Skip to main content
Premium Trial:

Request an Annual Quote

Predictive Oncology Shares Validation Data at AACR on Machine Learning Drug Development Platform


ORLANDO – Predictive Oncology's machine learning platform predicted tumor responses to a library of drugs with 92 percent accuracy, according to data the company shared at the American Association for Cancer Research's annual meeting on Sunday.

Predictive Oncology developed PeDAL, or patient centric discovery by active learning, to tackle the problem of attrition in drug development. "Most drugs never make it to the patient," Robert Montgomery, Predictive Oncology's director of bioinformatics, said during a presentation. "Many fail during clinical trials. It results in a waste in both cost and time."

By now, it's well known that more than 90 percent of investigational drugs fail to reach patients, and these failures have increased the cost of bringing a drug to market to between $1 billion and $2 billion. Part of the problem, according to Montgomery, is that early development programs don't account for patient diversity in drug response. To address this problem, Predictive Oncology spent 15 years building a biobank of more than 150,000 tumor samples. The firm then applied a machine learning algorithm to develop and improve a predictive model based on experimental data from a selection of those biobank samples and a set of drug compounds.

At AACR, the firm shared data from an internal pilot study comprising 130 ovarian tumor samples and 175 US Food and Drug Administration-approved drugs, a small number of which had been approved specifically for treating ovarian cancer. In total, there were 22,850 possible lab experiments for assessing each drug's activity in each tumor.

They carried out the study using a form of machine learning called active learning, in which the PeDAL platform produces predictions, then iteratively incorporates drug response data from laboratory experiments to improve the predictive model. Montgomery said the algorithm's selection of experiments for the researchers to carry out in the lab was guided by which predictions the algorithm deemed most uncertain, meaning that the results would therefore be most informative to the model.

The researchers strategically chose and completed 720 experiments in the lab to train the model, which Montgomery said was then able to make an additional 4,658 "confident drug response predictions," representing 20 percent of all possible experiments. Then, they used an independent set of tumor samples to validate the model. "Our model was able to accurately predict whether an experiment would be a hit or miss with 92 percent accuracy," Montgomery said.

"Then, we're able to ask questions like which compounds had the highest fraction of patient responses," or query whether certain subpopulations had heterogeneous drug responses that may point to the need for later biomarker discovery studies, Montgomery explained. A typical study, such as the one Predictive Oncology conducted to validate the predictive model, takes eight to 12 weeks using the PeDAL platform.

In addition to data from lab experiments, the predictive model incorporates other data on cancers of interest including extracted features from 2D tumor structure, molecular fingerprints from patient samples, biomarker information, historical drug response profiles, and basic tumor pathology characteristics.

To assess the predictive power of the PeDAL model, the investigators created a receiver operator characteristic curve, or ROC curve, and calculated the area under the curve (AUC). An AUC value of 1 indicates the model is able to make perfect predictions and 0.5 is the same as random chance. Montgomery said the model achieved an AUC of 0.98. They then recalculated this metric using samples that had not been previously selected or measured, and "the model was still capable of a very high AUC value," Montgomery said.

"In our study, by only testing 3 percent of our possible experiments, our model was capable of making confident predictions on an additional 20 percent of experiments, almost a sevenfold increase" over the number of lab experiments conducted, he said. "If we had decided to run these confident model predictions of 4,600 experiments in our lab, it would have taken us an additional 18 months to reach the same result that we did in 12 weeks."

The company believes its predictive model can improve the efficiency of drug discovery and development by reducing the time it takes to screen drugs and determine their chances of market success much earlier. Pamela Bush, Predictive Oncology's senior VP of sales and business development, previously said the company believes PeDAL has the potential to help drugmakers find optimal indications for their cancer therapies before they begin clinical trials and rule out agents that are likely to be ineffective.

The company is on the lookout for partnerships and collaborations to further test out the PeDAL platform's capabilities. In December, Predictive Oncology inked a deal with Cvergenx to develop radiotherapeutics and identify targets for radiosensitizers and radioprotective drugs.