NEW YORK – BioAI executives hope the company's recently signed deal with biotechnology firm Arbele will be the first of many collaborations centered on developing companion diagnostics out of its artificial intelligence-driven biomarker prediction platform.
Founded in 2019, BioAI aims to become a multimodal AI-driven precision medicine company. Prior to BioAI, CEO and Cofounder Thomas Colarusso had business development roles at the immuno-oncology diagnostics firm HalioDx, the digital pathology company PathAI, and the tissue imaging and data analysis company Definiens, which was acquired by AstraZeneca in 2014. At BioAI, Colarusso and his co-founders have assembled a group of experts to build models that can predict biomarkers associated with various diseases, including cancer, based on histopathology slide images.
More specifically, BioAI's platform uses a deep learning approach to identify genomic, proteomic, or other types of biomarkers from readily available immunohistochemistry (IHC) or hematoxylin and eosin (H&E)-stained slides. The predictive models are trained using image datasets from large clinical laboratories, according to Colarusso, and those images are annotated by pathologists to mark significant areas of the image such as tumor, immune, and normal tissue cells. The training data also include corresponding RNA and DNA sequencing data from patients, when available.
Initially, Manchester, New Hampshire-based BioAI has focused on working with drugmakers to build models that support early-stage pharmaceutical discovery and development. Colarusso said that work includes programs in gastric, lung, breast, cervical, ovarian, endometrial, and skin cancer, as well as inflammatory bowel disease and Crohn's disease.
For example, in August, researchers from BioAI and Eli Lilly published an analysis they collaborated on in AI in Precision Oncology. The goal of the project was to develop an AI model to detect RET alterations from H&E slides as a prescreening tool for enrollment in a non-small cell lung cancer clinical trial. The resulting model performed with 100 percent sensitivity and 63.3 percent specificity in an independent blind assessment, which the researchers deemed suitable accuracy for a method to prescreen patients prior to genomic testing.
Now, BioAI wants to validate models for clinical applications. "We've been focused on [pharmaceutical companies] over the last five years to build these early-stage biomarker assets … and some of the [companies] we're working with now have expressed interest to take it all the way to an [in vitro diagnostic test] with [US Food and Drug Administration] approval," Colarusso said.
In June, BioAI partnered with the molecular diagnostics firm Genomic Testing Cooperative to develop AI-based tests for both pharmaceutical researchers and clinicians. Within the collaboration, BioAI and GTC will develop biomarker screening algorithms and genomic profiling assays that utilize tissue and liquid biopsies.
While Colarusso believes that BioAI's models could eventually replace genomic testing for cancer patients, the company's vision for now is to use them to screen patients for specific biomarkers, which then can be confirmed by FDA-approved next-generation sequencing panels, qPCR tests, or fluorescence in situ hybridization assays. Using the AI screen to predict which patients may have biomarkers associated with targeted treatment could make the testing process more efficient and allow more patients to receive biomarker-directed precision medicine.
As an example, Colarusso said if a patient is diagnosed with lung cancer, the doctor will often put them on immunotherapy right away. If NGS panel testing, which has a turnaround time of about three weeks to four weeks, then shows an EGFR mutation, it's often too late to change course. Patients that are already on immunotherapy are unlikely to be taken off it. However, if the AI screen predicts that a patient has an EGFR mutation, Colarusso said that this could motivate the clinician to wait a few days for the panel test results to confirm the presence of a mutation before committing to a treatment plan.
In its latest collaboration with Arbele, a biotech specializing in developing treatments for gastrointestinal cancers, BioAI will be helping Arbele develop a companion diagnostic for detecting CDH17 expression in colorectal cancer patients that might respond well to the antibody-drug conjugates it is developing. But Colarusso said the test could also be applicable to other Arbele programs in solid tumor malignancies.
Although Arbele could theoretically develop a laboratory test for directly assessing CDH17 expression, Colarusso noted that such a test would require pathologists to manually interpret a stained slide under a microscope and assign an expression score, which can vary from pathologist to pathologist. An AI algorithm could bring more standardization to the process, in his view.
"[Arbele's] goal is to validate this to go from image to making a drug decision for a patient," Colarusso said. "And that's awesome for us [at BioAI] because that's the direction we're heading. We want to bring our applications to the clinic and that's why [Arbele has] decided to move forward with us."
Colarusso said BioAI wants to ink more development partnerships like the one with Arbele. BioAI's companion diagnostic pipeline, he said, will continue to grow out of ongoing collaborations with pharma as their drugs progress through clinical trials and sponsors see greater need for tests that can identify responders to treatments.
In conjunction with that effort, BioAI is building a network of clinical reference laboratories through partnerships similar to its deal with GTC. Colarusso envisions a network of labs that can handle testing tens of thousands to hundreds of thousands of patients per month. "Our goal is to be able to deploy specific tests for [pharmaceutical companies] into these labs that will be used for prospective screening of patients," he said.