The biotech trade is experiencing a rush of AI-powered instruments for a lot of elements of the complicated drug discovery course of. However one which has flown below the radar, more and more considered key to sure ailments however woefully understudied, is RNA. With $35 million in new funding, Atomic AI goals to do for RNA what AlphaFold did for proteins, and discover completely new remedies within the course of.
For those who can nonetheless recall your highschool biology, you in all probability keep in mind RNA as type of a center man between DNA (long run data storage) and proteins (the equipment of mobile life on the molecular stage). However like most issues in nature, it doesn’t appear to be fairly that straightforward, defined Atomic AI’s CEO and founder, Raphael Townshend.
“There’s this central dogma that DNA goes to RNA, which fits to proteins. However it’s emerged lately that it does far more than simply encode data,” he stated in an interview with TechCrunch. “For those who have a look at the human genome, about 2% turns into protein in some unspecified time in the future. However 80 p.c turns into RNA. And it’s doing… who is aware of what? It’s vastly underexplored.”
In comparison with DNA and proteins, little work has been carried out on this space. Academia has centered on different items of the puzzle and prescribed drugs have, partly as a consequence of that, pursued proteins because the mechanisms for medication. The result’s a extreme lack of understanding and knowledge on RNA constructions.
However what Atomic AI posits is that RNA is useful and value pursuing as a way of therapy. The key is within the “non-coding” areas of RNA, that are just like the header and footer on a doc. They do protein-like work however aren’t proteins — and so they’re not the one instance.
You possibly can take into consideration RNA strands as beaded necklaces, far more string than bead. The string is “floppy” and kind of what its detractors assume it’s: an middleman. However each on occasion you get a extremely attention-grabbing knot that appears unlikely to have shaped accidentally. As with proteins, should you can work out their construction, that goes a good distance in direction of understanding what they do and the way they are often affected.
“The hot button is to seek out these beads, these structured bits. It’s excessive data content material, it’s targetable, and it’s seemingly useful as effectively,” stated Townshend. “It’s seen in drug discovery as a key new frontier.”
An attention-grabbing concept for a graduate thesis, maybe (and it was for Townshend), however how are you going to construct a enterprise round it?
First, if the sphere is about to grow to be extra necessary, constructing out the strategies for learning has loads of worth. Then, should you do construct these strategies, you could be first in line to make use of them. Atomic AI is doing each concurrently.
A rotating 3D mannequin of an RNA strand construction predicted by PARSE.
The core of Atomic’s IP is, although that is one thing of a simplification, an AlphaFold for RNA. The biology is totally different, and the way in which the fashions work is totally different, however the concept is similar: a machine studying mannequin educated on a restricted set of a kind of molecule that may make correct predictions concerning the construction of different molecules of that kind.
What’s wild is that Townshend’s crew made simply such a mannequin, which outperforms others by a big margin, by feeding it the traits of simply 18 RNA molecule constructions “printed between 1994 and 2006.” This totally bare-bones mannequin wiped the ground with others, as disclosed in a front-page article printed in Science in 2021.
Since then, Townshend was fast so as to add, the corporate has vastly augmented its fashions and strategies with extra uncooked materials, a lot of which it has created itself in its personal moist labs. They name the up to date set of instruments PARSE: Platform for AI-driven RNA Construction Exploration.
“The Science paper represented an preliminary breakthrough, however we’ve truly generated an enormous quantity of… structure-adjoining knowledge,” he defined. “Not the total construction itself, however knowledge associated to the construction, tens of thousands and thousands knowledge factors; the identical scale of knowledge you’d want to coach large language fashions. And mixed with different machine studying work, we’ve been in a position to dramatically enhance each the velocity and accuracy from the paper.”
Which means Atomic AI is the one one who, publicly no less than, has a system that may take a RNA molecule’s uncooked knowledge in and spit out a fairly assured estimate of its construction. That’s helpful to anybody doing RNA analysis in or out of medication, and with gene therapies and mRNA vaccines, the sphere is unquestionably on the rise.
One other RNA construction (however rendered otherwise).
With such a software you possibly can go certainly one of two methods: license it as a “construction as a service” platform, as Townshend put it, or use it your self. Atomic has opted for the latter, and is pursuing its personal drug discovery program.
This strategy has a notable distinction from loads of the AI discovery processes on the market. The final concept is you’ve a protein, say one you need to inhibit expression of within the human physique, however what you don’t have is a chemical that binds reliably and completely to that protein, precisely the place and if you need it to (and cheaply, if attainable).
AI drug discovery efforts have a tendency to supply 1000’s, thousands and thousands, even billions of candidate molecules that may work, rank them, and let the moist labs begin working by way of the record as quick as they’ll. If you could find one which meets these above traits, you’ll be able to produce a novel drug or change a dearer one available on the market. However the important thing factor is you’re competing to seek out new binders to a identified protein.
“We’re not simply discovering binders, we’re discovering what’s targetable within the first place. The rationale that’s attention-grabbing is as a result of on the finish of the day, these large prescribed drugs care extra about novel biology than novel molecules. You’re enabling one thing that wasn’t doable earlier than by discovering this new goal, versus augmenting the variety of molecules out there to focus on it,” stated Townshend.
Not solely that, however some proteins have been discovered to be nigh undruggable for no matter motive, producing diseases immune to remedy. RNA might enable therapy of those similar diseases by making an finish run round the issue protein.
For the current, Atomic AI has narrowed down the record to sure cancers that end in pathological overproduction of proteins (and therefore good choices for preempting the mechanism), and neurodegenerative ailments that will additionally profit from upstream intervention.
In fact all this work is immensely pricey, necessitating because it does a considerable amount of each lab work and intense knowledge science. Happily the corporate has raised a $35 million A spherical, led by Playground World, with participation from 8VC, Manufacturing unit HQ, Greylock, NotBoring, AME Cloud Ventures, in addition to angels Nat Friedman, Doug Mohr, Neal Khosla, and Patrick Hsu. (The corporate beforehand raised a $7 million seed spherical.)
“Folks have picked all of the low-hanging fruit in protein land,” stated Townshend. “Now there’s new biology to go after.”
