A Really Amazing Use of Artificial Intelligence
I get a lot of science-related stuff in my email inbox. Only occasionally is it so cool that I’m willing to wade through a technical paper, and even less often does it seem worth sharing as a blog post. Science Magazine has such a paper this week.
I’m going to assume that anyone bothering to read this far has the middle-school biology knowledge that all cells are surrounded by a membrane made up of phospholipids and that this membrane is a sandwich with the greasy stuff between two layers of charged or otherwise water-loving stuff. As one side of a cell membrane faces the cytoplasm, and the other side faces outside of the cell, both watery places, it makes sense that membranes stick their water-loving parts into water.
Many important viruses (e.g., HIV, the virus that causes Covid-19, etc.) also have a phospholipid membrane around them. Relative to the size of cells, viruses are tiny.
Key point: because they are a tiny circle, virus membranes are more curved than a big cell’s membrane would be.
Some viruses are assembled in the cytoplasm of a cell, bind to the cell membrane, and induce the curvature that wraps the membrane around it. However, before the virus completes this process, the level of membrane curvature would be intermediate. It would also be intermediate for membrane-bound structures within or outside of cells that are smaller than cells.
Major question: How the hell does a virus or any other protein-based structure recognize/cause membrane curvature?
Related question: Could antivirals be designed based on ability to recognize a degree of curvature within membranes?
Imagine having one hundred of your child’s identical dolls all lined up in a row. Imagine you attach Velcro to each of their ears such that you connect all of them in a line. (This scenario also imagines you are a sufficiently insecure parent that you buy your child one hundred of something that they only need one of and that your child will not be traumatized when you steal and vandalize their dolls, but I digress.) Now, make a circle of the dolls. You can probably do it. Now, make a circle from only five of the dolls. You can’t because the force required to make a circle rips the Velcro apart and exposes the dolls’ bodies to what’s outside of the circle.
The doll circle is your membrane, and the doll bodies are the greasy part of the membrane.
The laws of chemistry do not favor water interaction with this greasy stuff. Greasy sticks to greasy; wettable sticks to wettable. Okay, with the dolls, I guess if you made your circles on the kitchen floor that was covered in fresh blood, the hundred-doll circle would keep the blood out of the center, protecting the doll bodies, and the five-doll circle would not.
Can you tell I’ve written some horror recently? I again digress.
Anyway, let’s say you are looking for something made of protein that will bind well to a curved membrane and perhaps force it to be even more curved. Well, biology gave us a starting point: most such things have a binding part that is a helix with two sides. Think of a helix as coiled like a spring. On one side of the coil, amino acids that like water are clustered. They keep the protein soluble in water. The other side is the business end. It tends to have a lot of amino acids that like to interact with the greasy stuff.
The scientists first used a computer program based on a concept from chemistry called free energy. Systems tend to favor the interaction of two things if that interaction lowers the system’s free energy. You can think of a release of heat out into the surroundings as one way to show a reduction in free energy of the system has happened. In their computational modeling, they modeled a small piece of protein (called a peptide) interacting with a model curved membrane (a surface with holes representing the exposed greasy part).
They used a type of code called a genetic algorithm. In their algorithm, they started with an initial chain of 24 amino acids, and then they had the computer randomly change it many, many times in parallel simulations. If the free energy for a particular simulated interaction went down (the system favored binding), the peptide sequence was kept. The resulting collection was subjected to an additional round of mutation and selection, again and again, until they ended up with a collection of peptides predicted to be really good at binding to curved membranes.
From looking at what they got, they found that if big, greasy amino acids (mostly tryptophan and phenylalanine) are on the side of the helix facing the membrane, the peptides not only bind the membrane well but can shove themselves in there and increase the curvature.
Here’s where AI comes in: they used this collection of predicted peptides as a training set to train a neural network, a form of artificial intelligence. This blog post is already too long, and too many people have done too good a job of explaining neural networks in layman’s terms for me to do that here.
The upshot is that when they used this model to screen through all possible peptide sequences to find those that would bind to curved membranes/induce curvature, the set they came up with included all known, proven biological examples that do this. And lots of new stuff.
As far as antivirals go, there was only one known example that bound selectively to tightly curved membranes. It wasn’t the tightest binder. Indeed, it didn’t bind tight enough to force curvature. Forcing curvature isn’t its mechanism of action; it binds the highly curved viral membrane and punches a hole in it, ultimately leading to virus destruction. (It works against hepatitis C; the authors suggest that if binding was too tight, hole-punching would be inhibited.)
At the end of the paper is the not-so-subtle pitch to get the author’s next grant funded (or corporate interest in funding them): they want to use this AI model to predict the best antiviral medicine candidates by selecting the peptides with this intermediate level of binding to curved membranes and make them. (The making is easy—machines have done this since the sixties.)
Will it work? I dunno. Maybe. I’d fund their next grant were I on the panel.
The point of doing this layperson’s explanation for a science fiction audience is to show how weak AI has become important in science while still being only one component of the process. You wouldn’t get that insight from the sensationalized headlines about AI changing “everything.” Writing fiction about near-future AI (unless for humor/satirical purposes—as anyone who reads my fiction knows, I’ve done that, too) requires a realistic perspective on what weak AI can and can’t do.
If this post helped some folks a bit toward that end, I did my job here.
For the aficionados, here’s the link to the paper: