University of Toronto scientists use machine learning to fast-track drug formulation development

From left: Zeqing Bao, PhD trainee in pharmaceutical sciences, and Riley Hickman, PhD trainee in chemistry, are co-authors on the study and supported study development and publication through data collection, results analysis, and preparation of the manuscript.

Making machine predictions useful to reduce ‘trial and error’ for new drug development

To investigate whether machine learning tools could accurately predict the rate of drug release, the research team trained and evaluated a series of eleven different models, including multiple linear regression (MLR), random forest (RF), light gradient boosting machine (lightGBM), and neural networks (NN). The data set used to train the selected panel of machine learning models was constructed from previously published studies by the authors and other research groups.

“Once we had the data set, we split it into two subsets: one used for training the models and one for testing. We then asked the models to predict the results of the test set and directly compared with previous experimental data. We found that the tree-based models, and specifically lightGBM, delivered the most accurate predictions,” said Pauric Bannigan, research associate with the Allen research group at the Leslie Dan Faculty of Pharmacy, University of Toronto.

As a next step, the team worked to apply these predictions and illustrate how machine learning models might be used to inform the design of new LAIs, the team used advanced analytical techniques to extract design criteria from the lightGBM model. This allowed the design of a new LAI formulation for a drug currently used to treat ovarian cancer. “Once you have a trained model, you can then work to interpret what the machine has learned and use that to develop design criteria for new systems,” said Bannigan. Once prepared, the drug release rate was tested and further validated the predictions made by the lightGBM model. “Sure enough, the formulation had the slow-release rate that we were looking for. This was significant because in the past it might have taken us several iterations to get to a release profile that looked like this, with machine learning we got there in one,” he said. 

Expectations around the speed with which new drug formulations are developed have heightened drastically since the onset of the COVID-19 pandemic. “We’ve seen in the pandemic that there was a need to design a new formulation in weeks, to catch up with evolving variants. Allowing for new formulations to be developed in a short period of time, relative to what has been done in the past using conventional methods, is crucially important so that patients can benefit from new therapies,” said Allen explaining that the research team is also investigating using machine learning to support the development of novel mRNA and lipid nanoparticle formulations.

More robust databases needed for future advances

The results of the current study are encouraging and signal the potential for machine learning to reduce reliance on trial-and-error testing slowing the pace of development for long-acting injectables. However, the Allen and the research team identify that the lack of available open-source data sets in pharmaceutical sciences represents a significant challenge to future progress. “When we began this project, we were surprised by the lack of data reported across numerous studies using polymeric microparticles,” said Allen. “This meant the studies and the work that went into them couldn’t be leveraged to develop the machine learning models we need to propel advances in this space,” said Allen. “There is a real need to create robust databases in pharmaceutical sciences that are open access and available for all so that we can work together to advance the field,” she said.

"We want this to be the start of something and not the end of the story for machine learning in drug formulation."

To promote the move toward the accessible databases needed to support the integration of machine learning into pharmaceutical sciences more broadly, Allen and the research team have made their datasets and  code and available on the open-source platform Zenodo.

“For this study our goal was to lower the barrier of entry to applying machine learning in pharmaceutical sciences,” said Bannigan. “We’ve made our data sets fully available so others can hopefully build on this work. We want this to be the start of something and not the end of the story for machine learning in drug formulation.”