Summary of older AI and Collective Intelligence discussions

        The recent progress in Artificial Intelligence provides both challenges and opportunities for scientific collective intelligence. Among the most notable examples of AI progress are ChatGPT and other Large Language Models, which have shown unexpected capabilities (Wei 2022, Mitchell 2023), and AlphaFold, which can predict the 3D shape of proteins from their genetic sequence with unprecedented accuracy (Jumper 2021). The development of AlphaFold has been recognized by the award of the 2014 Nobel Prize in Chemistry.

Challenges

          AI poses specific challenges for science. There are many reports of errors in statements from ChatGPT and from other AI systems. The types of errors and blind spots seem different from those more common in humans.

            These AI systems consist of neural networks with billions to trillions of parameters (Mitchell,…

Guy Salvesen and Giovanni Paternostro have spoken with Adam Godzik. Adam is the Bruce D. and Nancy B. Varner Presidential Endowed Chair in Cancer Research at the UC Riverside School of Medicine, Division of Biomedical Sciences. Adam was closely involved from an early stage in CASP, as a participant, and in the Joint Center for Structural Genomics (JCSG), one of the centers supported by the Protein Structure Initiative (PSI).

He sent the following comments:

                  I think that the success of PDB was driven by it being built by the crystallographic community itself, it was an effort from within, not from outside. It became widely accepted relatively early in its history, definitely before I got into the field. 

                  There was another development in bioinformatics that enabled AlphaFold – residue-residue interaction predictions from MSA (work of Debora S. Marks, for instance https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0028766) or a more general contact map prediction field…

Genentech and DeepMind are examples of start-ups that achieved major scientific advances, while parallel efforts by academic groups and by large companies on the same problems were not successful.

A comparative analysis might better show the reasons behind their achievements.

The research on DeepMind is described in the timeline, while the research on Genentech is based on published historical reconstructions, as the books by Stephen Hall (1987) and by Sally Hughes (2011), on interviews with many protagonists available from the Berkeley Library Digital Collections and on conversations with former employees, including Roberto Crea, who was one of the first five employees and even before that a key author in the papers describing the initial work done at City of Hope with Genentech support (Crea et al, 1978; Itakura et al, 1977; Hirose et al 1978; Goeddel et al, 1979).

The comparative analysis is ongoing, but the following points are…

Mohammed AlQuraishi is an Assistant Professor in the Department of Systems Biology at Columbia University. He is one of the leaders of the OpenFold consortium (https://openfold.io).

Thanks for reaching out about this.

With regards to additions, one piece from my own work is the RGN paper (https://www.sciencedirect.com/science/article/pii/S2405471219300766),

which was the first paper to do end-to-end differentiable learning of protein structure, and the first to show that a protein can be folded implicitly using a neural network. This ended up being the approach that AlphaFold2 ultimately took (with many more additions and elaborations on top of course). 

Jake Feala is a cofounder at Lila Sciences, a company unveiled in March 2025 aiming to use AI and autonomous labs to accelerate scientific discovery.

We invited him to contribute to our historical timeline, from PDB to AlphaFold. Specifically, we asked him about the reasons why the protein folding problem was solved by a VC-backed company, DeepMind, and not by an academic group, and also if he thinks that this achievement provides a general solution for the future of AI in science.

Thanks for the opportunity to contribute!

I think it's not exactly the right question to ask why the protein structure prediction problem was solved by a company and not an academic group. A more fitting question is why DeepMind solved it and not some other entity, academic or not. Back then DeepMind was a totally unique company and not only beat out academia but also the entire…

The timeline traces the historical milestones that led to AlphaFold, a landmark achievement in protein structure prediction powered by artificial intelligence (AI). It highlights key scientific, methodological, and cultural developments spanning over six decades, beginning with the first protein structures solved by Kendrew and Perutz (1958-1960).

Significant early milestones were the establishment of protein sequence and structure repositories, particularly the Protein Data Bank (PDB) initiated in the early 1970s by an effort including both senior and junior scientists. The PDB grew from these grassroots efforts amid debates about data-sharing practices, progressing gradually over several decades. The adoption of open data sharing policies was a consequence of community letters and petitions, initiatives prompted by the PDB leaders and decisions of scientific societies, journals (like Nature and Science) and funders (like HHMI and NIH).

Bioinformatics methods and computational tools evolved considerably, from algorithms for sequence alignment (1970s-80s), through other bioinformatics tools in…

Harold Varmus is a Professor of Medicine at Weill Cornell Medical College. His work has been recognized by the award of the 1989 Nobel Prize in Physiology or Medicine. He is a former Director of the NIH (1993-1999) and of NCI (2010-2015).

He sent the following comment about the change in policy at NIH in 1999 regarding the immediate release of structural data upon publication. We also asked him how this decision was influenced by his often-stated support for open science.

Like Tom Cech’s recollections, my memory of the dates and conversations pertinent to the history of making protein structural coordinates publicly accessible is a bit hazy.   But the decision to promote rapid release of such information was based heavily on the very successful adoption of a very similar policy for DNA sequence information (the Bermuda Rules), which was then being generated by the Human Genome Project.   

In addition,…

Thomas (Tom) Cech is a Distinguished Professor at the University of Colorado Boulder. He is a former President of HHMI (2000-2009). Dr. Cech's work has been recognized by the Heineken Prize of the Royal Netherlands Academy of Sciences (1988), the Albert Lasker Basic Medical Research Award (1988), the Nobel Prize in Chemistry (1989), and the National Medal of Science (1995). In 1987 he was elected to the U.S. National Academy of Sciences.

He sent the following comment about the change in policy at HHMI regarding the immediate release of structural data upon publication:

                  I appreciate your project, but my memory is hazy regarding the timeline of my own contributions (if any!) to the important requirement of depositing x-ray crystallography data on the PDB.

                  We began doing RNA crystallography in 1991, when Jennifer Doudna joined my lab as a postdoc, and began publishing structures in 1996.  I became…

Alexander Wlodawer is a Senior Investigator at the Laboratory of Cell Biology, NCI, National Institutes of Health.

He sent the following comment:

In the history you stress the role of the journals and their rules regarding deposition of both coordinates and structure factors, but I did not see one other crucial development. That was the requirement by first HHMI, and then NIH to deposit such data as a condition of being funded. I don’t remember the exact dates, but I do remember talking to Tom Cech, then the director of HHMI, convincing him that such requirements should be put in place. He also promised to talk to Harold Varmus, then director of the NIH. Very soon after that conversation the rules were officially announced.