A Recent Appraisal of Artificial Intelligence and In Silico ADMET Prediction in the Early Stages of Drug Discovery

In silico ADMET models have progressed significantly over the past ~4 decades, but still, the pharmaceutical industry is vexed by the late-stage toxicity failure of lead molecules. This problem of late-stage attrition of the drug candidates because of adverse ADMET profile motivated us to analyze the current role and status of different in silico tools along with the rise of machine learning (ML) based program for ADMET prediction. In this review, we have differentiated AI from traditional in silico tools because, unlike traditional in silico tools where the final decision is made manually, AI automates the decision-making prerogative of humans. Due to the large volume of literature in this field, we have considered the publications in the last two years for our review. Overall, from the literature reviewed, deep neural networks (DNN) algorithm or deep learning seems to be the future of ML-based prediction models. DNNs have shown the ability to learn from more complex data and this gives DNN an edge over other ML algorithms to be applied for ADMET prediction. Our result also suggests that we need closer collaboration between the ADMET data generators and those who are employing ML-based tools on this generated data to build predictive models, so that more accurate models could be developed. Overall, our study concludes that ML is still a work in progress and its appetite for data has not been sated yet. It needs loads of more quality data and still some time to prove its real worth in predicting ADMET.