For the past five decades, protein folding has posed a significant challenge in the field of life science. However, a major breakthrough occurred in 2020 when Google DeepMind successfully addressed this long-standing issue with AlphaFold. This achievement not only marked a turning point but also opened the door to numerous possibilities for applying AI in drug development and broader healthcare applications.
How AstraZeneca is Using AI Models for Drug Development
Following the success of AlphaFold, other players entered the scene. Meta introduced ESMFold, while the Chinese biotech company Helixon pioneered OmegaFold. Generate Biomedicines contributed Chroma, and Baker Lab brought forth RoseTTAFold and RoseTTAFoldDiffusion, thereby expanding the range of innovative solutions in this domain.
AstraZeneca, based in Cambridge, stands out as one of the leading players in the application of AI in healthcare. AIM reached out to Siva Padmanabhan, Managing Director at AstraZeneca India, to understand the pivotal role that AI plays in reshaping medical science. It serves as a platform for discovering, testing, and accelerating potential medicines, including protein folding.
However, integrating AI into drug development comes with both advantages and challenges. With the vast amount of data available today, the key lies in effectively analyzing, interpreting, and applying this information.
“In our research and development efforts, AI plays a vital role in decoding extensive datasets to enhance our understanding of specific diseases, identify new medicinal targets, guide molecule synthesis, and improve predictions of clinical success,” explained Padmanabhan. He highlighted that the application of AI extends beyond the laboratory and encompasses various clinical approaches.
Consider the clinical trial process, where AI and ML tools are used to extract valuable insights from trial data. Proficiency in utilizing trial data for safety and efficacy analysis has been demonstrated, and efforts are underway to maximize the potential of previously collected data. AI also contributes to event adjudication in clinical trials, streamlining processes at different stages with the overarching goal of reducing overall time investments.
The Role of AI In Protein Folding
AI’s role in protein folding research encompasses multiple facets, offering potential benefits and addressing certain limitations. Protein structures, including disease-causing proteins and antibodies, provide valuable insights into the development of therapeutics, aiding in the design of more effective and safer drugs.
Traditional methods like X-ray crystallography, NMR, and cryo-EM are essential for determining the structures required for drug discovery. However, the emergence of powerful AI models has introduced an alternative approach, capable of accurately predicting challenging, expensive, and time-consuming protein structures that are difficult to determine experimentally.
In the realm of predicting protein structures, optimizing folding simulations, identifying new proteins and their functions, and designing novel protein structures, Large Language Models (LLMs) have shown promise. These models, trained on a vast repository of known protein structures, can propose protein sequences that enhance functionality and desired properties.
In the context of antibody drug discovery, these AI models can suggest sequences with strong binding to target proteins and improved developability characteristics. Leveraging publicly available antibody datasets, along with information about the proteins they bind to, serves as a valuable resource for building and refining models tailored to specific targets of interest.
AI-driven structure prediction facilitates the structure-guided discovery of small molecules, peptides, and antibody therapeutics. However, it’s important to note that existing AI models have limitations, primarily in predicting only the overall protein fold. Their ability to predict changes caused by single amino acid mutations is restricted, and these tools provide a static snapshot of the protein without insights into its dynamic nature.
According to Padmanabhan, interdisciplinary collaborations are crucial for advancing AI applications in protein folding research. Effective protein folding models require a diverse team, including data engineers, data scientists, structural biologists, and machine learning experts. Additionally, the adoption of federated learning, where models are trained on data from various pharmaceutical companies and research centers without exposing the data to other entities, holds significant potential in transforming this field.
