Chapter 60
Obesity and Diabetes — 19
But that didn’t stop statin exuberance. Michael Brown and Joseph Goldstein, who were Nobel laureates for their work on the LDL receptor that led to statins, wrote “Heart Attacks: Gone with the Century?” forecasting that statins “may well end coronary disease as a major public health problem early in the next century.” Nearly three decades later, while there has certainly been progress, heart disease is still the number one cause of death, and there are over eight hundred thousand heart attacks a year in the United States.
Yet another parallel between statins and GLP-1s is the increasing potency of the drugs and their effect as the field moved forward. For statins, the LDL-lowering effect between the first available statins and more potent ones, like rosuvastatin (Crestor), that came along more than fifteen years later was striking. In the era of lower potency statins, there was no indication that these drugs could induce type 2 diabetes. In 2012, I wrote an op-ed in the New York Times that highlighted the risk of type 2 diabetes with potent statins, newly acknowledged by the FDA, of about 1 in 167 people taking Crestor 20 mg per day. While that’s a low absolute risk (0.6%), the number of people taking potent statins is high. Many studies since have reinforced that there is a small but unequivocal type 2 diabetes risk for potent statins, likely due to impaired insulin sensitivity. As I’ve emphasized, it’s imperative to avoid type 2 diabetes, if possible—we certainly don’t want that to be treatment induced. The main point is that it took many years after statins were first available, and increasing their doses and potency, to recognize the risk. For the GLP-1 drugs, it will take many years to sort out whether there will be unknown, unanticipated side effects (“unknowns”) for people taking increasingly potent preparations, like triple-receptor activators, for extended periods.
This chapter opened with the following questions: Why did we let twenty years slip before we recognized the true potential of GLP-1 drugs? What if we had AI in the early 2000s and could ask GPT-4 how to make a long half-life peptide that mimicked this naturally occurring incretin, and what would be its potential use?
The most far-reaching accomplishment in AI for life science, recognized by the 2024 Nobel Prize in Chemistry, has been the ability to predict the three-dimensional structure of the protein universe (more than two hundred million proteins) from a single, linear sequence of amino acids. This was called AlphaFold2 by a team of thirty scientists at Google DeepMind, led by Demis Hassabis and John Jumper. This required a jump from deep learning, using a convolutional neural network, to a transformer model, which corresponded to a striking improvement in predicting protein structure accuracy, from 59 to 92 percent, and from the median level of 6.6 angstroms to 1.5 angstroms (0.15 nm). The diameter of an atom is 2 to 3 angstroms, so this is predicting protein structure at the atomic level. The discovery of transformer models was made by a team at Google in 2017, distinctly different from previous deep-learning models. Instead of having to look back and forth from each word in a sentence (called a recurrent neural network), the transformer model could contextualize a whole sentence at once. Over time, with ingestion of hundreds of thousands of books, Wikipedia, and much of the content on the Internet, this gave way to large language models, such as ChatGPT and GPT-4.