PACpAInt: the path to personalized treatment for pancreatic adenocarcinoma

Pancreatic Adenocarcinoma is predicted to become the world’s second most deadly cancer by 2030.

For better patient prognosis and treatment, it is vital to diagnose the molecular subtype of a patient’s tumor.

To do so, Owkin has worked with Assistance Publique - Hopitaux de Paris - to create an AI model called PACpAInt.

PACpAInt identifies tumor molecular subtype directly from H&E slides, without the need for the expensive and time-consuming RNA-profiling that is currently used. PACpAInt works like this:

The model has two key steps: first an algorithm detects which regions of the slide contain tumor tissue.

Our algorithm does this with a high degree of accuracy - reaching area-under-the-curve levels of 0.99 and 0.98 when tested on our validation cohorts.

Next, we use a deep-learning model to predict the tumor molecular subtype from the tumor regions identified in the first step.

To reach an accurate level of prediction, we trained our model on 424 whole-slide images for which we had matching RNA profiles.

From the predicted RNA signature, our model could then classify the tumor as basal - higher risk - or classic - lower risk.

PACpAInt also uncovered a new subgroup of patients. A third of the tumors analyzed were a mix of both classic and basal features.

Patients with these mixed tumors were predicted to have a different prognosis by our model.

Finally, PACpAInt was also able to subtype patients based on specific non-cancer cells within the tumor - known as the stroma.

This opens up new possibilities for patient stratification in drug targeting trials.

PACpAInt provides a tool that is easy to implement, that could appreciate prognosis, and that could potentially decide treatment instantly, without lengthy and costly RNA sequencing analysis.

And in future PACpAInt could even be deployed worldwide, finally opening the way for patient stratification based on powerful molecular criteria.