A cryo-EM and GPCR expert, a virtual screening pioneer, and a virtual chemical library creator walk into a bar.
The result is not the opening line of a joke but rather the founding of a start-up that will rapidly discover novel small-molecule therapeutics through the virtual screening of AI-generated virtual libraries.
That company, Deep Apple, launched with a Series A commitment of $52 million from life sciences venture capital firm Apple Tree Partners.
Deep Apple was founded to develop a discovery engine that combines ensemble cryo-EM, deep learning, and molecular docking screens of ultra-large libraries that should speed up lead optimization—to under a year, they claim—and enable the pursuit of biological target signaling inaccessible to conventional discovery approaches.
“A recent [Journal of Medicinal Chemistry] publication suggested that only 1% of the clinical candidates come from virtual screens,” said Spiros Liras, PhD, founding CEO of Deep Apple and a venture partner at Apple Tree Partners. “We see this not as a limitation but as a major opportunity for a paradigm shift and a call for innovation.”
The drug discovery engine at Deep Apple is based on the knowledge and technologies of its three academic co-founders: Georgios Skiniotis, PhD, Stanford University, an expert in cryo-EM and GPCR structural biology; Brian Shoichet, PhD, University of California, San Francisco; UCSF, one of the first researchers to do virtual screening; and his UCSF colleague John Irwin, PhD, a computer library expert who made the popular ZINC free virtual library of more than 10 billion drugs.
The team at Deep Apple consists of more than 20 people, including a gaggle of computational chemists and machine learning engineers with experience in molecular dynamics and simulations. An exception is the head of drug discovery, who is a medicinal chemist. They are also building a team of biologists for in vitro pharmacology using assays designed in-house.
Straight to the core
Liras said that it is typical for the raw, two-dimensional cryo-EM to get filtered by the human eye, resulting in much of the data getting thrown away. Deep Apple is utilizing deep machine learning to analyze all of the raw, two-dimensional cryo-EM data to accelerate the creation of 3D maps with better quality and to extract dynamics pertaining to conformational states, such as whether they are stable or fleeting.
With the help of all this data, they can identify cryptic or fleeting pockets that seem to stabilize a desired protein for screening by their in-house virtual library collection, Orchard.ai, which primarily consists of novel compounds derived from models they have developed against GPCR subfamilies. To select and prioritize compounds, Deep Apple has created a proprietary large-scale docking-based scoring algorithm.
Liras said that Orchard.ai was created for several reasons. One reason is that pre-existing virtual libraries, like the ZINC virtual library—a free database of commercially available compounds for virtual screening—serve as models for quick and reasonably priced synthesis and testing in the search for biologically active molecule analogs. Deep Apple only makes a compound when its computational approach convinces them that it has some intrinsic biological value, as opposed to producing all of the compounds for physical screening.
Though broadly applicable, Deep Apple’s platform is best suited for targeting any integral membrane protein, including receptors, transporters, and ion channels. Deep Apple is focusing its platform on GPCRs, in part due to the expertise of its founders and because there is a lot of opportunity. The company is advancing multiple programs focused on GPCR modulators, a proven target class with applications in metabolic disorders, inflammation, immunology, and endocrine diseases.
The team is still in the early stages of its drug discovery process, with a portfolio of seven targets. Liras said they are looking to deliver their first clinical candidate for a target in inflammation in the second quarter of 2024. Deep Apple is looking to add more clinical candidates in early 2025 and has several targets in metabolic diseases that pertain to weight loss management.
Another AI varietal?
Liras strongly prefers Deep Apple not to be labeled “another AI company.” How he sees it, Deep Apple applies rational, deep learning models to answer contemporary questions in drug discovery.
“The genesis of Deep Apple is based on observations that biological signaling may give us an opportunity to explore drug discovery differently,” said Liras. “We are aware that deep learning can accelerate a lot of the computation. It’s a very rational way to exploit all the data that we get from structures that seem capable of inducing a particular signaling cascade.”
There’s a phrase: “Anyone can count the seeds in an apple, but only God can count the number of apples in a seed.”
Much in the way that anyone can count the seeds in an apple, drug discovery can be a laborious process of chewing and digesting data down to the core. Perhaps it’s only through tools like AI that the number of drugs can be culled from the data like apples being counted from seeds. That’s exactly what Deep Apple hopes to do, at least within their protein slice of GPCRs. The question is whether it’ll be a nice, clean bite or one with a worm.