The application of human genetics to drug discovery has long focused on protein coding sequences. However, a fundamental problem in disease genetics is to understand variation in DNA sequences in the 98% of our DNA that does not code for proteins, and the biological consequences of this variation for disease. AlphaGenome1, a deep learning model recently described by DeepMind in Nature, addresses this challenge, alongside the genetic approaches at Nucleome.
Nucleome, as outlined in a recent MedNous article2, is pioneering the use of non-coding human genetics to uncover the molecular drivers of complex disease and to use these insights to develop first-in-class drugs targeting such mechanisms.
Nucleome goes beyond computational prediction (such as AlphaGenome) by combining deep learning with multiple experimental functional genomics technologies. This combination of technologies enables the construction of experimentally verified mechanistic links between cell-type specific changes in non-coding DNA and their downstream consequence. At the core of this technology is Micro-Capture-C, a single base-pair resolution 3D genome assay which precisely links non-coding variants to the genes they affect.
The “Nucleome genetic score”
Integration of the multiple streams of experimental and computational evidence, with explicit consideration of the stringency of each piece of evidence, allows the generation of the “Nucleome genetic score” capturing the confidence in the chain of causality between disease associated genetic variation and the downstream causal genes and the associated cellular contexts. This measure of confidence in causal gene identification can be used for drug target identification and selection, directly informed by the mechanistic link to human disease.
The validity of this integrated technology has been demonstrated (see Figure 2) by a striking correlation between clinical success rates and Nucleome genetic scores, generated for nearly one hundred historical drug targets in inflammatory diseases. Comparison of success rates for target-indication pairs with a high genetic score versus those with a low genetic score demonstrates the ability of Nucleome’s technology to increase confidence in causal gene identification and thus conviction in drug target selection.
Application of the technology
Nucleome is applying its technology to build a pipeline of first-in-class drugs for inflammatory disease. Nucleome’s lead discovery program, NTP464, exemplifies the power of the approach and technology that discovered a novel mechanism driving restoration of immune homeostasis likely applicable to multiple autoimmune indications.
Models such as AlphaGenome are designed to be extended through fine-tuning on custom datasets or via the integration of additional data modalities. A great model, but Nucleome can go further.
Nucleome’s unique dataset of single base-pair resolution 3D genome contacts across disease-relevant primary human cell types, represents a class of training data that could substantially extend the precision of such models: bridging the gap between computational variant effect prediction and the experimentally verified, cell-specific mechanistic links needed to drive drug discovery.
References
