Day 2
AlphaFold – how machine learning changed structural biology forever (or not?)
Saal Zuse
Jan Gebauer
Getting first-hand insights into the impact of machine learning on life science
In 2020, the scientific community was shaken when the results of a special contest for protein prediction, known as the Critical Assessment of Protein Structure Prediction (CASP), were revealed. A relatively new competitor emerged as the champion, surpassing all other teams that had been participating in the game for decades. This new competitor was Google and their predictor was a neuronal network called "AlphaFold". Their new approach caused significant waves in the field of structural biology, even capturing the attention of the mainstream media. Several news channels featured reports on AlphaFold, with one German magazine, "Der Spiegel," declaring that "The year 2020 will be known [...] as the year when machines began to outstrip us in research." Join me as we explore the background behind this transformative development and assess the magnitude of machine learning's impact on science, with a particular focus on structural biology.

In 2021, Google published the methodology and source code for AlphaFold and within days, scientists adapted the code to allow virtually everyone to predict their own protein structures without prior knowledge.

Now, two years after its public release, AlphaFold has established itself as an essential tool in structural biology. Yet, with time, we've also gained a deeper insight into its limitations.

In this talk, I would like to delve into AlphaFold and similar machine learning techniques and explore their impact on science and structural biology. To truly appreciate their significance, we will first need to understand the role of protein structures and how they shape our daily lives. Additionally, we’ll have to examine how protein structures were traditionally solved prior to the advent of AlphaFold. We’ll then touch upon the concepts of protein evolution to better understand the biological basis behind this breakthrough, before we’ll look at the intricacies of the neural network itself and discuss the training data necessary to achieve its remarkable capabilities. Drawing from my experience as a practicing structural biologist, I will illustrate these points with real-life examples, showcasing instances where AlphaFold has succeeded and where it has encountered challenges. Lastly, we will peer into the future and speculate on the potential trajectory of this scientific journey and its potential to transform science and our approaches towards it.