Evaluation of network architecture and data augmentation methods for deep learning in chemogenomics

Abstract

Among virtual screening methods that have been developed to facilitate the drug discovery process, chemogenomics presents the particularity to tackle the question of predicting ligands for proteins, at at scales both in the protein and chemical spaces. Therefore, in addition to to predict drug candidates for a given therapeutic protein target, like more classical ligand-based or receptor-based methods do, chemogenomics can also predict off-targets at the proteome level, and therefore, identify potential side-effects or drug repositioning opportunities. In this study, we study and compare machinelearning and deep learning approaches for chemogenomics, that are applicable to screen large sets of compounds against large sets of druggable proteins. State-of-the-art drug chemogenomics methods rely on expert-based chemical and protein descriptors or similarity measures. The recent development of deep learning approaches enabled to design algorithms that learn numerical abstract representations of molecular graphs and protein sequences in an end-to-end fashion, i.e., so that the learnt features optimise the objective function of the drug-target interaction prediction task. In this paper, we address drug-target interaction prediction at the druggable proteome-level, with what we define as the chemogenomic neuron network. This network consists of a feed-forward neuron network taking as input the combination of molecular and protein representations learnt by molecular graph and protein sequence encoders. We first propose a standard formulation of this chemogenomic neuron network. Then, we compare the performances of the standard chemogenomic network to reference deep learning or shallow (machine-learning without deep learning) methods. In particular, we show that such a representation learning approach is competitive with state-of-the-art chemogenomics with shallow methods, but not ultimately superior. We evaluate the most promising neuron network architectures and data augmentation techniques, such as multi-view and transfer learning, to improve the prediction performance of the chemogenomic network. Our results shed new insights on the design of chemogenomics approaches based on representation learning algorithms. Most importantly, we conclude from our observations that a promising research direction is to integrate heterogeneous sources of data such as various bioactivity datasets, or independently, multiple molecule and protein attribute views, instead of focusing on sophisticated, yet intuitively relevant, encoder's neuron network architecture.

Comprehensive Survey of Recent Drug Discovery Using Deep Learning

“…Transfer learning, as part of lifelong learning, is inspired by how quickly humans acquire new knowledge from other similar experiences in the past. Transfer learning can improve many problems of insufficient data by fine-tuning a pre-trained model with a large dataset in another or a general field to an actual small-scale dataset [ 181 ]. Bonggun et al [ 23 ] imported a molecule representation model learned from the PubChem database and applied it to their DTI model to improve performance.…”

“…Other metrics such as concordance index (CI or C-index) and Spearman's correlation coefficient (ρ) quantify the quality of rankings by comparing the order of the predictions and the order of the ground truths. A frequently used ranking metric in the DTA prediction is the CI [25,130,181]. When predicting the binding affinity values of two random DTPs, the CI measures whether those values were predicted in the same order as their actual values.…”

Modeling DTA by Combining Multiple-Instance Learning with a Private-Public Mechanism

“…Traditionally, an experimental assay is the surest way to obtain the desired binding affinity, but it is expensive and time-consuming to use this approach to analyze many possible DT pairs. A plethora of drug-like compounds and latent protein targets pose greater challenges because multiple drugs can be associated with multiple targets [ 1 , 2 ]. As a result, drug–target affinity (DTA) prediction has attracted considerable attention in recent years [ 3 ].…”

“…Actually, there are a variety of scenarios in which the classifications of objects (bags) can only be determined by some key components (instances), such as medical diagnoses; that is, some instances trigger the bag label. Following this concept, DTIs can be characterized by an MIL framework: the private representation contains abundant information that has been proven to be effective for DTA prediction [ 6 , 9 , 11 , 19 ], as does each public feature obtained via early fusion [ 2 , 12 , 13 , 14 ] and each public feature obtained via concatenation [ 4 , 5 , 7 , 20 ]. However, the exact contribution of each instance to the final DTA value of the bag is unknown.…”

Modeling DTA by Combining Multiple-Instance Learning with a Private-Public Mechanism

“…Actually, there are a variety of scenarios in which the classifications of objects (bags) can only be determined by some key components (instances), such as medical diagnoses; that is, some instances trigger the bag label. Following this concept, DTIs can be characterized by an MIL framework: the private representation contains abundant information that has been proven to be effective for DTA prediction [6,9,11,19], as does each public feature obtained via early fusion [2,[12][13][14] and each public feature obtained via concatenation [4,5,7,20]. However, the exact contribution of each instance to the final DTA value of the bag is unknown.…”