Generative AI to Find the Elixir of Long Life

Di Gabriele De Palma

Among Silicon Valley entrepreneurs, the specter of an age-old idea looms—the elixir of life. Today, instead of being called amrita like in the Indian Vedas of the 15th century BC or ambrosia as with the ancient Greeks, it is referred to as anti-aging. While immortality may not be the goal, the hope is to develop products that can rejuvenate cells and extend both life expectancy and quality.

Peter Thiel, the entrepreneur behind PayPal and former director of eBay, has provided an apt description of this challenge: "Anti-aging innovation will repair the problems caused by aging just as a patch fixes a software bug." Thiel is one of the most active players in this field and has heavily funded the Methuselah Foundation (a reference to the biblical figure Methuselah). He is not alone—Jeff Bezos has also ventured into this space, supporting Altos Labs'cellular reprogramming project.

Now, a new key player enters the scene—none other than Sam Altman, CEO of OpenAI, who promises to revolutionize the field. OpenAI claims to have developed a language model capable of imagining proteins that transform normal cells into stem cells—totipotent cells capable of assuming various functions and developing into different tissues. This AI model has reportedly outperformed human scientists in this task.

This is OpenAI’s first model focused on biological data, 4b micro. The protein engineering project began over a year ago when Retro Biosciences, a longevity research company based in San Francisco, approached OpenAI for a collaboration. This was not a random connection—Sam Altman had already personally invested $180 million in Retro Biosciences.

Retro Bio’s stated goal, as per its corporate tagline, is to “extend human life by 10 years.” The company focuses on Yamanaka factors—a set of proteins discovered by Japanese Nobel laureate Shin’ya Yamanaka. When added to skin cells, these proteins transform them into stem cells.

Yamanaka factors are considered an ideal starting point for achieving tissue rejuvenation, which is why Altos Labs is also pursuing research in this area, parallel to Retro Bio.

However, this type of cellular reprogramming is currently far from efficient. The process takes several weeks, and less than 1% of treated cells complete the rejuvenation process.

GPT-4b micro was trained specifically to suggest new and more effective ways to redesign protein factors. Researchers at OpenAI and Retro Bio used AI-generated insights to modify two Yamanaka factors, increasing their effectiveness by more than 50 times—at least according to preliminary measurements.

According to researchers John Hallman and Aaron Jaech of OpenAI and Rico Meinl of Retro Bio, the AI-generated proteins appear to be superior to those scientists have been able to produce manually.

Until the publication of their findings, the scientific community remains doubtful because faith alone cannot be considered a scientific method.

At this stage, GPT-4b micro is not yet an official product. It currently exists as a custom demo for this specific project, and it is unclear whether it will evolve into an independent model or wheintegrate its functionalitiesto other, more general products.

The algorithm has been trained on a database of protein sequences from various living species and has been programmed with information about protein interactions. While this dataset may seem extensive, it is only a fraction of the data used to train OpenAI’s flagship chatbots, like ChatGPT. This makes GPT-4b not a large language model but rather a “small language model” designed for a specialized dataset.

It is still unclear how GPT-4b precisely formulates its hypotheses, as is often the case with AI models. Even its developers do not fully understand the underlying mechanisms. “It’s like when AlphaGo (Google DeepMind’s AI for the game of Go) defeated the world champion, but it took a long time to figure out why,” said Betts-Lacroix, CEO of Retro Bio. “We are still trying to understand what GPT-4b does, and we believe that our current application is just the tip of the iceberg in terms of its potential.”