Artificial Intelligence (AI) for drug discovery and development has entered a growth phase. This presentation will focus on the commercial elements and call out major current and emerging applications of AI/ML within the sector. Benchmarking of the AI-enabled competencies of big tech and start-ups will be followed by a brief assessment of the recent trends, growth opportunities and strategic imperatives. An overview of the emerging ecosystem, partnerships, and selective case studies will be discussed. Key insights will be shared including: 1) themes driving AI application across the drug discovery & development value chain; 2) dynamics around regional trends, vendor eco-system, innovative business models and game-changing companies, and 3) investment opportunities and growth potential.

Current gene therapies are mostly limited to plasmid-based and adeno-associated virus variants with inefficient response rates and limited use. Better viral delivery methods would expand the available pharmacopeia to produce precision medicines for intractable and incurable genetic diseases. We are building VirCAD (Virus Computer-Aided Design), an HPC-powered, cloud-accessible, AI/ML-capable bioCAD platform to mine, design, model and test new viral drugs for improved cell and gene therapies, vaccines, oncolytics, antibiotics and other categories.

Drug discovery is experiencing a fundamental shift in terms of evolving from in silico to in machino which refers to novel molecules designed by machine learning technologies. In recent years, the role of data science in drug discovery is evolving rapidly. We expect this evolution to continue at a rapid clip for the foreseeable future. In this talk we will discuss tools, technologies, and languages that data scientists are using to gain extra insights and value from biomedical data. The presentation will explore them in the context of three different data analytics domains - descriptive, predictive, and prescriptive.

Studies of human genetics have largely focused on populations from WEIRD (Western, Educated, Industrialized, Rich, Democratic) countries which has resulted genomic insights that are not generalizable to all populations. Most studies, trials and papers conclude with a call to action to recruit and use more diverse genomes, and yet the proportion of non-European ancestries in genomic studies is diminishing. To address this gap, we must work across the whole pipeline of genomic research and health care delivery, from the populations we work with and the data we collect, to the analyses we carry out and the availability of genetic testing. This talk will cover the complex challenges taking an end-to-end approach to diversifying genomics involves, and what we might do to reduce bias in our data-driven systems in precision medicine.

Finding druggable targets for complex chronic diseases is particularly difficult using existing drug discovery approaches. This leaves significant pockets of patients with unmet medical needs, but also many opportunities to reposition active molecules originally designed to treat another disease, which act on mechanisms relevant to the needs of currently ineffectively treated patients. High-resolution patient stratification based on combinatorial analytics offers an accurate and scalable route to map effective drugs to patient subgroups across multiple new disease indications, providing a faster, cheaper, and derisked route to their approval.

Precision dosing, or delivering the right dose for every patient, is a key enabler of precision medicine. However, nowadays recommended dose for most of medicines is not personalized. There are many valid reasons behind the "one dose fits all" approach of today's practice for testing and finally recommending medicine dosage both during R&D (lean development) and marketed use (minimal complexity of drug usage). Exploring precision dosing in R&D is key for leveraging the full potential of a therapeutic intervention, enhancing the change of development success and providing a therapeutic benefit personalized to each patient's needs. To explore precision dosing in R&D, methodologies needs to be developed and tested, in particular, those which can take advantage of highly dimensional biomarkers characterizing today's era of digitalization. Learning from multidimensional data to generate dynamic rules is a primary benefit from reinforcement learning. In special cases such as sedation in intensive care units, reinforcement learning is applied and works by automatically adapting dosing through continuous monitoring of patient’s consciousness within a feedback loop process. We further explore the problem of optimal control of sedation with Propofol by considering inherent factors characterizing model-based drug development. In particular, the opportunity to use pharmacokinetic-pharmacodynamic (PKPD) model as a simulation engine to generate knowledge from dosing scenarios which cannot be tested experimentally; the partial knowledge of patient’s drivers of drug effect; and finally, the presence of difference level of variability in the data. Through simulation, we show that learning purely from retrospective trial data, even with a sample size up to 500 patients, gives worst precision than heuristic dosing strategies. However, embedding PKPD model of drug effect within the RL algorithm provides best accuracy. Extending the definition of reinforcement learning agent state leads to further increase accuracy of optimal dosing policy. In this study case, residual variability has a large impact on the efficiency of the algorithm. Model-based reinforcement learning constitutes a promising approach for precision dosing with a need for additional research to realize its potential for application to real data from clinical settings and drug development.

We have created one of the largest resources related to COVID-19 by compiling datasets from 50 publicly available sources [https://tinyurl.com/DrugX-pipeline]. An expert curation strategy was combined with a text mining system to screen over 552700 abstracts that are relevant to SARS CoV-2. We obtained 107 interactions amongst the SARS-CoV2 genes, 820 interactions between SARS-CoV2 genes and drug molecules, 4045 interactions between drug molecules and side effects and 8403 interactions amongst the drug molecules. Utilizing all these interactions, we constructed 5 heterogeneous networks covering the molecular pathophysiology of COVID 19. These networks include CoV-Human, CoV-CoV, drug-CoV, drug-drug, and drug-side-effect. We also obtained a scale-free multi-layer network of 189 nodes and 554 edges, composed of 20 SARS CoV-2 genes, 148 human genes, 7 drugs, and 14 side effects. Further, we deployed an automated docking pipeline to conduct large-scale docking studies (~1 million docking experiments) involving several drugs and targets to identify network-based binding propensities [Rawal et al 2022 A]. Next, we developed a multi-modal pipeline named Cov-DrugX (http://drugx.kamalrawal.in/drugx/) to repurpose drugs for COVID-19 treatment. It incorporates several independent modules used for ranking and characterizing the drugs.  Each module utilizes data from diverse sources, such as the DGIdb, DrugBank, SIDER, etc. We have labeled these modules as DL-200, D-Side_Effect, D-Target, D-Gene_Expression, D-Dock, D-Circuit, D-AI_ranking, and D-Symptoms. We have incorporated a deep-learning (DL) system (multi-layered neural network) in modules such as DL-200 and graph neural network in D-AI_ranking. A new scoring mechanism is introduced to standardize heterogeneous outputs generated by different modules. We have shown applications of COV-Drug-X on COVID-19 related drugs such as molnupiravir, ritonavir, cilostazol, and rabeprazole. Further, we have conducted screening of 2700+ FDA-approved drugs and identified ramipril, imatinib, mitoxantrone, and rifampicin as the top-ranking hits. Next, we also screened large chemical and drug libraries using our platform. This exercise has revealed cyclosporine as one of the top-ranking candidates. To support our predictions, we created evidence against each top-ranking hit using the literature search [Verma and Rawal et al 2022].