CASP13

For decades, structural biology was locked in a painfully slow, expensive waiting game.

Determining the 3D shape of a single protein required months—sometimes years—of grueling lab work using X-ray crystallography or cryo-electron microscopy. This bottleneck wasn’t just an academic headache; it directly stalled drug discovery, disease research, and our fundamental understanding of life.

Then came CASP13 in December 2018.

The Critical Assessment of Structure Prediction (CASP) is the ultimate blind test for protein folding. Every two years, organizers release amino acid sequences of proteins whose structures have been experimentally determined but not yet made public. Computational teams from all over the world compete to predict how those sequences will fold.

Historically, progress in CASP was measured in inches. In 2018, an outsider took a mile.

The Pre-CASP13 Landscape

To understand why CASP13 caused shockwaves, we have to look at how protein structure modeling used to work. Teams generally fell into two camps:

• Template-Based Modeling: If a target protein looked similar to a known protein already in the Protein Data Bank (PDB), algorithms could map the new sequence onto the old structure.
• Free Modeling (Ab Initio): If a protein shared no evolutionary history with known structures, algorithms had to predict the shape from scratch based entirely on physics and evolutionary contact patterns.

For nearly a generation, Free Modeling accuracy lagged behind. Algorithms routinely failed to resolve complex folds, leaving a massive gap between the proteins we could sequence and the proteins we could actually visualize.

Enter AlphaFold: The Day Everything Changed

At CASP13, London-based AI lab Google DeepMind entered the competition for the first time with a system called AlphaFold.

The results weren’t just a narrow victory; they were a total routing of the traditional structural biology paradigm. AlphaFold secured first place in the rankings, vastly outperforming veteran academic groups.

“What just happened?” became the defining quote of the conference, famously penned by computational biologist Mohammed AlQuraishi in his viral post-mortem blog.

The Global Distance Test (GDT)—the metric used to evaluate prediction accuracy where 100 represents a perfect match—showed AlphaFold achieving unprecedented accuracy on difficult “Free Modeling” targets.

Predictor Group	Median GDT (Hard Targets)	Approach Type
AlphaFold (DeepMind)	~60-70+	Deep Neural Networks (Spatial Distances)
Runner-Up Academic Teams	~45-50	Traditional Fragment Assembly / Co-evolution

While a score of 90+ is considered competitive with physical lab experiments, AlphaFold’s jump into the high 60s and 70s on previously “un-modelable” targets was the largest single-iteration leap in CASP history.

How Did the AI Do It?

Prior to CASP13, deep learning in biology was largely used to predict simple contact maps—binary guesses of whether two parts of a protein chain touched.

DeepMind’s engineering team shifted the paradigm by training a deep neural network to predict continuous physical distances between pairs of amino acids rather than binary contacts. They converted these distance distributions into a geometric potential score, which was then optimized using a traditional gradient descent algorithm.

Instead of trying to mimic the physical timeline of how a protein folds in a cell, AlphaFold treated the challenge as a massive geometric data-puzzle.

The Legacy of CASP13

It is rare that a single scientific conference marks a definitive “before and after” boundary for a discipline, but CASP13 did exactly that.

• The Blueprint for AlphaFold 2: The lessons and data from CASP13 directly informed DeepMind’s complete architectural redesign that led to AlphaFold 2 at CASP14, which fundamentally solved the 50-year-old protein folding problem.
• The Democratization of Structural Biology: Today, tools born from the CASP13 revolution allow researchers worldwide to instantly look up or generate highly accurate models of virtually any protein structure.
• Accelerated Therapeutics: From designing custom enzymes that degrade plastic to predicting the structures of viral proteins during health crises, the speed unlocked at CASP13 fundamentally compressed decades of basic research into days.

CASP13 proved that deep learning wasn’t just a tool for sorting photos or playing board games—it was capable of mastering the intricate, chaotic geometry of nature itself.