Every year, Nature Biotechnology highlights corporations which have obtained sizeable early-stage funding within the earlier yr. Atomic AI generates the construction information it wants for its AI mannequin with the goal of RNA drug discovery.
Connecting AI and RNA for drug discovery takes greater than to “one way or the other sprinkle some AI mud on issues and it’ll be nice,” says Atomic AI CEO and founder Raphaël Townshend. Some RNA therapeutics startups and spinouts select to associate with large pharma pretty early on, however that’s a path Atomic AI, based in 2021, has not publicly disclosed to this point. With its $7 million seed spherical and $35 million sequence A financing, ‘AI-native’ Atomic AI is producing the information it wants to coach its synthetic intelligence (AI)-based RNA construction mannequin.
Atomic AI
Atomic AI founder and CEO Raphaël Townshend.
To coach AI algorithms, says Townshend, “95% of the work is getting the suitable information within the first place.” Atomic AI’s groups do their very own wet-lab experiments to complement the sparse public information on RNA construction. In contrast to the Google DeepMind crew that developed AlphaFold, which predicts protein buildings and used information within the Protein Knowledge Financial institution for its coaching information, Townshend and colleagues began from solely 18 identified buildings once they co-developed a software with the lab of Stanford biochemist Rhiju Das to foretell the complicated three-dimensional construction of RNA from an RNA sequence. Their neural community extrapolated molecular options from the coaching set and predicted RNA buildings on the idea of sequence higher than different present platforms, with their Atomic Rotationally Equivariant Scorer printed in 2021 in Science1. The software program is available on Zenodo.
Atomic AI groups have been gathering information by means of wet-lab experiments to feed their Platform for AI-driven Construction Exploration (PARSE). PARSE consists of ATOM-1, its mannequin that predicts RNA construction and performance from chemical mapping information. PARSE and ATOM-1, says CSO Manjunath Ramarao, “are the inspiration for the corporate and the work that we do.” He spent 12 years at Bristol Myers Squibb and joined Atomic AI in October 2023. ATOM-1, they be aware in a preprint2, lets them mannequin RNA buildings regardless of the dearth of floor fact information.
Successes that buoy new entrants within the RNA therapeutics house embody mRNA-based vaccines for COVID-19 and authorised medicine akin to Spinraza (nusinersen), an antisense oligonucleotide that modifies RNA splicing to deal with spinal muscular atrophy. But, in accordance with a report3 by the Nationwide Academies of Sciences, Engineering, and Drugs, “using RNA-based applied sciences for treating human illnesses is in its infancy.” Atomic AI desires to assist the sector develop up because it eyes functions in oncology and neurological and neurodegenerative illness.
Townshend’s profession may need led to a trajectory with self-driving vehicles given his PhD-focus on pc imaginative and prescient with Ron Dror at Stanford College. Throughout his rotation by means of a structural biology lab, he remembers, “this appears to be like like the suitable place to be.” His sense about structural biology intensified in a analysis stint with DeepMind’s AlphaFold crew. As he stayed on the interface between structural biology, biophysics and AI, he deepened his expertise in biology and computing and constructed the platform that Atomic AI now makes use of to foretell the three-dimensional construction of RNA.
RNA is more durable to drug than DNA, says College of California San Diego researcher Gene Yeo, who advises Atomic AI and who was interviewed collectively with Townshend. RNA kinds complicated three-dimensional buildings, and maybe as many as 140 totally different chemical modifications can tune RNA perform and metabolism. For now, it’s much less clear how one can learn or alter most of those modifications, however Yeo thinks Atomic AI is effectively positioned for progress.
Atomic AI scientists use a library of compounds to change particular RNA nucleotides, learn out the adjustments with sequencing and assess how RNA’s two- and three-dimensional construction was altered. Its core neural community, says Townshend, is a basis mannequin, akin to those who underpin massive language fashions like ChatGPT. On this case, the mannequin is tuned to foretell RNA’s purposeful and structural properties. Quite than depend on public information, which could be of variable high quality, Ramarao says, “We needed to do our personal experiments, generate our personal information that may feed into the mannequin constructing.” Atomic AI goals to map the transcriptome to allow a scaled-up hunt for RNA on-targets and off-targets.
The wet-lab and modeling groups collaborate and iterate to construct selectivity and efficiency traits right into a lead molecule, says Ramarao. They handle specifics about chemistry or construction collectively, says Atomic AI scientific advisor Percy Carter, who’s CSO of Blueprint Medicines and who was interviewed along with Ramarao. With RNA, scientists can goal regulation of translation, mRNA stability, RNA degradation or splicing, which could be mediated by proteins or by proteins interacting with RNA. A small molecule may work together immediately with RNA or intervene with RNA–protein interactions to modulate perform or degrade RNAs. Goal sorts will differ, and so there may be direct binding to RNA, says Carter, or one has to think about exercise of a protein-driven-process on RNA. Proteins that mediate splicing function in a posh and bind to RNA as ribonucleoproteins. Thus, as they assess RNA construction, Atomic AI’s groups additionally heed protein construction and habits. RNA buildings could be experimentally resolved, as an illustration with nuclear magnetic resonance (NMR) spectroscopy.
It’s identified what makes chemical compounds and ligands extra probably than others to work together with RNA, says Yeo. Such traits embody that they’re positively charged, intercalating aminoglycoside-like, versatile molecules. However this information alone is inadequate to seek out probably the most promising ones. In some circumstances, a protein, slightly than an RNA, is a greater goal in one of many cell’s many protein–RNA interactions. In different circumstances — as an illustration, the place protein focusing on has failed — RNA focusing on gives a possible therapeutic profit for beforehand undruggable targets.
Concentrating on RNA additionally means navigating RNA’s loops and bulges that type in two and three dimensions, says Matthew Disney from the College of Florida Scripps Institute for Biomedical Innovation & Know-how, who will not be affiliated with Atomic AI. The Protein Knowledge Financial institution will not be an RNA databank, and way more structural details about RNA is required, particularly ligand-bound buildings, he says. Forcefields for small-molecule docking have been optimized for protein docking. RNA binding could be electrostatically related, however it’s much less well-defined and might embody high-energy interactions. Simulated buildings assist to evaluate forcefield interactions, however experimental validation with NMR, X-ray crystallography or different approaches is required, says Disney, to flesh out the atomic particulars and be taught what is actually “the language of the perfect power discipline.”
Efficiency of an RNA therapeutic is vital, says Disney, however so, too, is selectivity. Disney and colleagues developed a computational strategy to seek out on- and off-targets for small molecules that focus on RNA and think about how one may optimize them. When treating most cancers, an off-target impact may be tolerated, he says, however that scenario shifts in a genetic illness when sufferers take a drug all their lives. One other problem is getting sufficient of the molecules into diseased tissue with out an excessive amount of toxicity.
With RNA therapeutics, as in biotech extra usually, startups should steadiness platform-building with their hunt for bioactive compounds to in the end get into the clinic, says Disney. It’s higher to have a compound in hand to advance towards a objective, he says, “however it could come again to platform,” as balancing platform and bioactive compounds, he says, is vital for constructing worth for buyers.
