Deep Neural Networks for ADMET properties' prediction

Introduction

The absorption, distribution, metabolism, excretion and toxicity properties of any compound contributes to its potential as an orally administered drug. A set of criteria defined by Rule of Five or the Pfizer's Rule of Five stats that any compound meeting these criteria are more likely to have ADMET properties. The importance of these properties has increased in modern medicinal chemistry and new machine learning techniques are employed to predict these properties. The predictive ADMET models have helped in discovery of small molecules with improved safety and dose.

Deep Neural Networks are showing usefulness in such predictions due to improved computational efficiency, larger datasets and adaption of image processing in the chemistry.

Early Models

The early DNNs employed showed improved prediction performance. Many researchers raised concerned that this improvement is in fact has no relation with better computational capacity, rather it is result of mere memorizing of molecules regardless of relevance. This situation can be potentially be beneficial if molecules are closely related, however it get worsen when molecules are not related.

A fully connected DNN was trained using liver microsome stability dataset of three species. It was found that the prediction improved for Human Liver Microsome (HLM), however it degraded for Rat and Mouse Liver Microsomes (RLM and MLM respectively). With addition of 8 new species in the dataset, the model show mixed performance, improved for one while degraded for another. Thus these approaches were mere regarded as 'trial and error' type of work.

Improvements

One of the reason for such memorization by the model was because early DNNs required 'fingerprint' representation of each compound. The fingerprints are vectors of discrete numbers where each number represents presence (0 or 1) or the number of chemical fragments. Hence a graph CNN (GCNN) was developed to dynamically learn a fingerprint optimized for most relevant chemical information.

Numerous GCNNs were tested using different industrial and public ADMET datasets. Overall, the GCNN showed 14% improvement from the traditional static fingerprint models.

Measuring Generalization

The performance of model in drug discovery also depends on the type of split. Generally, time split and scaffold splits are considered in case of drug discovery rather than random split. In time split, the earlier data is used for training while the recent data is used for testing. The scaffold split on the other hand uses core structures around which the compound is built to split the data. This ensures that compounds with similar core structures are kept together. The compounds with one scaffold are used in training to predict related properties, while compounds with another scaffold is used during test to analyse model's generalization.

Another type of split designed for the domain was time plus molecular weight split. Here molecules with weight less than 500 Da are used for training and molecules with weight more than 600 Da are used for testing. The GCNNs can be better at learning structure-activity relationship of smaller molecules which can be extrapolated to larger molecules.

The Future

It is believed that the GCNNs will greatly improve the drug discovery process in the future. It's predictive efficiency has already surpassed fully connected DNN and other ML techniques including RF and SVM. Also, if GCNN are made to learn the biophysical properties of the compound than it performance may increase drastically. The research is made continuously to improve the efficiency of DNNs in the domain.