Look behind the lecture Gurpreet Singh delves into machine learning in drug discovery

In this interview, we sit down with Gurpreet Singh, Head of Applied Machine Learning at Bayer (Leverkusen, Germany), to explore his talk on machine learning presented at the AI & Machine Learning in Drug Discovery & Development Summit (5–7 July 2022, virtual). Gurpreet shares his insights on AI in therapeutic target identification, the impact of AI in drug repositioning and much more in an interview not to be missed.

What sparked your interest in AI in medicine and how has the field progressed since?

The capacity of these algorithms to manage vast, multidimensional datasets sparked my interest in artificial intelligence. I pursued my doctoral thesis before the recent boom in computer vision with the ImageNet challenge. I was leveraging some of these techniques to construct a battery of machine learning-based tools to aid doctors in disease diagnostic differentiation. The methods utilized achieved a high level of prediction accuracy, especially when applied to photographs of the disease's earliest stages. The key research challenges addressed were high dimensional data processing, automated subject classification and disease trajectory determination.

Since then, the introduction of deep learning algorithms has transformed the landscape of machine learning applications in this and numerous other fields. Currently, clinical facilities are introducing tangible tools intended to enhance the detection and efficiency of radiology-related procedures.

How have recent advances in AI improved the accuracy of therapeutic target identification?

With the conclusion of the Human Genome Project in 2003, a vast amount of data was gathered in an effort to comprehend the role of individual genes in observable phenotypes. This was followed by a second wave of initiatives that sought to comprehend the function of transcriptomics and proteomics.

Simultaneously, advancements in the capacity of AI tools to handle big datasets, the increasing availability of computational resources and progress in the domain of graph-based deep learning networks have shifted the focus of research towards knowledge graphs. The purpose of these graphs is to create a much more comprehensive understanding of the disease network, human biology and disease–drug interactions than was previously known or achievable.

In my opinion, it is now much easier to realize a gene–disease relationship. However, human biology is much more intricate than that and hence understanding this relationship in the context of the right tissue and right modality still remains to be addressed. The completion of the single cell transcriptomic atlas through the Tabula Sapiens Consortium is a very promising beginning to address these questions.

How has AI impacted drug repositioning for rare diseases?

This area is quite promising. The most recent example of drug repurposing is the rediscovery of Halicin as an antibiotic by utilizing a graph neural network in 41 days. However, the timelines for bringing a repurposed medicine to market are comparable to those of a new drug. There are numerous aspects that must be enhanced before the total benefit can be realized.

How do you see the future of AI in medicine and surgery?

In this field, AI is beginning to make major strides. The shortage of physicians in nearly all specialties creates a compelling case for deploying AI as assistants to enable physicians to:

• Manage the patient load efficiently by focusing on the most urgent cases.
  ‍
• Improve patient care.  This reduces the false positive rates that have plagued the healthcare industry since the dawn of time.
  ‍
• Decrease the cost and increase the speed of patient service. The greatest influence from this will be seen in radiology.

Nonetheless, it has been observed that practitioners still have difficulty trusting the predictions made by these models. Therefore, it is necessary to:

(1) Improve the interpretability of these models so that physicians can have greater trust in them.

(2) There is a need for senior management to educate physicians on the advantages and limitations of these assistants.

(3) TAI companies and the FDA must progressively investigate the claims made regarding the capabilities and limitations of these systems.

My ideal future is one in which doctors may in person or remotely check the vitals  of a patient who is at home or on vacation with their smart devices. In the event of an emergency, the patient can visit a nearby clinic where their preferred physician can examine them in person or remotely through a hologram where robots enable physical exams. If a surgery is necessary, the doctor should be able to execute it in person or have a robotic arm perform it while they are at home or sipping margaritas on a beach in Florida without the danger of infection to the patient.

What were the main three take-home messages from your talk at the AI Drug Discovery & Development Summit?

(1) Current technologies for target identification lack depth and our grasp of human biology, particularly disease biology, is severely lacking.

(2) Although the future that AI technologies promise is spectacular, we are far from realizing its full potential. At best, we have only scratched the surface of the possibilities.

(3) Organizations must undergo change management to enable the emergence of organizations in which working with AI as an assistant is comparable to working with a colleague who is adept at some tasks but inept at others.

Interviewee profile:

I am Gurpreet Singh, the Head of Applied Machine Learning at Bayer Pharmaceuticals, where I support drug discovery. Previously, I was the Director of Data Science at GSK (London, UK), where I led initiatives to modernize the drug discovery pipeline through machine learning. My team was tasked with creating protein language models that can learn from protein sequence data to predict structure, function and evolutionary fitness. We utilized these models as in silico replicas for predicting function and designing new drugs. I led efforts to apply machine learning to research and development (R&D) workflows, such as suggesting novel vaccine candidates for polymicrobial antimicrobial resistance, optimizing drug formulation, and conducting immune profiling and repertoire analysis.

I have substantial experience leading multidisciplinary R&D teams. In addition, my work at Weill Cornell Medicine (NY, US) centered on utilizing machine learning to improve patient outcomes by utilizing trial data, registry information and databases of real-world evidence. I advocate for machine learning-based applications to improve patient care and have been published in leading medical journals. My PhD dissertation focused on utilizing machine learning to improve the early identification of Parkinson's and Alzheimer's disease utilizing imaging biomarkers.

‍The opinions expressed in this feature are those of the interviewee/author and do not necessarily reflect the views of Future Medicine AI Hub or Future Science Group.