Artificial intelligence has spearheaded lasting change in the healthcare industry, with arguably the biggest breakthrough being AlphaFold. Created by Google DeepMind, AlphaFold is a deep neural network that predicts a protein’s structure simply by analyzing its amino acid sequence. The system gained global recognition in 2020 when it solved the “protein-folding problem,” later winning the Nobel Prize in Chemistry in 2024.
AlphaFold, Explained
AlphaFold is an algorithm that gained international recognition in 2020 when it cracked the “protein-folding problem,” accurately predicting protein structures based on their amino acid sequences. By calculating predictions in a matter of seconds, AlphaFold promises to accelerate drug discovery and development, improve disease modeling and transform the healthcare industry in other ways.
Cracking the mystery of protein structures has unlocked a range of possibilities, from accelerating drug development to producing new treatments for diseases. Below, we’re taking an in-depth look at AlphaFold, including how it works, its significance, its potential upsides and some limitations to keep in mind.
What Is AlphaFold?
AlphaFold is an advanced algorithm that can predict the three-dimensional structure of a protein by studying its one-dimensional amino acid sequence. In particular, it is an artificial neural network, consisting of multiple layers of simulated nodes that mimic the behavior of neurons in the human brain. Working together, these nodes can process massive data sets and reference them to uncover complex patterns in new data.
These abilities made AlphaFold ideal for solving the intricacies of amino acid sequences, enabling it to accurately predict a protein’s structure in a matter of seconds. For peace of mind, AlphaFold also shares a confidence score alongside its predictions, giving users a sense of how reliable its calculations are.
Why Protein Structure Matters
Knowing a protein’s structure is crucial because its function is closely linked to its shape. The ability to predict protein structures then allows researchers to gain a deeper understanding of how a protein works and the exact role it plays in a cell.
After all, proteins are the building blocks of life, enabling cells to form and maintain their structure. Any genetic-level mishaps during protein folding could disrupt cell development, leading to disease or even death for an organism. Predicting protein structures not only provides insights into how proteins function in general, but also aids research in areas like drug discovery and genetic diseases.
The Problem With Protein Folding
A protein is made up of 20 amino acids linked together. While these amino acids can assemble into a protein in mere milliseconds, processing all the possible combinations of amino acids and predicting the shape of the protein is incredibly complicated. In fact, it would take the lifespan of the universe to complete this calculation. This problem is known as Levinthal’s paradox — named after Cyrus Levinthal, who discovered the issue in the 1960s. It doesn’t help either that little is known about the actual process of how proteins fold into their final structures.
In 1972, biochemist Christian Anfinsen contributed an important addition, though, claiming that a protein’s structure depends on its exact sequence of amino acids. In other words, researchers should be able to use a protein’s amino acid sequence to predict its structure, even without understanding the process of protein folding. This hypothesis opened the door for AlphaFold to leverage AI and navigate the computational complexity of protein folding.
How Does AlphaFold Work?
To handle the vast number of amino acid combinations, AlphaFold was trained on several large, publicly available data sets. These include the Protein Data Bank, which contains more than 256,000 protein structures, and UniProt, which lists nearly 150 million protein sequences. Using machine learning and deep learning, AlphaFold can reference these databases to recognize patterns in new amino acid sequences.
DeepMind researchers built AlphaFold with a transformer architecture as well, so it specializes in sequence-to-sequence tasks. In this case, AlphaFold houses a transformer called Invariant Point Attention, which excels at solving problems related to 3D structures.
Despite its extensive training data and design, AlphaFold must still calculate two variables to predict protein structures: The distance between each pair of amino acids and the angles of the chemical bonds between them. For distance, AlphaFold generates a range of possible distances with varying probabilities, combines them into a performance score and applies gradient descent to refine its answers and gradually improve the accuracy of its predictions.
AlphaFold vs. AlphaFold 2 vs. AlphaFold 3
Google DeepMind researchers introduced the original version of AlphaFold at the thirteenth Critical Assessment of Protein Structure Prediction (CASP) competition in 2018. It demonstrated the ability of neural networks to predict protein structures by using the calculations of three neural networks to produce accurate predictions. Since then, AlphaFold has made massive strides over its next two generations:
- AlphaFold 2: In 2020, AlphaFold 2 solved the protein-folding problem that had stumped scientists for 50 years. The model upgraded its predecessor’s architecture to a neural network called Evoformer, which processes multiple sequence alignments to predict the structures of single proteins.
- AlphaFold 3: AlphaFold 3 goes a step further than AlphaFold 2, replacing the Evoformer with a Pairformer architecture and a diffusion network to predict the structures of DNA, RNA, ions and other molecules besides proteins.
Another key difference between AlphaFold 2 and 3 is their availability. AlphaFold 2 has an Apache 2 license, allowing it to be used for both academic and commercial purposes. While the source code for AlphaFold 3 also has this license, the model’s trained weights and parameters are not open source. Instead, AlphaFold 3 is subject to terms of use that restrict it to non-commercial contexts for now.
What Is AlphaFold Used For?
AlphaFold has revolutionized various scientific fields, from drug discovery to environmental conservation:
- Drug Development: With improved knowledge of proteins, scientists can quickly identify which proteins are ideal drug targets and develop customized medicines.
- Disease Modeling: AlphaFold has helped model proteins in heart disease, leishmaniasis and Chagas disease, accelerating research on potential treatments.
- Pandemic Response: AlphaFold predicted protein structures of the Covid-19 virus, contributing to efforts to design an accurate test.
- Genetic Diversity: AlphaFold can trace the relationship between DNA changes and physical traits, shedding light on neurodivergent conditions like autism.
- Environmental Conservation: Researchers are using AlphaFold to study a protein in honey bees that enables them to better handle stress and disease, with the eventual goal of promoting a more genetically healthy honey bee population.
Benefits of AlphaFold
Previously, researchers spent decades and millions of dollars on equipment to decipher protein structures. AlphaFold can predict these structures in seconds, making the process much cheaper and faster. DeepMind has also open-sourced the code behind AlphaFold and provided free access to its protein structure database of more than 200 million predictions, allowing all scientists to enhance their research with AlphaFold’s findings.
Broader knowledge of the proteins that make up the human body has spearheaded faster drug discovery, bolstered pandemic responses and strengthened research around neurodivergence. Without AlphaFold, scientists would have to rely on methods that take months or even years, severely hindering the quality of care that society has come to expect.
Limitations of AlphaFold
While AlphaFold is a major step forward in fully comprehending proteins, it’s still inherently flawed. It inherits any missing values and errors from its training data, making it susceptible to hallucinations that impact its accuracy and reliability.
Another shortcoming is that AlphaFold’s predictions are based on static snapshots of proteins, which are actually dynamic and constantly evolving within an organism. AlphaFold then fails to account for chemical modifications, mutations and other changes that proteins undergo. It has a hard time with edge cases, too, like intrinsically disordered regions — parts of a protein that don’t have a clear structure. As a result, its scope remains limited.
Frequently Asked Questions
What is AlphaFold?
AlphaFold is an algorithm built by Google DeepMind that famously solved the “protein-folding problem” in 2020. Trained on several massive data sets, AlphaFold uses its vast knowledge of proteins to identify patterns in new amino acid sequences. It then calculates the distance between each pair of amino acids and the angles of their chemical bonds to predict protein structures. This breakthrough has led to advances in drug development, disease modeling and environmental conservation, among other areas.
Is AlphaFold free?
It depends. AlphaFold 2 is freely available for academic and commercial use under an Apache 2 license. Researchers can also access AlphaFold 3’s features via the AlphaFold Server database, and its trained weights and parameters are available on GitHub — but the model can only be used for non-commercial purposes.
What is AlphaFold 3?
AlphaFold 3 is the latest version of the AlphaFold algorithm, succeeding earlier models that focused on predicting the structures of single proteins. By combining a neural network called Pairformer with a diffusion network, AlphaFold 3 can accurately predict the structures of more complex proteins and other molecules like DNA, RNA and ions.
Does AlphaFold replace lab experiments?
No. While AlphaFold’s predictions support hypotheses and accelerate research involving protein structures, scientists still need to conduct experiments and test its predictions to ensure the most accurate and reliable results.
How is AlphaFold used in drug discovery?
AlphaFold has provided the scientific community with in-depth insights into how protein structures work. Researchers have then leveraged this broader knowledge to more accurately target specific proteins when repurposing existing drugs or designing new ones, accelerating the drug discovery and development process.
