More Recent Comments

Thursday, November 14, 2024

The success of protein structure prediction software depended on the solved structures deposited in the Protein Data Bank (PDB)

The development of protein structure prediction programs began fifty years ago and culminated in the remarkable success of AlphaFold, developed by Google DeepMind. Demis Hassabis and John Jumper of Google DeepMind received the Nobel Prize in Chemistry (2024) for their work on AlphaFold.

AlphaFold and its predecessors were trained on a database of known protein structures called the Protein Data Bank (PDB). PDB began in 1971 as a collaboration between the Cambridge Crystallographic Centre in the UK and Brookhaven National Laboratory in the US. It utilized standardizing software for collecting and storing atomic coordinates and allowing researchers to search the database from remote locations. It soon became a requirement for researchers to deposit their data in PDB when they published.

The Wikipedia entry on PDB has a brief history of the key players and institutions.

It's important to know that AlphaFord would not have been possible 30 or 40 years ago because there were not enough known crystallographic structures for it to have learned the general rules and principles of protein folding. The development of successful prediction software absolutely required a tremendous investment in protein structure laboratories and the collection of data from all around the world. Most of the money for this kind of (mostly) basic research came from governments. And most of the money for developing and supporting PDB comes from government sources.

Many university departments, including my own Department of Biochemistry at the University of Toronto, hired protein structure researchers in the 1980s and 1990s and thousands of graduate students and postdocs were trained in these labs. This investment only began to pay off in the 21st century, due partly to the development new technology that made solving structures easier and cheaper.

I'm pleased that Nature recently published an interview with Helen Bergman, a former director of PDB, in order to point out how important the database was in the success of AlphaFold [The huge protein database that spawned AlphaFold and biology’s AI revolution]. We should not lose sight of all the basic science work that contributed to the success of Google DeepMind. Here's an excerpt.

Do you think we would have had AlphaFold without the PDB?

Knowing what I think I know about how AlphaFold works, it would have been extremely difficult. Two things were important about the PDB data: it’s checked and validated by expert curators. The other thing is that the data are completely machine readable.


2 comments :

Graham Jones said...

Another important contribution to Alphafold's success comes from molecular evolution and bioinformatics. Alphafold uses multiple sequence alignments (MSAs) as well as known protein structures. The following gives a flavour of the work that went into Alphafold. I don't claim to understand it.

"The MSA loss is intended to force the network to consider inter-sequence or phylogenetic relationships to complete the
BERT task, which we intend as a way to encourage the model to consider co-evolution-like relationships
without explicitly encoding covariance statistics (this is the intention but we only observe the outcome that
it increases model accuracy)."
https://www.nature.com/articles/s41586-021-03819-2

Joe Felsenstein said...

Some decades ago I was in a hotel in Palo Alto, California, there to attend a meeting of the editorial committee of the Annual Review of Ecology and Systematics. Eating in the nearly-empty hotel restaurant, I joined an older woman scientist who was also there for a similar reason. When computational biology came up, she said that she was the one who had, manually, typed in the coordinates of the first entry in the database of 3D protein structure coordinates. I could not help saying "that's important!". It sounds like this may have been Helen Berman.