Artificial intelligence technology applied to scientific research opens up new possibilities to accelerate the development of innovative treatments in the pharmaceutical industry.
One of the highlights of this story is AlphaFold , the foundational model and AI platform that enabled DeepMind and Google to make incredible strides in the biological realm. In particular, its developers set out to alleviate some of the key challenges researchers face in the scientific process; one of these highly relevant problems is the ability to determine the structure and sequence of proteins, the fundamental components of living organisms.
Although the 300 million proteins that exist on Earth are just combinations of around 20 basic amino acids, the sequence and folding of these elements are crucial to essential life functions. Therefore, the inspiration and creation of AlphaFold arose to unravel the mystery of protein folding and thus unlock new potential advances in science.
To achieve this, the scientists trained AlphaFold with nearly 100,000 known proteins. After two major iterations and in collaboration with Isomorphic Labs, AlphaFold 3 is the latest model released, which stands out for its high accuracy in understanding protein interactions and its modeling capabilities.
Dr. John Jumper , PhD , is one of the leading pioneers behind the development of AlphaFold, and enthusiastically describes the “AlphaFold story” as a significant milestone in science, especially with regard to understanding protein structure and folding – what once took years can now be completed in just minutes thanks to this platform.
Why is it important to the general public?
Because this technology has numerous potential applications. One of the most relevant is its enormous contribution to the pharmaceutical and drug discovery industries. A drug is, in essence, a small molecule that binds to a protein in the body in a specific configuration, and triggers a series of events that seek to attack a specific pathology.
AlphaFold and DeepMind’s work in this field has opened up unprecedented capabilities to identify these proteins for drug discovery and development purposes. As scientists gain a better understanding of protein structures and interactions, they can create more precise drug targets, better understand side effects, and delve into new areas of protein-drug interactions that were previously unimagined.
As Max Jaderberg , Director of AI at Isomorphic Labs , describes , this work marks a monumental chapter in human history, allowing scientists to «rationally develop therapies against targets that were previously difficult or considered intractable to modulate.»
Despite the incredible progress these organizations made, the process faced numerous challenges. The rapid growth of AI technology and foundational models produced a variety of concerns about reliability and trust, especially as they are increasingly used in crucial applications. Pushmeet Kohli , who founded the Trustworthiness Team at DeepMind and is now Vice President of Research at the company, explains that while some of the biggest problems in the life sciences can be solved with machine learning and AI, a lot of time and resources are also invested in ensuring that models are trained and developed as carefully as possible.
In particular, Kohli notes that he is always thinking about how to make the system more reliable in producing consistent and accurate results, and, most importantly, finding ways to implement safeguards that detect when systems make mistakes.
While Google and DeepMind are certainly not the only innovators in this space, they are probably among the most established players in the market. Scientists and developers around the world have recognized that artificial intelligence has immense potential in this field and are quickly working to create their own products. For example, Genomenon , a startup founded at the University of Michigan , has developed AI products that leverage genomic data to support precision diagnostics and therapies.
Facebook ‘s parent company , Meta , has also invested significant resources in this sector, developing its own basic protein folding model, called ESMFold . As of March 2023, the platform claimed that its latest update could predict nearly 772 million protein structures.
The lack of competitors to AlphaFold and the significant advancement of other market players are indicative of how challenging this work is, especially when setting such a high standard in terms of security and reliability. However, despite the challenges, this field has the potential to make a considerable impact on the world of medicine and healthcare.
