Founded by Flagship Pioneering in 2019, Cellarity seeks to embrace biology’s complexity, as diseases are often not driven by a single target, but emerge from a cascade of cellular interactions that shift over time as cells develop, differentiate, and malfunction, says Parul Doshi, Cellarity’s chief data officer.
The Somervile-based clinical-stage biotech integrates high-dimensional transcriptomics with AI models to develop cell-state correcting therapies. Recently, Cellarity has targeted its technology to the prediction of drug-induced liver injury (DILI), a poorly understood late-stage challenge in drug development, costing an estimated $350 million annually per pharmaceutical company.
“We should share more information on safety so that patients get safe drugs,” emphasized Doshi in an interview with GEN. “I hope we compete on efficacy, not safety.”
In a study published in Nature Communications titled, “A large-scale human toxicogenomics resource for drug-induced liver injury prediction,” Cellarity researchers present an integrated AI model, called ToxPredictor, which evaluates toxicogenomics to predict dose-related DILI risks.
At the core of ToxPredictor is DILImap, a transcriptomics library in primary human hepatocytes which illustrates the transcriptional signature of 300 compounds linked to DILI at multiple doses. DILImap characterizes liver injury mechanisms, such as mitochondrial dysfunction, oxidative stress, immune activation, and metabolic changes. Cellarity describes DILImap, as the “largest known toxicogenomics dataset available for DILI modeling.”
Validation results demonstrated 88% sensitivity, how accurately a compound is predicted to lead to DILI, at 100% specificity, how accurately a compound is predicted to be safe. ToxPredictor also identified numerous Phase III clinical safety failures that had been undetected in animal studies. The model and validation data is open-source for non-commercial use.
”Accurately pinpointing at which specific dose DILI occurs will better direct go/no-go decisions for drug discovery,” highlighted Sreenath Srikrishnan, senior scientist at Cellarity and co-author of the paper.
Cellarity’s active learning framework leveraging transcriptomics recently published in Science. The first drug candidate emerging from the platform, CLY-124, is a globin-switching oral medicine, which is currently under evaluation in a Phase I clinical trial.
Save billions
Traditional pre-clinical methods for characterizing DILI, such as quantitative structure-activity relationship (QSAR) models, offer low specificity and binary predictions, which lacks mechanistic insights. In contrast, transcriptomics provides a multi-dimensional system-level view capable of detecting diverse DILI mechanisms not captured by conventional assays.
Among the list of Phase III clinical failures flagged by ToxPredictor include Evobrutinib, a selective inhibitor of Bruton’s tyrosine kinase (BTK) to treat multiple sclerosis (MS), TAK-875, a G-protein-coupled receptor 40 (GPR40) agonist for type 2 diabetes, and BMS-986142, a reversible Bruton’s tyrosine kinase inhibitor for rheumatoid arthritis. Predicting these failures early in the development pipeline could save the industry billions of dollars.
To address the call from regulators to reduce reliance on animal models for drug testing, DILImap leverages primary human hepatocytes to improve the translation gap. While Srikrishnan says the next step is to expand to 3D liver systems, such as organoids, these complex models have limitations, including scale, with current throughput restricted to screening 20-30 compounds, and low-dimensional readouts, such as a limited panel of metabolic markers, which fail to capture the full spectrum of molecular responses.
“You need transcriptomic and mechanistic breadth to tease out subtle signals of how these DILI trajectories are starting in cells, which helps scale across the translation gap,” Srikrishnan told GEN. “We envision growing the platform to multi-scale by integrating different models and omics modalities to provide a holistic picture of DILI. Transcriptomics provide a good starting point.”
ToxPredictor enters a growing ecosystem of transcriptomics models that predict cell state, which the field has placed under the term, “virtual cell.” Srikrishnan says a gap still remains when applying these models for discovery.
“Even for traditional pre-clinical discovery tasks, a lot of these models do struggle. Safety with DILI is one piece that will help us build this holistic view of the entire spectrum of drug discovery,” he said.
