Identifying the Mechanism and New Indications for ONC201


OneThree Biotech was born out of the largest precision medicine institute in America with the purpose of decreasing the failure rate of bringing therapeutics to market using biology driven artificial intelligence. Our technology was built in a lab setting; allowing us to engineer our methodologies around real-world results. While working with major biotech and pharma partners; we’ve validated our predictions through work in target discovery, compound and combination therapy identification, toxicity prediction, biomarker selection, and much more. In this white paper, we’ll discuss our OTB Target Deconvolution and Gene Essentiality modules and how they were used for ONC201 (initially developed by Oncoceutics and recently acquired by Chimerix for up to $400M):

  1. Identify a novel target for first-in-class compound ONC201 and
  2. Pinpoint optimal indications/ biomarkers for future clinical development.

OneThree helped solve problems that we had been working on for years and potentially saved several years of additional work. I expect that this approach will be integral to the future of drug discovery

Joshua Allen, Ph.D, Chief Scientific Officer, Oncoceutics


One of the most challenging parts of drug development is understanding the mechanism through which compounds act. This is especially true for compounds discovered through phenotypic screening based approaches– like ONC201. Without a clear target and MoA, Oncoceutics was not able to position ONC201 for the ideal cancer subtypes and patient populations for future development.

Target Deconvolution

Beginning with the data ingestion engine, the OneThree Biotech (OTB) approach starts by integrating over 40 distinct data types from 160+ different public and proprietary data sources. The backend OneThree database is fed into 7 different modules that make up the OneThree’s AI platform – each designed to address a different stage in the drug development pipeline.

Built using a Bayesian network1, the OTB Target Deconvolution module integrates multiple different data sources (e.g. structure, efficacy, transcriptional response) on a single compound to identify similar drugs and most likely binding targets. When tested on 2,000 compounds with known targets it achieved an AUC of .91 at identifying drug-target pairs (Figure 1). When applied to a library of 50,000 orphan compounds the OTB Target Deconvolution module was able to identify a subset of novel microtubule binding molecules that were efficacious against cell lines resistant to know treatments. 

Figure 1– AUC measures as more data is integrated into the Target Deconvolution method. The final value is with all data integrated

Gene Essentiality

Integrating data from high throughput shRNA and CRISPR screening results, we created a multi-class learning method that  combines data both at the gene level as well as the sample level to predict whether loss of a given gene would lead to a decrease in fitness for a specific cancer sample. When validated using external data, we found the Gene Essentiality module could accurately pinpoint high confidence essential genes (AUC = .98, AUPRC = .75). Most importantly, the Gene Essentiality module can be used to differentiate efficacious vs. non-efficacious oncology drugs, by predicting which targets will be most effective for specific cancer types (Figure 2). In model testing we found that the Gene Essentiality module was able to identify 90% approved oncology inhibitors, significantly outperforming using shRNA or CRISPR experimental results (71% and 77%, respectively).

Figure 2– AUC measures as more data is integrated into the Target Deconvolution method. Final value is with all data integrated

Identifying the Target of ONC201

We began by integrating 3 datasets on ONC201 (obtained from Oncoceutics) into our backend database: 1) chemical structure of ONC201 (in SMILE format), 2) growth inhibition data against the NCI60 cell line panel, and 3) pre and post treatment gene expression data. Using these datasets (and the backend OneThree database) our Target Deconvolution module predicted that the top ranked binding target of ONC201 was selective antagonism for DRD2 (Dopamine Receptor 2). This finding was confirmed through both in vitro assays and clinical readouts.

Figure 3– Validation of DRD2 antagonism by ONC201 through A) pure dopamine binding assays and B) prolactin levels in treated patients

Finding new Indications for ONC201

Following validation of DRD2 as the first binding target for ONC201, we then used our Gene Essentiality module to identify the specific cancer subtypes where DRD2 inhibition is most likely to lead to cell death. Across all 300+ cancer models in the backend platform we identified glioblastoma and pheochromocytoma as two subtypes with the highest predicted sensitivity to DRD2 inhibition. Additionally using the Biomarker Identification Module we specifically identified methylation of lysine 27 on histone 3 (H3K27M) as a predictive biomarker for ONC201 efficacy in glioblastoma. 

Validating H3K27M Glioblastoma

Working with Oncoceutics, we validated H3K27M glioblastoma as a new indication through an in vivo animal model. Following this, Oncoceutics started new Ph. II clinical trials in H327M mutant gliomas. Final data from these trials is expected in 2021 but results to date have been incredibly promising and have validated H3K27M GBM as a top indication for ONC201 (Figure 4). 

Figure 4– Press clippings and patient stories from ONC201 H3K27M GBM trials2

  1. Madhukar, et al. “A Bayesian machine learning approach for drug target identification using diverse data types”, Nature Communications, 2019
  2. Gilvary et al. “A machine learning approach predicts essential genes and pharmacological targets in cancer”, BioRXiv, 2019
  3. Obtained from Oncoceutics public releases