Vijayan Ramaswamy is an Institute Research Scientist at the University of Texas MD Anderson Cancer Centre.

Vijayan earned a PhD in molecular modelling and has worked with renowned scientists. 

His research involves using computational approaches for enabling cancer drug discovery.

INDIAai interviewed Vijayan to get his perspective on AI.

As a drug discovery scientist, it is unquestionably necessary to have a breadth of research knowledge. Could you tell us about what it needs to succeed in drug discovery?

Drug discovery is a long, expensive, and high-risk process. Successful drug discovery combines science and business and requires well-coordinated multi-disciplinary teamwork. To be successful as a drug hunter, one needs to be a jack of many trades and a master of at least one. Having in-depth knowledge in one's scientific domain is necessary to excel in your job. To lead cross-functional teams in a matrix environment and become a successful pharma R&D leader, one needs to have a breadth of scientific knowledge spanning the entire drug discovery continuum.

Could you describe your role as a research scientist at MD Anderson Cancer?

My job responsibility as a scientist is to help advance and accelerate our internal discovery programs by applying state-of-the-art computational and AI/ML methods. My typical day involves working closely with medicinal chemists and collaborating with scientists from cross-functional teams. I conduct predictive modelling using AI/ML techniques to forecast activity, toxicity, and other pharmaceutical endpoints of interest for small organic molecules. In addition, I use deep learning-based generative modelling algorithms to ideate and explore novel ideas for designing compounds. Besides, I also perform molecular simulations on high-performance computing clusters to simulate how small molecules bind to drug targets, how strong they bind, and how they modulate the motions of the target protein.

What challenges do you face as a drug discovery scientist working with AI?

There are many challenges in implementing AI into drug discovery. First, data is the bedrock for building any AI or ML model. Unfortunately, pharmaceutical data presents a significant challenge, as the data generated is often noisy, sparse, incomplete, and truncated. In reality, the quality and sparsity of labelled data restrict the implementation of AI in drug discovery. However, newer approaches like Few-Shot Learning (FSL) promise to tackle this problem. The other concern with using AI-based systems is the lack of transparency, as it's impossible to explain precisely why an AI/ML model arrived at this decision. This process is called a "black box" approach, lacking transparency and human interpretability. Explainable AI (XAI) will help address this explainability crisis in the foreseeable future. Apart from these technical challenges, the other challenge is managing expectations. Given the hype and inflated expectations created around AI, it's essential to sift the hyperbole from realism to keep expectations realistic.

How difficult would it be for a student of chemistry or biology to understand AI and apply it to their research?

I don't think it would be hard for chemistry or a biology student with a basic understanding of mathematics and statistics to learn AI and apply it in their research. Learning a new skill set is increasingly accessible nowadays, as many online resources and coding boot camps exist. From my own experience, I can tell you that I learned scripting languages like R, Python, and Shell on-the-job. Unfortunately, most non-technical students fear coding to be their biggest obstacle. However, open-source libraries available in Python for statistics, ML, DL, data mining, and data exploration obviates the need to write complex codes and handle memory allocation. Also, researchers can integrate many cheminformatics toolkits into various environments to solve chemistry-related problems. These low-code tools and libraries allow those without a technical coding background to create applications and write codes with minimal coding knowledge.

What, in your opinion, are the essential prerequisites for students interested in pursuing a career in AI?

The technical skills one requires to pursue a career in AI will vary, depending on the job. However, most of them require some common technical skills. In my opinion, an aspiring student needs to have a firm understanding of statistics and probability to understand complex algorithms and analyze data. Statistical learning theory provides the framework for most machine learning algorithms. Having an experience in statistics also helps you in evaluating machine learning models. AI is programming language agnostic. However, Python is the go-to choice as it offers many open-source libraries and frameworks that significantly democratize AI. Handling a high volume of data and processing complex numerical calculations requires a lot of computing power. GPU-based systems that provide robust parallel processing architecture have become commonplace in recent years and are to accelerate memory-intensive operations. Therefore, knowledge of GPU programming is indispensable if one aspires to become an AI/ML engineer. Given the need to handle and store structured data in some databases, SQL database skillsets are also necessary. Most database platforms are after SQL, so begin by learning the basics of SQL. With ever-changing technology, cloud computing is becoming the new standard as many organizations transfer their data, workflows, and infrastructure to cloud services. Hence, having a basic understanding of popular cloud-based deployment systems (such as AWS, GCP, Microsoft Azure, and IBM Cloud) is pertinent.

Could you recommend some good books and research publications on AI for life sciences?

A few books I would recommend for reading and getting up to speed are 

• Practical Statistics for Data Scientists: 50+ Essential Concepts Using R and Python 2nd Edition by Peter Bruce (Author), Andrew Bruce (Author), Peter Gedeck (Author)
• Python for Data Analysis by Wes McKinney
• Deep Learning for the Life Sciences by Bharath Ramsundar, Peter Eastman, Patrick Walters, Vijay Pande
• Chemoinformatics: Basic Concepts and Methods Thomas Engel (Editor), Johann Gasteiger (Editor)

In addition, many excellent review articles have on this topic. For example, a compiled list of interesting papers is available in the "Practical Cheminformatics" blog by Dr Patrick Walter. I also recently co-authored a review article titled "Enhancing preclinical drug discovery with artificial intelligence" with my collaborators. This review article in Drug Discovery Today's journal provides an overview of the many roles of AI in preclinical drug discovery.