In an AI hype-filled biopharma industry, one company is taking a back-to-basics yet supercomputer-powered approach — using Bayesian analysis on massive patient datasets to guide drug discovery. The company crunches trillions of data points per patient. “It’s massive, which is why we use a supercomputer,” said Niven R. Narain, Ph.D., BPGbio CEO. The company has an exclusive relationship with Oak Ridge National Labs, using its Frontier supercomputer to perform complex computational tasks, including the analysis of multi-omics data, the development of predictive models, and the simulation of biological systems. Frontier is hailed as the world’s first exascale supercomputer, meaning it can perform more than 1 quintillion calculations per second.
BPGbio’s AI-powered platform, known as NAi Interrogative Biology, illustrates its approach to drug and diagnostic discovery. The platform includes a lmassive biobank of multi-omics data on more than 100,000 patients. By analyzing this expansive patient dataset, the NAi platform takes a biology-first big data approach to enable an objective exploration of human biology to identify potential new drug and diagnostic targets. Since its development, NAi has discovered more than 100 targets. Partners who have used NAi include the Department of Defense, Sanofi, and Harvard Medical School.
Bayesian approach a rare one in an ML-centered biopharma landscape
Eschewing the more conventional machine learning (ML) techniques favored by many of its peers, BPGbio is one of the few players in the biopharma industry using a Bayesian approach to AI. Bayesian networks, which derive from the work of the 18th century British mathematician Thomas Bayes, visually represent the relationship between variables. They are like a flowchart where each variable is a node, and the connections between nodes show how one variable can influence another.
Machine learning methods can offer powerful capabilities, but often rely on substantial datasets with known outcomes to train algorithms. This approach can introduce biases if the training data reflects faulty assumptions from the past. Dealing with this bias often involves rigorous data curation and model testing. In contrast, BPGbio uses a Bayesian AI approach to support non-obvious hypothesis generation by allowing the data itself to lead the analysis. “Philosophically and clinically, Bayesian methods help us avoid bias by enabling the data to produce hypotheses, rather than letting hypotheses shape the data,” Narain said.
BPM31510 shows promise in phase 2a pancreatic cancer study
BPGbio’s lead candidate, BPM31510 for pancreatic cancer and glioblastoma, exemplifies the promise of its Bayesian AI discovery platform. Originally uncovered by analyzing multi-omics data from the company’s biobank, BPM31510 modulates tumor metabolism by targeting mitochondrial respiration. This novel mechanism of action may sensitize tumors to chemotherapy and radiation. The compound is now in phase 2 trials, showing early signs of clinical benefit including a doubling of progression-free survival in pancreatic cancer.
The company’s pipeline also includes clinical-stage programs for squamous cell carcinoma, epidermolysis bullosa and mitochondrial disorders. “For example, in epidermolysis bullosa, we are analyzing a rare childhood disease associated with acute mitochondrial deficiency,” Narain said.
Promising data propel BPM31510 towards phase 2b pancreatic cancer trial
BPGbio plans to unveil updated data from its phase 2a trial of BPM31510 for advanced pancreatic cancer at the 2024 ASCO GI Symposium in San Francisco. The 19-patient study combined BPM31510 with the standard chemotherapy gemcitabine in patients with refractory disease. Early findings show promising signs of efficacy — the combo doubled progression-free survival to a median 7.2 months, compared to 3.6 months with gemcitabine alone. BPM31510 also demonstrated a favorable safety profile with manageable side effects. Researchers say the latest data supports the proposed mechanism of action of BPM31510 in modulating cancer cell metabolism. On the heels of these positive phase 2a results, BPGbio now plans to initiate a larger phase 2b trial evaluating BPM31510 and gemcitabine as a first-line treatment for pancreatic cancer.
Speaking of its oncology approach, Narain said when building cancer models, the company takes data from hundreds and thousands of patients along with millions of analytes on each patient. “Then, we’re allowing all of that data to be subjected to a Bayesian analysis,” he said.
Filed Under: Biotech, Data science, machine learning and AI, Oncology