KAMDA: A Multi-View Kolmogorov-Arnold Network Integrating Pre-trained BERT Embeddings and Similarity-Based Imputation for Microbe-Drug Association Prediction

The prediction of drug-microbe interactions faces difficulties because of three main factors, including incomplete networks, missing data and restricted capabilities of single-network models. To solve this problem, in this paper, we proposed a KAMDA framework integrating multi-view graph learning with Kolmogorov-Arnold Networks and pre-trained language models. The KAMDA framework presents three main innovations: a dual-view encoding system that merges biological similarities with Gaussian Interaction Profiles to extract both intrinsic biological patterns and topological patterns; KAGAT layers which use learnable B-spline activation functions to provide better non-linear modeling capabilities than fixed MLPs; a BERT-based imputation system which uses ChemBERTa and DNABERT to fill in missing information. The experiments conducted on aBiofilm showed noticeable results with both the AUC and AUPR scores reached remarkable values of 0.9774 and 0.9719, respectively. It could be noticed to be a potential powerful tool for microbe-drug association prediction. The code and data are available at: https://anonymous.4open.science/r/KAMDA-1F16 • A KAMDA framework integrating multi-view graph learning with Kolmogorov-Arnold Networks and pre-trained language models for Microbe-Drug Association Prediction with three main contributions: • A dual-view encoding system that merges biological similarities with Gaussian Interaction Profiles to extract both intrinsic biological patterns and topological patterns. • KAGAT layers which use learnable B-spline activation functions to provide better non-linear modeling capabilities than fixed MLPs • A BERT-based imputation system which uses ChemBERTa and DNABERT to fill in missing information