Unveiling Nature’s Secrets: AI’s Journey Through 500 Million Years of Evolution to Create a Revolutionary Fluorescent Protein

Researchers are considering whether evolution might have developed in alternate ways. For example, was it unavoidable for humans to arise, or are we merely the result of a sequence of natural coincidences that could have never occurred, leading to an entirely distinct world? Although there is no conclusive response, artificial intelligence (AI) has now the ability to conduct evolutionary experiments. A recent study, published last week in the journal Science, unveils alternative routes in the construction of a certain type of protein that nature never ventured to explore. This technology may yield significant insights for creating new therapies and various applications

In his 1989 publication Wonderful Life, evolutionary biologist Stephen Jay Gould posed a philosophical inquiry: If the timeline of life’s evolution on Earth could be reversed to the beginning and replayed, would the conclusion be identical to the one we know, or something entirely different? Gould supported the latter viewpoint. Using video game analogies, he argued that evolution would have navigated a markedly different trajectory, and humans would not have appeared. “Replay the tape a million times … and I doubt that anything akin to Homo sapiens would ever evolve again,” he stated.

Gould’s proposition has ignited extensive debate ever since, with some endorsing determinism while others favor contingency. In his 1952 short narrative A Sound of Thunder, science fiction writer Ray Bradbury described how a time traveler’s seemingly trivial act of stepping on a butterfly during the age of the dinosaurs altered the future’s course. Gould expressed a similar sentiment: “Modify any early occurrence, even slightly and without clear significance at the time, and evolution cascades into a fundamentally different trajectory.”

The language of proteins

Scientists have been addressing this dilemma through experiments aimed at reconstructing evolution either in laboratory settings or in nature, or by comparing species that have arisen under similar circumstances. Presently, a new frontier has emerged: AI. In New York, a team of former researchers from Meta — the parent organization of social networks Facebook, Instagram, and WhatsApp — established EvolutionaryScale, an AI startup dedicated to biology. The EvolutionaryScale Model 3 (ESM3) system developed by the company is a generative language model — similar to the platform that drives ChatGPT. However, instead of generating text, ESM3 produces proteins, the essential building blocks of life.

ESM3 processes sequence, structure, and functional data from existing proteins to comprehend the biological language of these molecules and fabricate new ones. Its developers have trained it utilizing 771 billion data packets from 3.15 billion sequences, 236 million structures, and 539 million functional attributes. This amounts to over one trillion teraflops (a metric for computational capability) — the highest computing power ever applied in biology, as cited by the company.

“ESM3 represents a move towards a future in biology where AI serves as a tool for constructing from fundamental principles, akin to how we build structures, machines, and microchips,” states Alexander Rives, co-founder and chief scientist of EvolutionaryScale and leader of the recent study. He perceives biology as the most advanced technology ever developed, and one that is programmable using a universal alphabet — the genetic code, which is converted into amino acids, the building blocks of proteins. “ESM3 comprehends all this biological information, translates it, and articulates it fluently for generative purposes.”

The protein that never was

Rives and his team employed ESM3 to craft a new green fluorescent protein (GFP). GFP is a naturally occurring protein that emits a green glow under ultraviolet light and is frequently utilized in research as a marker. The initial GFP was identified in a jellyfish, though other variants can also be detected in corals and anemones. The scientists trained ESM3 to produce a new GFP, and the outcome astonished them: a fluorescent protein, which they designated esmGFP, showing just 58% similarity to the most closely related GFP. According to the researchers, this corresponds to simulating 500 million years of evolution. ESM3 is now available to the scientific community as a novel tool for designing proteins with therapeutic roles, environmental remediation abilities, and additional potential applications.

Thus, AI has uncovered a pathway that nature might have explored 500 million years ago, but for reasons unknown, did not pursue. Rives and his collaborators clarify that merely a few mutations in GFP can eliminate its fluorescence. However, ESM3 has identified a new domain of fluorescent proteins that could have existed yet never did. As they express, “At the foundation of these sequences lies a fundamental language of protein biology that can be interpreted using linguistic models.”

Rives and his team explain that only a few mutations in GFP can eliminate fluorescence; yet ESM3 has discovered a new area of fluorescent proteins that might have been produced but were not: “At the base of these sequences is a fundamental language of protein biology that can be interpreted using linguistic models,” Rives affirms.

As noted by Jonathan Losos, a professor at the University of Washington who investigates the notion of evolutionary rewinds by observing species in their natural habitats, “This research exemplifies brilliantly that there are numerous ways evolution could have occurred.” He regards the findings as supporting the idea of contingency, a perspective famously championed by Gould.

This perspective is also echoed by Zachary Blount, a professor at Michigan State University, who showcased the contingency of evolution in a noteworthy bacteria-growing experiment initiated in 1988 by his previous supervisor, Richard Lenski. The experiment continues to this day, encompassing over 80,000 generations.

“The study illustrates that there are viable biological avenues that haven’t evolved (as we believe) on Earth, suggesting real paths that evolution could have taken but didn’t due to the absence of necessary history,” Blount asserts. He also acknowledges the existence of some level of determinism in nature, evidenced by the fact that the ESM3 experiment is 42% similar to other GFPs. While Blount is skeptical that AI will ever completely resolve the rewind dilemma, he recognizes its capacity to assist in discerning what is contingent, what is not, and the reasons behind it. “It provides us with methods to investigate the realm of biological possibilities, enabling us to compare what’s biologically feasible with what exists or has existed.”

Subscribe to our weekly newsletter for more English-language news coverage from EL PAÍS USA Edition