Bringing a new drug to market is a series of long processes involving multiple phases like early discovery, preclinical research, clinical development and regulatory approval, before culminating in market entry. The journey through this funnel typically spans a decade or more and costs billions of dollars. It is estimated that only 1 in 10,000 compounds initially considered, finally makes it to the market. There is clearly a need for solutions that speed up go-to-market times and reduce the development costs of drugs and vaccines. This need has an immediate poignancy in the current healthcare crises faced by the world.
There are several companies that currently apply Artificial Intelligence techniques to the discovery of new molecules or for repurposing existing ones. What are less prevalent today are tools that help shepherd the next phases that are equally critical. For example, in the discovery phase, Deep Learning-based models can be employed to predict potential toxic effects of a new drug molecule under investigation. These models can analyse data from multiple modalities to help with the crucial go/no-go decisions in the selection of the safest drug molecule. Identifying or predicting the risk road map and failure points significantly reduce the time and resources needed in this early phase of drug development. Importantly, this also helps reduce the need for animal sacrifice, and provide a boost to humane animal research in line with the 3R principles of animal testing (Reduce, Refine, Replace).
In addition, solutions can be applied to help improve the evaluation of parameters developed to assess efficacy of a drug molecule against a disease process. The requirement for accurate assessment and comparison of features between diseased and control tissues, in an accelerated manner, is a critical aspect of this phase. Deep learning solutions are providing promising results in this area, since they have the potential to provide objective, quantifiable and reproducible inferences on the efficacy of the drug molecule.
In addition to demonstrating efficacy in the discovery phase, drug candidates also go through a significant number of safety assessment studies, including the pre-clinical toxicologic pathology phase. In this phase, pathologists need to screen and analyse thousands of tissue images included in toxicology studies – needless to say, a manually intensive, time consuming task. Deep learning techniques suit this problem well by automating sorting and triaging of images. Results can be made available faster, are evidence based, and reproducible, leading to savings of several person-months that are otherwise spent on reporting toxicology studies.
New drug candidates also need to be tested for toxicity against specific biological systems like cardiovascular, pulmonary, reproductive, etc. These can often be tedious tasks and prone to errors. As an example, analysing a single image from testicular tissue sections, in Spermatogenic Staging assessment studies for the male reproductive system, currently takes several days! Crucially, very few pathologists have the expertise to attempt this task. Deep Learning solutions currently exist that provide automated classification of different stages of spermatogenesis. This helps in providing important clues towards disturbances in spermatogenesis that are attributable to toxic effects of drug molecules. Importantly, application of deep learning techniques for analysis has reduced the time taken for assessment of a tissue section from several days to a few minutes.
Pharmaceutical companies are constantly looking for ways to reduce the risks and corresponding investments needed at various points of the drug development funnel. Deep Learning techniques are well suited to make a difference in complex, expensive processes like drug development. A reduction in both risk and investment will help bring safe, efficacious drugs to market in a shorter time.