AlphaFold 1 was DeepMind’s first major success in protein structure prediction and, looking back from 2026, also serves as a particularly interesting example of how to combine deep learning, bioinformatics, statistics, and numerical optimization within a single system. Before AlphaFold 2 revolutionized the field in 2020, the initial version had already demonstrated at CASP13 that a neural network could learn distance distributions between residues and then use them to reconstruct 3D structures with markedly higher accuracy than competing methods.
Key points of AlphaFold 1 in 20 seconds
- DeepMind started working on protein structure in 2016 and presented AlphaFold at CASP13 in 2018.
- The first version predicts distance distributions between residue pairs based on sequence and evolutionary signals.
- It then converts these probabilities into a differentiable potential and searches for compatible structures using optimization.
- At CASP13, it achieved high-accuracy models in 24 of 43 free modeling domains, compared to 14 by the next best method.
- AlphaFold 2 later replaced much of this pipeline with a more integrated neural architecture.
For a technical reader, AlphaFold 1 has an instructional advantage over its successors: its components are quite clearly distinguishable. The system does not attempt end-to-end learning of three-dimensional geometry. Instead, it first constructs sequence-derived representations and related proteins, then uses a neural network to estimate structural constraints, and finally feeds those constraints into an optimization process.
In some ways, it is a hybrid architecture.
From AlphaGo to CASP13: how AlphaFold started
Google DeepMind traces the project’s origin to March 2016, after AlphaGo’s victory over Lee Sedol. The company assembled a small team to explore whether machine learning techniques could apply to one of biology’s classic problems: predicting a protein’s 3D structure from its sequence.
Two years later, a major public demonstration was achieved.
AlphaFold competed in CASP13, held in 2018. CASP, Critical Assessment of protein Structure Prediction, is a blind evaluation created in 1994, where teams are given sequences whose structures have not yet been published and are asked to predict them.
AlphaFold ranked first.
The technical paper was published later in Nature in January 2020 under the title Improved protein structure prediction using potentials from deep learning. It describes the system used in CASP13 and highlights a clear difference: AlphaFold obtained models with TM-scores ≥ 0.7 in 24 of 43 free modeling domains, compared to 14 for the second-best method.
This was significant because free modeling targets are exactly those for which there is no close structural template available to copy or adapt.
Input still a sequence of characters
From a computational perspective, a protein can initially be represented in a surprisingly simple way.
Each protein is made of an amino acid sequence. The 20 standard amino acids are typically represented by single-letter codes:
A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y
A small protein might look like this:
NLYIQWLKDGGPSSGRPPPS
This is the sequence of Trp-cage, a small peptide often used as an example of folding.
The challenge is transforming this one-dimensional string into a three-dimensional structure.
AlphaFold 1 doesn’t simply predict XYZ coordinates for each atom. Its core idea is to first infer what distances should exist between residue pairs.
For a protein of length L, this naturally leads to an L × L matrix of relationships.
Each position (i,j) contains information about the possible geometric relationship between amino acid i and amino acid j.
MSA transforms evolution into training data
A key component is the Multiple Sequence Alignment (MSA).
For a target protein, AlphaFold searches for evolutionarily related proteins in large sequence databases. DeepMind used resources like Uniclust30 and other sequence collections, along with tools such as HHblits.
The core idea is quite elegant.
Suppose a particular position in a protein changes over evolution. If another position tends to change at the same time, there might be some relationship between them.
Like:
A L K G T A L K G T B L R C T B L R C T
The first and fourth positions seem to evolve in a correlated manner.
This might occur because the residues end up close together in space. If a mutation alters the chemistry of one, a compensatory mutation may be needed in the other to maintain structural stability.
Not all correlations imply contact, but millions of sequences contain enough statistical signal to make these relationships useful.
AlphaFold 1 leverages this coevolution data to build pairwise features. In the original paper, couplings between 20 amino acids plus a gap state produce 21 × 21 matrices, totaling 441 values for each pair of positions.
Protein as an image with hundreds of channels
This introduces a very helpful analogy for machine learning engineers.
The protein can be represented as a kind of square image:
L × L × features
Each cell (i,j) encodes information about residue pair i,j, with channels containing data from the MSA, sequence profiles, and other signals.
AlphaFold 1 processes this representation via a deep residual convolutional neural network with dilated convolutions.
This makes sense if you think of the matrix as an image.
A CNN can integrate local information, and through depth and dilation, propagate signals over large distances within the pair matrix.
The most important output is a distogram.
Instead of predicting a single distance, the model predicts a distribution over distance intervals for each residue pair.
Conceptually:
residue 21 ↔ residue 874-5 Å 0.045-6 Å 0.116-7 Å 0.487-8 Å 0.248-9 Å 0.08...
The network indicates which geometries are much more plausible than others.
This is far more informative than a binary contact/non-contact prediction typical of earlier methods. The Nature paper emphasizes that predicting distances offers more structural insight than just contact prediction.
The second output aids torsion angle prediction
AlphaFold 1 also predicts backbone torsion angles.
The backbone bonds can’t rotate arbitrarily. Two key degrees of freedom are represented by the φ (phi) and ψ (psi) angles.
Instead of predicting them independently, the model can learn distributions reflecting which angle combinations are chemically common or plausible.
These predictions serve as starting points for the next step but are not the final structure.
And this highlights a fundamental difference with AlphaFold 2.
The neural network doesn’t finish the job: L-BFGS steps in
After obtaining the distance distributions, AlphaFold 1 constructs a statistical potential.
The idea can be simply stated.
If the network predicts two residues should be about 7 Å apart, a structure placing them at 25 Å will incur a high cost.
Placing them in a high-probability zone results in a lower cost.
For each pair (i,j), a function similar to this can be derived:
V_{ij}(d) = -log P_{ij}(d)where P_{ij}(d) is the predicted probability that residues i and j are separated by distance d.
Since the original distributions are discrete, AlphaFold uses interpolation to create continuous functions suitable for gradient-based optimization.
The contributions from many pairs are summed, along with other structural terms, to produce an overall potential:
minimize V(φ, ψ)
This means finding the set of angles that yields a structure with the lowest potential value.
The system employs L-BFGS, a quasi-Newton optimization algorithm designed for large problems without requiring a full Hessian matrix.
For machine learning practitioners, this architecture is especially appealing because it separates two problems:
- The network learns which geometries seem probable.
- The optimizer searches for a structure satisfying those probabilities.
Not everything is solved inside the neural network.
The potential corrects background biases
Here's an interesting technical detail.
Residues in a protein don't have a completely random distance distribution even without sequence information. Chain connectivity and physical constraints introduce biases.
AlphaFold 1 accounts for some of this by comparing the learned distribution for a specific protein with a reference distribution based on general properties like sequence length.
The idea is akin to calculating how much more likely a particular distance d given that sequence is versus what would be expected from overall protein geometry.
This approach helps convert neural predictions into a more usable signal for the optimizer.
Training on 31,247 proteins
The original work used known structures from the Protein Data Bank (PDB), carefully splitting datasets to prevent overly similar proteins from contaminating evaluations.
The Nature paper details the PDB versions, CATH classifications, and sequence databases used, including the specific splits for experiments.
This is crucial in computational biology.
A random split can be misleading because proteins from the same family may share similar structures. Evaluation must measure whether the model generalizes to unseen folds that are just different sequences.
CASP adds an external validation layer: predictions are made before experimental structures become public.
AlphaFold 2 was not just a bigger AlphaFold 1
After CASP13, DeepMind redesigned almost the entire system.
AlphaFold 2 competed in CASP14 in 2020, achieving a much larger leap. DeepMind clarified that the techniques used were new compared to version 1.
The architectural differences can be summarized as:
AlphaFold 1
MSA → features → residual CNN → distances + torsions → potential → numerical optimization → structure
AlphaFold 2
MSA + pairwise representation → Evoformer / attention → Structure Module → 3D coordinates → recycling
In AlphaFold 2, geometry is much more integrated within the neural model itself.
The Evoformer propagates information between sequence and pair representations, while the Structure Module directly generates 3D structures, and the system can recycle its predictions for refinement.
This approach led CASP14 to declare that AlphaFold achieved unprecedented accuracy, and it became the breakthrough for which Demis Hassabis and John Jumper would later receive half of the Nobel Prize in Chemistry in 2024, shared for protein structure prediction.
Quick timeline of AlphaFold
Highlights include:
2016. After AlphaGo, DeepMind forms a small team to explore protein structure prediction.
2018. AlphaFold participates in CASP13 and ranks first.
2020. Nature publishes the architecture used in CASP13: deep residual networks predicting distance distributions converted into potentials.
2020. AlphaFold 2 debuts at CASP14 with a redesigned, more accurate architecture.
2021. AlphaFold 2 is published in Nature with open-source code and models released.
2021-2022. DeepMind and EMBL-EBI launch the AlphaFold Protein Structure Database, providing predictions for over 200 million proteins.
2024. AlphaFold 3 expands to include protein complexes, nucleic acids, ligands, and ions.
2024. Demis Hassabis and John Jumper jointly receive half of the Nobel Prize in Chemistry for protein structure prediction. David Baker receives the other half for computational protein design.
Why AlphaFold 1 still merits study
No one in 2026 would choose AlphaFold 1 over AlphaFold 2 or 3 for structure prediction.
Its value lies elsewhere.
For a developer, it’s an almost ideal architecture to see how a scientific problem can be decomposed using different tools.
There are biological databases.
Sequence searches.
Feature engineering.
Probabilistic models.
Deep residual CNN architectures.
Differentiable geometry.
L-BFGS optimization.
Embedded physical knowledge in the objective function.
AlphaFold 1 also shows that scientific AI doesn't need to replace all prior knowledge with a giant network.
Deep learning works here because it builds on decades of bioinformatics, turning evolution’s regularities into mathematical constraints that are usable.
This combination proved enough to win CASP13.
Two years later, DeepMind almost rebuilt it from scratch and achieved with AlphaFold 2 the breakthrough that the Nobel later described as the practical resolution of a half-century-old problem.
Frequently Asked Questions
What exactly was AlphaFold 1?
It was the system presented by DeepMind at CASP13 in 2018. It used deep neural networks to predict distance distributions between residues and converted those into potentials for 3D reconstruction.
Did AlphaFold 1 use Transformers?
Not in the way AlphaFold 2 does. The first version mainly relied on deep residual convolutional networks and features from multiple sequence alignments.
Why was predicting distances better than contacts?
A contact prediction only indicates whether residues are likely close, but a distance distribution provides richer information about geometry, leading to more precise constraints for structure modeling.
Was AlphaFold 1 the system that received the Nobel?
No, the 2024 Nobel in Chemistry honored Demis Hassabis and John Jumper for protein structure prediction—with particular emphasis on AlphaFold 2's contributions in 2020. AlphaFold 1 was the precursor that won CASP13 and set the stage for this evolution.
Source: Noticias Inteligencia Artificial
