Paper short abstract:

This talk focuses on the adoption of artificial neural networks in the search for protein structures in the late 1980s. By presenting on journal articles and their authors’ methodical reflections, it will be argued that ML was especially akin to the research culture of chemistry.

Paper long abstract:

Computer simulations in chemistry started as early as the 1950s but started to be more regularly used from the 1970s onwards. To name just a few developments, in 1970, the computer program GAUSSIAN developed by the group of John A. Pople at Carnegie Mellon University was released as the first commercially available program for computational chemistry. From the late 1980s onwards, artificial neural networks were used to help predict at least secondary protein structures. Lastly, in the 1990s, Density Functional Theory (DFT) changed the way quantum mechanical computations were approximated in computational chemistry.

In my talk, I will focus on the second of these developments, i.e., the use of artificial neural networks and machine learning for the determination of protein structures. I will argue that machine learning did fit the research culture of chemistry especially well -- perhaps better than the mathematical formalism of quantum mechanics. It was the learning by pattern recognition and the avoidance of direct appeals to theory that particularly made the use of machine learning very akin to chemical thinking. I will argue this point historically by presenting on journal articles concerned with the use of artificial neural networks in the search for protein structures. A special focus will lie on the reflections of the authors on the usefulness of neural networks and their comparisons with other computational tools, as well as their relationship to theory.