Proteins are a fundamental building block of life, with a dizzying variety of structures and functions. Since the 1960s, scientists have understood how genes are translated into long chains of amino acids. But predicting how those amino acids fold together to create three-dimensional proteins has been a more elusive puzzle. Solving this puzzle was a task of considerable importance, since the shape of a protein determines the function (or dysfunction) of these molecular machines.

Then, in 2020, Demis Hassabis and John Jumper released an AI model (AlphaFold) capable of accurately predicting the structure of a protein based on its amino acid sequence. It transformed entire fields of research and medicine. Structural biologists rely on laborious methods such as x-ray crystallography or cryogenic electron microscopy, dedicating months and years to solving these spatial puzzles, one at a time. After many years solving these puzzles, structural biologists have now generated a large enough database of protein structures to train AlphaFold 2 to generate “astonishingly accurate” predictions of protein structure. In October 2024, the Nobel Prize in Chemistry was awarded to the creators of AlphaFold for this ground-breaking feat of computational biology.

The Promise for Rare Diseases

The full benefits of AlphaFold are still unfolding. One particularly promising area is the study and treatment of rare diseases. According to the U.S NIH, there are over 7,000 rare diseases affecting 30 million people or roughly 8% of the American population. Many of these diseases are seriously understudied, or ignored all together, falling into a class of pathologies colloquially referred to as “orphan diseases.” Research on rare diseases is frequently slowed by the twin bottlenecks of limited funds and insufficient manpower.

AlphaFold can help alleviate both constraints by streamlining efforts to understand the proteins responsible for driving rare disease pathogenesis. For example, mutations in the genetic instructions for a protein called alsin are implicated in several rare motor neuron diseases that affect children. But without an understanding of the protein’s structure, scientists had no foothold from which to study the disease. This changed within the last few years. Using AlphaFold, and other predictive algorithms based on AlphaFold’s code, researchers began to map this protein’s geography and highlight regions believed to be crucial to its physiological function. While this is only the first step of many towards treating these diseases, it creates a path forward from what had previously been a dead end.

Alsin is not alone in its structural mysteries. In fact, of the approximately 20,000 proteins found in the human genome, the Protein Data Bank only had validated structures for 7,074 of them (as of 2021). With AlphaFold, we now have the predicted blueprints for the remaining 13,000. And while AI-predicted models must still be verified, they can serve as invaluable launching platforms for rare disease research that has struggled to get off the ground. And even more importantly, AlphaFold can continue to provide momentum for these underfunded projects as they move into drug discovery.

The Impact on Drug Discovery

While there are many approaches to disease treatment, a significant portion of drugs rely on protein structure. These drugs can work to inhibit harmful proteins, restore functionality to beneficial proteins, and some drugs are even proteins themselves. Regardless of approach, the over-arching goals of drug development remain the same: specificity and efficiency. If the compound is not precisely targeted, the drug is likely to have off-target effects and severe side effects in patients. Of course, the ability to reliably target a protein is of little use if the compound does not induce a potent effect on protein activity. To find the right molecular key for unlocking therapeutic potential, companies invest millions of dollars combing through thousands of compounds, hoping to find the optimal fit.

For companies working in rare disease spaces, this can be an investment of time and money they simply can’t afford. Today, predictive algorithms offer a solution. An updated version of AlphaFold (AlphaFold 3), released at the beginning of 2024, introduced the ability to include other molecules (such as drug compounds) in the virtual mapping of protein structures. Companies can now use this software to screen new drugs virtually, avoiding the costly experiments that would have otherwise been necessary to rule candidates out.

In theory, this streamlines the screening process, narrowing the list of potential drug candidates from thousands to hundreds. AlphaFold 3 could be the lifeline that rare disease therapies need, bringing life-enhancing treatments out of the lab and into preclinical and clinical trials. However, while this much-anticipated achievement has brought new hope to drug discovery prospects, the reliability of these new hybrid structure predictions has yet to be tested.

Cautious Optimism

The Nobel Prize-winning AlphaFold 2 is an open-source software, meaning that the code underneath the software can be verified, modified, and understood. This transparency is crucial for computational and structural biologists, as it provides a measure of confidence in AlphaFold’s results. As chronicled by hilarious screenshots found across social media, AI is far from foolproof. The open-sourced nature of AlphaFold 2 allowed scientists to better understand the limits of the software, and to more deeply explore the areas in which it excels.

To the disappointment of many researchers, the code for AlphaFold 3 has not yet been released. In fact, as an open-letter to Nature points out, even reviewers of the AlphaFold 3 announcement article were not given access –– a break from the traditional peer-review process employed by Nature. However, it’s worth noting that the release of code for AlphaFold 2 was also delayed until several months after its debut, and DeepMind has said it intends to release the AlphaFold 3 code by November 2024.

For now, the reliability of AlphaFold 3 rests on the reputation of its Nobel Prize-winning predecessor. AlphaFold 2 was a major leap forward in structural biology. AlphaFold 3 appears poised to catapult the field once more into a world of exciting possibilities. These predictive algorithms can be a boon to underserved populations, and a powerful tool for unlocking new discoveries and breakthroughs. For rare disease research, it can jumpstart projects and streamline screening in the quest to define underlying causes and create viable treatments for the 7,000+ rare conditions that have been defined in the U.S. Combined with today’s improvements in precision medicine, manufacturing, and more, the future of rare disease is beginning to take shape.