AI, bio-research: a match made in digital heaven

As some of you may know, I’m a huge sci-fi fan. Some of the things I’m going to describe may seem like science fiction, but it’s all either already real, or extrapolated based on existing science.

For decades, scientists sought to understand what’s called protein folding.

A tiny bit of background: nearly all life on earth has a genetic code composed of DNA— except retroviruses such as influenza, which use RNA—that instructs the cell how to build amino acids, which then link together to form proteins.

Because of the electrical attraction and repulsion of various atoms in each amino acid that makes up a protein, it inevitably folds into some sort of shape other than a straight line. Until very recently, it wasn’t possible to predict how a new protein would fold. However, scientists from the University of Washington were recently successful in designing software that allowed them to accurately predict the shapes into which proteins will fold, given a particular amino acid sequence.

You’re probably asking: “Why would this matter?”

Simply put: they can now create proteins that don’t necessarily exist in nature. Imagine a designer protein that could help out those afflicted with gluten sensitivity or even Celiac disease that would chop up the gluten to be digestible. Or a molecular cage that could more efficiently deliver a drug to cells. Or even molecular scale sensors. The possible applications are vast.

Another important, growing area of research is known as machine learning. This is a broad subject but I will focus on one particular part of it. In science fiction, such as Terminator or the Matrix, we often see the dark side of artificial intelligence. What those movies are referring to is often called Artificial General Intelligence. I believe we’re many years from figuring that out, but much of what you seen in the news— and on commercials describing a particular brand of AI— is a very specific, single-goal AI. It’s created with the aim of solving a specific problem, and possesses nothing remotely close to human level intelligence. Instead, its job is to take tasks for which humans would require a vast amount of time, knowledge and training and automates them.

There exists something called a generative adversarial network— GAN for short. There are also many variations on them, trained with different goals. GAN consists of two main parts which are— as the name indicates— adversaries.

The task of the first, the discriminator, is to determine whether or not the data supplied to it represents a real photograph or a fake. It is first trained on anywhere from a few hundred to thousand images of whatever it is supposed to be detecting, such as human faces.

The second part’s job, as generator, is to start with random visual noise— an image consisting of pixels with random color and brightness— and then, over hundreds or even many thousands of repetitions, manipulate that random noise until its output fools the discriminator, which in this example has been trained on human faces, so the generator’s goal is to create something that looks like a realistic photo of a face.

Now, you may wonder, what could this ever be useful for?

We’ve already sequenced the human genome, though we don’t yet know all the intricate details of what every gene and gene combination actually does. Let’s step briefly into the border between reality and science fiction, but still in the realm of what I think is entirely possible.

Let’s imagine a time not very far in the future in which we understand the precise relationship between genes and physical appearance; a time in which we know exactly which genes are responsible for every single aspect of how a person’s face looks.

Imagine a GAN, trained on a combination of DNA sequences, photographs and perhaps other data I haven’t thought of, that is then given an unknown sequence of DNA—say, from a crime scene.

It’s job is to create, as accurately as possible based on the provided sample, an image of the person to whom that sample belongs, most likely the perpetrator of a violent crime.

Perhaps someone is already working such a thing, I don’t know. But I firmly believe that it’s not only possible but on the way to eventually becoming reality. Perhaps it could save lives in ways we can’t even imagine.

Now, imagine an AI of some sort applied to disease research—armed with knowledge of protein folding and also being applied to that other area of research as well to improve its accuracy—that could design cures for diseases that would take humans much longer to devise by experimentation. Those results could then be tested in a real world laboratory, greatly shortening the time it takes to reach a cure.

Some of these ideas may sound a bit out there, and to some, maybe even scary, as people tend to fear what they don’t understand. In my view, the best cure for fear is knowledge.