Predicting protein-ligand binding free energy with AI : does the neural network architecture really matter ?

Boris TOUZEAU^1,2*, Lin Jung-Hsin^2,3,4,5

¹Chemical Biology and Molecular Biophysics, Taiwan International Graduate Program, Academia Sinica, Taipei, Taiwan
²Research Center for Applied Sciences, Academia Sinica, Taipei, Taiwan
³Institute for Biomedical Sciences, Academia Sinica, Taipei, Taiwan
⁴School of Pharmacy, College of Medicine, National Taiwan University, Taipei, Taiwan
⁵College of Engineering, Chang Gung University, Taoyuan, Taiwan

* Presenter:Boris TOUZEAU, email:boris.touzeau@gmail.com

In this report, we explore the performances of various deep learning (DL) approaches for predicting the binding free energy of protein-ligand complexes and put it in comparison to the impact of the size and the quality of the database on the results. Among the state-of-the-art techniques available in artificial intelligence, the convolutional neural network (CNN), a network widely used in image recognition, has been, and still is, one of the best1. Its last successful implementation made use of 2 improvements ; the residual link and the inception block2 and achieved a 3.57% error on the ImageNet testset, (as a comparison the human brain error is 5.01%). Hence it has been adopted by various laboratories; physics, biology, but also in structural biochemistry, where several articles have been using a CNN in order to build a predictive model on binding free energy prediction3-5. However, while CNN is a powerful DL algorithm that can take “pictures” (or grids) as input to its network and extract the features from them, it is still viewed as a ‘black box’. Indeed, the rational is understood but we still are unable to know which features are deemed crucial for a given network and which should be discarded. While various architecture models, namely, AlexNet, VGGNet, ResNet, etc, perform differently, the core principle behind is rather similar and relies on 3 aspects: feature extraction, feature training; and prediction. Thus, the learning process is similar in consequence, a true improvement in feature learning might come from somewhere else more so than a change in architecture. It has been shown in by Halevy et al. in 20096 and Mahajan et al. in 20187, respectively, that the size of the dataset may actually matter more than the architecture itself. This is why we will present the result of several experiments we did in order to answer the following questions ; 1. What are the differences between ResNet and VGGNet architectures when predicting the binding affinity of differently sized database? 2. Is the size of the database as impactful as the chosen architecture? Do two networks have different trends in prediction? We concluded that the correlation coefficients of predictions with the ResNet-like networks are somewhat superior to those of VGGNet-like networks at equally-sized databases. It will provide us with some insights on the roles of the datasets and the architectures.

Keywords: Deep Learning, Convolutional Neural Network, AI