[Reading Science] Is the Savior of New Infectious Diseases AI, Not 'Antibody Holders'?

Successful Improvement of Existing Antibody Therapeutics Performance Using Generative AI
Active Research on Developing New Antibody Therapeutics

The 2007 global blockbuster film I Am Legend. The protagonist, Will Smith, who survived thanks to possessing immune antibodies alone, strives to create a cure for the zombie virus using his own blood. It was similar during the recent COVID-19 pandemic. Scientists made all-out efforts to isolate antibodies from the serum of patients who had contracted and recovered from COVID-19 to develop effective treatments. However, in the future, such efforts may no longer be necessary. Scientists are opening the way to develop powerful antibody therapies against dangerous diseases like COVID-19 and Ebola virus more quickly and easily by using artificial intelligence (AI).

Virus. Stock image. Photo by Pixabay

On the 24th of last month, a research team from Stanford University School of Medicine in the United States published their findings in the international journal Nature Biotechnology. The team confidently announced that they developed an algorithm called 'neural networks' using the generative AI ChatGPT, significantly improving the efficiency and performance of antibody design. They claim that their pattern-learning-based generative AI, 'neural networks,' will accelerate the development of new antibody therapeutics and help create novel drugs that were difficult to develop using traditional design methods.

Antibody therapies are widely used for major diseases such as breast cancer and rheumatoid arthritis, generating over $100 billion in global sales annually. This is because antibodies are a key component of the human immune system's response to viral infections. Biotechnologists are working to reconstruct antibody proteins to create treatments that can address various diseases. The problem is that developing antibody therapies requires screening each candidate one by one, which demands enormous time, labor, and capital.

To shorten this process, the research team created 'neural networks,' which share the same structure as large language models like ChatGPT that have learned vast amounts of text. However, instead of text, they trained the model on about 10 million protein sequence structures. This is called a 'protein language model.' They also used a protein language model developed by the big tech company Meta to predict small-scale antibody mutations. As a result, neural networks demonstrated remarkable performance even after learning only a few thousand out of over 100 million protein structures. It proposed designs that enhance the ability of existing antibody therapies against COVID-19, Ebola virus, and influenza to recognize and block these viruses. Interestingly, the AI suggested protein structure modifications not in the regions that detect and respond to targets but in areas located outside those regions. Peter Kim, a professor at Stanford University School of Medicine who participated in the research, stated, "The information provided by this AI reaches areas that even antibody engineering experts do not know or generally fail to understand," adding, "I still have not figured out the reason behind this."

Scientists expect that the use of AI will not only improve existing antibody therapies but also lead to the development of entirely new drugs. Charlotte Dean, a professor of immunoinformatics at the University of Oxford, UK, said, "What has been developed this time is excellent for improving antibodies," but also expressed hope that "instead of simply improving existing antibodies, generative AI could be used to create completely new therapeutic antibodies." Suj Biswas, co-founder of Navla Bio in Boston, USA, said, "It could help develop drugs that effectively act on viruses for which antibody design has been difficult," adding, "For example, AI could assist in designing antibody therapies that target various objectives such as removing cell membrane-adherent G-protein-coupled receptors related to neurological disorders and heart disease, or binding to tumor proteins or immune cells capable of killing tumors."

There has been actual progress in such research. In March, researchers at the biotech company Absci in Vancouver, Canada, published research results on the preprint server bioRxiv, reporting progress in creating new antibodies using AI. They succeeded in newly designing several important regions of antibodies for breast cancer treatment using AI models based on protein sequence structures and experimental data. However, creating entirely new antibodies still faces many challenges. One major difficulty is that the protein structures, such as receptors used by antibodies to recognize specific targets, are very complex, making AI modeling extremely challenging.