1. Abstract Protein-protein interactions (PPIs) are a vital phenomenon for every biological process. Prediction of PPI can be very helpful in the probing of protein functions which can further help in the development of new and powerful therapy designs for disease prevention. A lot of experimental studies have been done previously to study PPIs. However, lab-based experimental studies of PPI prediction are resource-extensive and time-consuming. In recent years, several high throughput, computational approaches to predict PPI have been developed but they could be fallible in terms of accuracy and false-positive rate. To overcome these shortcomings, we propose a novel approach AE-LGBM to predict the PPI more accurately. This method is based on the LightGBM classifier and utilizes the Autoencoder, which is an artificial neural network, to efficiently produce lower-dimensional, discriminative, and noise-free features. We incorporate conjoint triad (CT) features along with Composition-Transition-Distribution (CTD) features into the model and obtained promising results. The ten-fold cross-validation results indicate that the prediction accuracies obtained for Human and Yeast datasets are 98.7% and 95.4% respectively. This method was further evaluated on other datasets and has achieved excellent accuracies of 100%, 100%, 99.9%, 99.2% on E.coli, M.musculus, C.elegans, and H.sapiens respectively. We also executed AE-LGBM over three important PPI networks namely, single-core network (CD9), the multiple-core network (The Ras/Raf/MEK/ERK pathway), and the cross-connection network (Wnt Network). The method was successful in predicting the pathway with an impressive accuracy of 100%, 100%, and 98.9% respectively. These figures are significantly higher than previous methods that are based on state-of-the-art models and models including LightGBM or Autoencoder, proving AE-LGBM to be highly versatile, efficient, and robust.