A key enabling breakthrough at Nosis Bio has been the development of RNA medicines that can target any cell type via receptor-mediated endocytosis. Achieving targeted delivery requires more than just strong binding to a target receptor—it demands true specificity. To design safe, effective therapies, our medicines must avoid binding to the vast amount of non-target proteins across the human proteome. Our latest innovation, Deep Kernel Inversion (DKI), was designed specifically to meet this challenge by enabling the design of therapies with optimized on-target affinity and minimized off-target interactions.
The Overlooked Challenge of Off-Target Specificity
Machine learning has made significant strides in drug design, but most efforts have focused on maximizing a drug’s affinity for its intended target. While important, on-target affinity is only half of the equation. Equally crucial—and often overlooked—is ensuring low or negligible affinity for the thousands of non-target proteins. At Nosis we have identified approximately 120,000 bindable extracellular epitopes across the human proteome, each of which a drug could bind with unintended consequences, potentially causing side effects and/or reducing therapeutic efficacy.
Designing specificity has remained largely unexplored because existing computational methods lack the scale and precision needed to predict interactions across the entire proteome. Designing a drug with true specificity, one that binds only to its designated target, requires assessing its interactions with all other proteins—a task that has traditionally been computationally prohibitive.
How Deep Kernel Inversion (DKI) Enables True Specificity
At Nosis Bio, we invented DKI to address this fundamental challenge. DKI is a deep learning-based framework that transforms molecular surfaces into high-dimensional vector representations. In this vector space, molecular interactions can be accurately predicted with a simple dot product calculation, which reduces the complexity of predicting an entire molecular interaction network from O(n2) to O(n). In doing so, we have turned the computational prediction of molecular interaction networks from a task that would have required millions of GPU hours (using methods like Alphafold2) to something that requires less than 5 hours of runtime.
With DKI, we can now efficiently screen any molecule’s affinity across all extracellular epitopes. This enables us to design truly cell- and gene-specific medicines with exceptional potency and minimal off-target activity. By achieving this balance, DKI allows us to design RNA medicines with true specificity, ensuring they act precisely where needed without unintended interactions.
A New Standard in Precision Drug Design
By enabling off-target affinity predictions at proteome scale, DKI moves us beyond conventional drug design focused solely on on-target affinity. With this approach, we can create RNA medicines that are both potent and precise, reducing the risk of off-target effects and improving therapeutic outcomes.
Today we are releasing a new white paper, "Deep Kernel Inversion: Rapid and Accurate Molecular Interaction Prediction for Drug Design," which dives into the science and potential applications of DKI in mapping biological pathways, target discovery, and developing the next generation of RNA-based therapies. In doing so, our goal is to draw attention to the need for specificity in drug design, and we hope this will lead to others taking this challenge on as well.
We will be talking more about some of the other technologies we have developed at Nosis Bio in the coming months, so stay tuned!