In late 2020, DeepMind announced that its AlphaFold 2 system had effectively solved the protein folding problem — a grand challenge in biology that had stumped scientists for over 50 years. At the CASP14 competition, AlphaFold 2 achieved a score of 87 on the hardest class of proteins, compared to just 58 two years prior, and beat its closest competitor by a staggering 26 points. In practical terms, it matched the predictive accuracy of experimental methods like X-ray crystallography — methods that cost around $120,000 per protein and take up to a year to complete. That's not hype. That's a genuine scientific leap.
Did DeepMind Actually Solve Protein Folding?
The word "solved" deserves some unpacking. What DeepMind achieved is that AlphaFold 2's computational predictions now rival the accuracy of slow, expensive experimental methods for determining protein 3D structures. This doesn't mean every protein is solved forever, or that all edge cases are covered — but for a large and important class of proteins, the system performs at a level once thought to require physical lab work.
02:15
CASP14 competition scores showing AlphaFold 2 at 87 vs nearest competitor — a 26-point gap
Watch at 02:15 →
To put the competitive performance in perspective: in 2018, the first AlphaFold scored 58 at CASP (the Critical Assessment of Protein Structure Prediction competition). In 2020, AlphaFold 2 hit 87 — and its next closest competitor was still 26 points behind. In a field where incremental progress is the norm, this is extraordinary.
There will likely be Nobel Prize-winning discoveries that emerge directly from work enabled by these computational methods. It's also possible we'll see the first Nobel Prize where a machine learning system played a central role in the underlying discovery — even if the prize itself goes to the humans who built and directed it.
What Is Protein Folding and Why Does It Matter?
To understand why this is such a big deal, you need to understand what proteins actually are and what they do. At the most basic level, proteins are chains of amino acids — in humans, there are 21 different amino acids that serve as the building blocks of life. These chains fold into complex 3D structures, and that 3D structure determines the protein's function.
05:40
Visualization of how an amino acid sequence folds into a 3D protein structure
Watch at 05:40 →
Proteins are the workhorses of living cells. They act as catalysts for chemical reactions, transport molecules across membranes, provide structural support, and regulate virtually every biological process your body runs. They are simultaneously the building blocks and the active agents of life.
Here's the fascinating part: a given amino acid sequence almost always folds into one specific 3D shape. Not sometimes — almost always. That's a weird and wonderful property. And the number of possible ways a protein could fold is staggeringly large — we're talking 10 to the power of 143 possible configurations, a number that dwarfs the estimated number of atoms in the observable universe (10 to the 80). This is known as Levinthal's paradox: how does a protein fold into its correct shape so quickly when the number of possible configurations is essentially infinite?
Nature solves this problem in milliseconds. Computational science struggled with it for decades.
Which Diseases Are Caused by Protein Misfolding?
One of the most direct reasons this breakthrough matters to human health is the link between protein misfolding and disease. When proteins don't fold correctly, the consequences can be severe. Misfolded proteins are implicated in Alzheimer's disease, Parkinson's disease, cystic fibrosis, type 2 diabetes, and many cancers, among others.
10:22
AlphaFold 1 architecture diagram: CNN input, distance matrix output, and gradient descent optimization step
Watch at 10:22 →
If we can understand the correct structure of proteins — and identify what goes wrong when they misfold — we open the door to designing drugs and treatments that target the root cause of these conditions rather than just managing symptoms. That's a profound shift in how medicine could work.
AlphaFold 1 vs AlphaFold 2: What Changed?
At the time of the announcement, DeepMind had not yet published a full paper on AlphaFold 2, so some of what follows is informed speculation based on the blog post and known trends in deep learning. But the general shape of the improvements is clear.
How AlphaFold 1 Worked
AlphaFold 1 was a two-step process. The first step used a convolutional neural network (CNN) that took the amino acid sequence as input — along with a rich set of features including multiple sequence alignments (MSA) of evolutionarily related proteins. The network outputted a distance matrix: a grid showing the predicted distances between pairs of amino acids in the final folded structure, with confidence distributions.
The second step was non-learning-based. It used gradient descent optimization to find a 3D folded shape that best matched the distances specified by that matrix. Effective, but limited.
What AlphaFold 2 Appears to Do Differently
The most significant architectural change in AlphaFold 2 is the move from convolutional neural networks to attention mechanisms and transformers — the same paradigm shift that has revolutionized natural language processing. If GPT-3 showed what transformers could do with text, AlphaFold 2 may be showing what they can do with biology.
There also appears to be a more integrated, iterative learning process. In AlphaFold 1, the multiple sequence alignment data was used as a feature engineering input — essentially pre-processed information fed into the model. In AlphaFold 2, the MSA seems to be part of the learning process itself, with information passing back and forth between the sequence representation and the residue-to-residue distance representation in an iterative loop.
The blog post also references a "spatial graph" representation rather than a simple distance matrix, suggesting the model may be capturing richer geometric relationships between amino acids. The overall lesson from recent deep learning history applies perfectly here: use attention, and make as much of the problem end-to-end learnable as possible. Both lessons seem to be baked directly into AlphaFold 2.
How Does AlphaFold 2 Work?
While the full technical paper wasn't available at launch, the core architecture appears to work roughly as follows:
- Input: Amino acid sequences plus evolutionarily related sequences gathered through multiple sequence alignment
- Processing: Transformer-based attention mechanisms that learn relationships between residues — both in sequence space and in physical 3D space
- Iterative refinement: Information cycles between the sequence-level representation and the pairwise distance/spatial representation, each pass refining the prediction
- Output: A predicted 3D structure for the protein, with high enough accuracy to rival experimental methods for most protein classes
The dataset context matters here too. We currently have about 200 million mapped proteins but only around 170,000 known 3D structures — because experimental methods like X-ray crystallography are so expensive and time-consuming. AlphaFold 2 could potentially expand that known structural database by orders of magnitude, fundamentally changing what structural biologists can study and discover.
What Are the Real-World Applications of AlphaFold?
The implications ripple outward in every direction from this breakthrough:
- Drug discovery: Designing proteins that correct misfolded ones — directly targeting the causes of diseases like Alzheimer's and Parkinson's
- Agriculture: Engineering insecticidal proteins or frost-protective coatings for crops
- Tissue regeneration: Self-assembling proteins for medical applications
- Biomaterials: Novel proteins for use in textiles and industrial materials
- Genomics: Identifying the unknown functions of genes encoded in DNA by predicting the structure — and thus function — of the proteins they encode
Looking further out, the long-term trajectory of this work points toward accurate physics-based simulation of entire biological systems — cells, organs, and beyond. If a protein is the simplest biological system, then the ability to model it precisely is the first step toward modeling vastly more complex systems. That's not science fiction; it's a logical extension of end-to-end learning applied to the real world.
Is AlphaFold 2 the Biggest AI Breakthrough Ever?
This is the kind of argument best settled over several beers, but let's lay out the contenders honestly. In the AI field, the short list of landmark breakthroughs includes: AlexNet and the ImageNet moment that launched the deep learning revolution in computer vision; AlphaGo and AlphaZero demonstrating superhuman performance through reinforcement learning and self-play; the transformer architecture and its applications in NLP, culminating in GPT-3; and real-world deployments like Tesla Autopilot and Waymo's autonomous driving systems.
AlexNet was technically a demonstration on a simplified image classification dataset — impressive, but not directly impactful on human lives. AlphaFold 2 is different. It operates on a real-world scientific problem with direct implications for medicine, agriculture, and materials science. On the axis of direct real-world impact, it's hard to argue against AlphaFold 2 being near or at the top of the list.
In structural biology specifically, this may be the most significant methodological advance in one or two decades. And in AI broadly, it demonstrates something important: that the tools being refined in game-playing and language modeling are now mature enough to make serious contributions to science itself. That transition — from AI as a tool for narrow tasks to AI as a partner in scientific discovery — may turn out to be the real headline here.
2020 was a difficult year for civilization. But breakthroughs like this are a reminder of what human ingenuity, backed by powerful computational tools, is still capable of achieving. The protein folding problem is not fully closed — but the door to understanding life at its most fundamental level just opened much, much wider.
