While AI can lift competition and productivity, it also can act as a great leveler, putting smaller players on the same footing as goliaths.
Take pharmaceutical research, for example.
Large companies have the budget and resources to physically test millions of drug candidates, giving them an advantage over startups and researchers. But smaller labs can achieve similar results by harnessing neural networks that simulate how a potential drug molecule will bind with a target protein.
Deep learning can help smaller companies and other researchers discover promising drug treatments by improving the speed and accuracy of molecular docking, the process of computationally predicting how and how well a molecule binds with a protein.
“You don’t actually need to have the molecule in hand,” says David Koes, assistant professor at University of Pittsburgh. “You can screen billions of compounds and they don’t actually have to exist.”
A Perfect Match
When scientists look for the perfect molecular structure for a drug treatment, they look at a few laws of attraction.
A drug molecule should have an attraction, or affinity, to the protein that researchers want it to bind to. Too little affinity, and the drug is too weak for the pair to work.
A familiar principle applies here: opposites attract. The lesson is universally learned in elementary school science classes — and from unsolicited relationship advice. Now Koes and his fellow researchers are imparting this principle to their neural network.
The match should be specific, too — if the molecule is too general, it could bind with a hundred proteins in the body instead of just one. “That’s usually a bad thing,” says David Koes, assistant professor at University of Pittsburgh.
Screening these molecules virtually could speed up the years-long process of identifying a candidate good enough to bring to clinical trials.
As Koes puts it, “When you discover better drugs to begin with, you fail less later.”
This method further opens up researchers’ horizons to test molecules that don’t even exist yet. If a particular molecular structure looks promising, it can be synthesized in the lab.
Koes sees immense potential in this field. He envisions a future world where researchers could use sliders to activate molecular features like solubility, or whether a molecule can pass the blood-brain barrier.
It will take time to get there, he concedes. “It’s quite challenging because you need to make something that’s physically realistic and chemically realistic.”
Unleashing Deep Learning
The researchers’ convolutional neural network looks at the physical structure of a protein to infer what kind of drug molecule could bind as desired.
Choosing a non-parametric method, the team did not instruct the algorithm on which features of molecular structure are important for binding — like “opposites attract.” So far, the results are encouraging and show the neural network is able to infer these laws from training data.
The deep learning model, using the cuDNN deep learning software, improved prediction accuracy to 70 percent compared with the 52 percent of previous machine learning models.
“If we can get it to the accuracy point where people are motivated to synthesize new molecules, that’s a good indicator that we’re useful,” Koes says.
Koes has been using NVIDIA GPUs for almost a decade. He says this work used an array of NVIDIA GPUs, including Tesla V100, TITAN V and TITAN Xp.
Though the team has not yet optimized the model for inference, GPUs have been used in both the training and inference phases of their work.
Koes says the process of virtually screening a test molecule is so complex — the model must sample multiple different 3D positions to determine a molecule’s affinity — that “it’s not really usable without a GPU. It’s like a self-driving car, constantly processing.”