Skip to content

Predictive drug development: toward efficacy from the get-go

Experienced drug hunters are learning how to apply machine learning and proprietary exper­i­mental assays to the design of therapeutic molecules that have predictably low toxicity, potentially expanding the pipeline of treatments for both common and rare diseases.
Creyon Bio

The DCVC Deep Tech Oppor­tu­ni­ties Report (which debuted this year) summarizes our thinking about the deep tech investment areas we consider the most exciting, important, and conse­quen­tial. It’s also a guide to the inspiring work innovators inside and outside the firm’s portfolio are doing to extend human capa­bil­i­ties, save the environment, and make everyone’s lives longer, healthier, and easier.

Three of this report’s oppor­tu­ni­ties bore on the augmen­ta­tion of human abilities. What follows is a summary of one of them.

Opportunity 3.1: Predictive Drug Development

With the emergence of oligonu­cleotide-based medicines (OBMs), we entered the era of information drugs” engineered to overcome specific flaws in the body’s genetic program. Once we understand how a mistake in our DNA or RNA causes a disease, we can, in principle, design an OBM to fix it. There’s just one big problem. Inside the body, many potential drug molecules do more harm than good. That’s why traditional pharma companies must sift through thousands of drug candidates, and spend months or years on toxicity studies in animals, to find just a few molecules that don’t have dangerous side effects. What’s needed is a new generation of tech­nolo­gies for engineering compounds that drug makers know in advance will be safe and effective, eliminating the old trial-and-error approach to drug screening and vastly speeding up safety studies and clinical trials. 

DCVC Bio has invested in a collection of startups exploring new compu­ta­tional approaches to this problem, for both OBMs and traditional small-molecule drugs. One example is Creyon Bio, which has built a library of survey compounds repre­sen­ta­tive of the kinds of nucleotide sequences and added chemical units found in OBMs. It studies how those compounds affect the livers, kidneys, cardio­vas­cular systems, and nervous systems of mammals. At the same time, test-tube assays reveal whether the survey compounds are effective in patient cells. The result is a proprietary dataset so large that the company can use it to train machine-learning algorithms to predict whether novel OBMs will have toxic or immune-stimulating side effects. 

We’re trying to solve the oligonu­cleotide design problem, which is that the space of possi­bil­i­ties is so huge, trial and error completely fails,” says Swagatam Mukhopad­hyay, Creyon’s cofounder and chief scientific officer. We said, What is the minimal set of sequences we need to study to be maximally informative about the phar­ma­co­logic design space?’ That gener­al­iza­tion allows us to come up with molecules that are orders of magnitude better, in the sense that we have high confidence in their safety and efficacy from the get-go,” before it’s ever delivered to a patient. 

The appli­ca­tions of machine learning and artificial intel­li­gence in the phar­ma­ceu­tical sector are still very much in their infancy,” says DCVC Bio managing partner John Hamer. But with a rational, predictable way to design low-toxicity medicines, he says, drug makers could create treatments for the millions of Americans with rare, untreated diseases: We call that the long tail of drug development, and no one’s touching it today.”

Read and/or download the DCVC Deep Tech Oppor­tu­ni­ties Report here.

Related Content