You have to come as a shock to hear that you are dead. In the (presumably apocryphal) story of the Nobel Prize's (and perhaps apocryphal) origins, Alfred Nobel is so appalled with premature obituaries calling him the “death merchant,” that he is determined to change his reputation and donate his fortune to mankind. But at least if you're reading your Obit, you know you can do something about it. A bigger and perhaps more likely shock to scientists is hearing that your field is dead.
That's because it's the inevitable outcome of research. Because new techniques, technologies, ideas come in and replace what you've done before. And it's a pattern that appears to be happening at an increasing pace as scientists realize the possibilities of AI.
DFT's death?
For example, last year's Nobel Prize was partially awarded to Alphafold, an AI tool for unraveling protein structures developed by Deepmind, owned by Google. When Alphafold first appeared a few years ago, many people working on solving protein structures felt that writing was on the wall for their work. A similar death note arrived last month at the EUCHEMS Inorganic Chemistry Conference.
Meta-built models provide DFT-like accuracy at a fraction of the cost
Reiher was referring to the work that was released earlier this year by Facebook and Instagram owner Meta. Meta has released the Open Molecule Dataset or Omol25, the largest dataset for quantum chemical calculations covering the widest range of molecules and materials generated from millions of DFT simulations.
The dataset itself is an impressive feat, but the reason for building it is to train machine learning models that can perform DFT jobs faster. Similar to how AlphaFold learns from protein structure databases and devises algorithms that can predict folding from only amino acid sequences, models trained in a quantum chemistry calculation databases can predict molecular properties based on atomic coordinates. The models released simultaneously by META appear to be in good condition to match their promises. This provides DFT-like accuracy at just a small portion of the computational cost.
Datasets can democratize chemistry
These changes are dramatic as they not only give researchers powerful tools, but also reduce barriers to accessing the tools. Its democratic effects are also complemented by other applications of AI. For example, many groups around the world have built things like OMOL25 into “agent AI.” This is a system that allows you to autonomously execute commands on your behalf. These often provide easy-to-use chatbot-style interfaces (represented by large language models such as ChatGPT) that allow chemists to use the tools without becoming domain experts. Along with cheaper and faster calculations, these represent major steps in the capabilities of computational tools and availability to researchers. In the near future, we can look for our coverage of this.
It's only the biggest companies with the deepest pockets that can afford to step on the cost
There are still many tasks that cannot be done with AI. For example, it had limited success in drug discovery, because it was unable to address the most difficult part of the task yet, and therefore could not create molecular works in humans. However, many people predict that solving this and other problems is also a problem of gathering sufficient data to train the model. As OMOL25's efforts show, building these datasets is a considerable task, even more so when it involves carrying out physical experiments. Generating them becomes a huge investment of time and money, and is just the biggest company with the deepest pockets that can add up those costs and wait for revenue.
At this point, Meta, Google and others have databases and tools at their disposal (although the latest release of Alphafold no longer allows commercial use). If this continues, the prospects of democratizing discoveries and research may be immeasurable.
It may still be revealed that reports of DFT's death are exaggerated. But we certainly see the birth of a new kind of chemist
