task, often complicated by the close numerical values associated with knot invariants. Reported neural network models for knot classification treat polymers as one-dimensional sequences, neglecting the rich spatial structural information on knots. In this study, we address the challenges by proposing a feature fusion model that integrates graph convolutional network and bidirectional long short-term memory network for enhanced knot classification in polymer science. Our model treats the conformation of a polymer chain with a knot as a graph structure to describe the topological connection relationships among neighboring monomers. It fuses the characteristics of bonded monomers (one-dimensional sequence information) and the features of neighboring nonbonded monomers (two-dimensional graph information), thereby achieving higher prediction accuracy for knot-type recognition. Machine learning (ML) experiments are conducted on 18 different knot types for N = 100 and 36 different knot types for N = 200, demonstrating the superior accuracy of our model compared to previous methods. Furthermore, our study demonstrates the superior computational efficiency of our model compared to traditional approaches based on Alexander polynomials, indicating the potential of ML techniques to revolutionize knot classification in polymer science. These findings enhance the understanding of factors influencing knot classification accuracy in polymer chains, positioning our model as a valuable tool for advancing the analysis of intricate knot configurations in polymer science and related fields.