Artificial Intelligence for the drug discovery market is forecasted to reach US $8,419 million by 2026. This traction is due to the growing conviction in the potential of Artificial Intelligence (AI) in reimagining drug discovery. Traditional R&D efforts often take 11-15 years, costing more than $2.6 billion. In addition, nine out of ten drug candidates fail phase I trials and regulatory approval, making this process ineffective and expensive.

Leading biopharmaceutical companies believe in AI due to growing awareness in the pharmaceutical sector and rising investment in drug development. AI works with the latest advanced biology and chemistry processes to develop state-of-the-art algorithms; it has the potential to move drug screening from the bench to a virtual lab without the need for extensive experimental input and manpower.

Most biopharmaceutical companies have started internal programs or are collaborating to develop AI platforms for increasing the research power of immuno-oncology drugs, metabolic-disease therapies, cancer treatments, and many other therapeutic targets.

AI is Changing Drug Discovery

Applications of AI in drug discovery are extensive and have the following classifications:

1) Target selection and validation

Target Identification deals with identifying the function of the possible molecular targets (genes/proteins of a small molecule) and their role in a disease, aimed at finding the efficacy target of a drug. Selection and validation require the evaluation of functional genomics, structural genomics, proteomics, cell-based assays (in-vitro), and animal research (in-vivo) assays.

AI analyzes Drug Information Bank (which contains drug candidates, gene expressions, protein-protein interactions and clinical data records) from a public library for predicting therapeutic potential. For example, when applying feature engineering by deep autoencoder, relief algorithm, and binary classification (using the Xgboost algorithm on a genome-wide protein interaction network), drug information can produce scores of potential targets and help with prioritization.

Chemicals can be encoded into continuous latent vector space, which permits gradient-based optimization in molecular space and allows predictions using a graph-convolutional network based on binding affinity and other properties to find the drug target site.

The AI platform also relies on training computer vision and machine learning models on cryo-EM microscope data (2D structures) to understand the detailed spatial 3D structure of proteins and molecular complexes.

The selection of drug candidates requires a series of desired properties. These properties include pharmacokinetics, pharmacological and toxicity profiles. AI drug-designed algorithms that reinforce learning can be deployed for a simplified molecular-input line-entry system string, which is a specification in the form of a line notation for describing chemical species using short ASCII strings. It includes potential energy measurements, molecular graphs with varying weights for atoms or bonds, coulomb matrices, molecular fragments or bonds, atomic coordinates in 3D, etc.

2) Compound screening and lead optimization

Compound screening and the lead optimization step involve hits followed by lead identification to select drug candidates through combinatorial chemistry, high throughput screening, and virtual screening.

AI-based Virtual Screening is the compound database that is made by pulling mass quantities from publicly available chemogenomics libraries, which includes tens of millions of compounds annotated with information about their structure. This model uses Naïve Bayesian Classifiers, k-Nearest Neighbors, Support Vector Machines, Random Forests and artificial neural networks algorithms to enable medicinal chemists to efficiently find potential lead molecules among millions, speeding up initial stages of drug development.

Planning Chemical Synthesis with AI Retrosynthesis Pathway Prediction is an ambitious plan to exploit AI to automate chemical syntheses with minimal manual operation. Synthesis robots, combined with AI, can be used here. An AI platform named 3N-MCTS combines three different deep neural networks with Monte Carlo Tree Search for computer-aided organic compound synthesis and can select only well-known reactions by filtering out the most promising building blocks for the synthesis of target compounds.

For cell target classification, the AI model must be trained to quickly and automatically recognize the different features of cell types. The principal component analysis is employed to reduce the dimensions of the extracted features. AI-based methodologies, like the least-square support vector machine, can be trained to classify various cell types.

To accurately separate different cell types in the sample during the cell sorting step, Intelligent Image-Activated Cell Sorting devices are helpful, which measure optical, electrical, and mechanical cell properties through AI-based deep neural network algorithms.

3) Preclinical studies

Preclinical Studies or non-clinical studies are laboratory tests for new drug substances in in-vitro and in-vivo to establish their efficacy and safety profile.

To establish a molecular mechanism of action, an unsupervised approach of clustering-based machine learning tools evaluates RNA sequencing technologies, reducing the time to capture relevant significant quantities of biological information. It also uncovers many previously unexplored links between diverse stimuli and the cytokines they affect.

The pharmacokinetic/Pharmacodynamic ML Modelling approach works on in-vitro studies and preclinical PK studies to predict dose concentration (exposure) response relationship. In addition, ML models in Drug-Dose response improve efficacy predictions for effective multidrug combinations using only a few experiments.

In evaluating the toxicology profile of a compound, the most expensive parameter and time-consuming task, Deeptox Algorithm has already assessed more than 10000 environmental chemicals and drugs for 12+ different toxic effects in explicitly designed assays. As such, it can add value to drug development by accurately predicting the toxicity of compounds.

Deep learning algorithms used in ‘In-Silico’ methods can predict pharmacological properties by using transcriptomic data, including various biological systems and conditions.

4) Clinical trials

The development of AI tools for clinical trials would be ideal for recognizing patient diseases, identifying gene targets, predicting the effect of a designed molecule, and on and off-targets. One AI mobile application, when compared to traditional direct observation therapy, increased medication adherence by 25% in Phase II clinical trials.

AI in risk-based monitoring – a clinical trial monitoring technique that fulfills regulatory requirements but moves away from 100% source data monitoring – can significantly improve the conduct of clinical trials in all phases. In Phase II and III clinical trials, AI can identify and predict human-relevant disease biomarkers to select and recruit specific patient populations, which would increase the success rate of clinical trials.

Technical Obstacles and Prospects

A limitation in using AI to predict drug targets is translating traditional basic research performed in global labs into a language that a computer understands. In many cases, the data quality or balance is still an issue. While data augmentation techniques and improving image quality and variation are studied, this remains challenging.

The patent of AI solutions for drug discovery is an intense process. Security of medical/drug discovery data is also a big challenge, and proper security measures are essential.

In the present scenario, it is challenging to comprehensively test a new lead compound in combination with all available drug molecules. It takes thousands of studies to analyze known side effects and unknown interactions. However, if available, an AI algorithm approach would prove invaluable in further hastening drug development efforts.

AI will revolutionize how drugs are discovered and will reinvent the pharmaceutical industry.

Madhu Bala
Solution Architect, Wipro Holmes

Madhu focuses on applying AI methods and Automation techniques to solve business problems in the Healthcare industry. She is an expert in Natural Language Processing, Machine Learning and Deep Learning technologies. She has over a decade’s experience as a Scientist in Pharma R&D.