Machine Learning

Deep dive into mouse brain

3D light-sheet image data analysis enables analysis of the entire organs and not just a small subset of the tissue. We built a tailored statistical pipeline, which includes machine learning and deep learning components, for automated analysis of costume readouts.

Investigation overtime prediction based on a mixed logistic regression model

Overdue investigations can delay the release of an investigational medicinal product to clinical trials. These delays could affect the start of new clinical studies and impact existing ones.

A significant percentage of good manufacturing practice (GMP) deviation investigations are not closing on the specified time. For this reason, we created a tool based on a mixed logistic regression model that predicts which investigations will be overdue and which will be closed on time with high accuracy (around 70%). Further, the tool which has already been tested by the Quality Assurance helps to better understand the reason behind the delay. Based on this there is an optimization mechanism which allows the user to make adjustments that can reduce the probability of going overtime. Additionally, except from the prediction analysis part the tool includes a live performance tab and automatic mechanisms, like email notifications and pop-up windows in case of significant drop of the accuracy and it can be updated automatically including new data on the training set. The tool is already tested for more than 6 months and is patent pending.

Enhancing In-Line Microscopy Monitoring and Particle Size Analysis with Neural Networks: Lessons from a Challenging Project

Chemical engineers in MSD utilize the crystallization process for material (drug) development. They examine the effects of process conditions on material properties such as particle size.

The current state of the crystallization process of our interest is that the experiments must be run at full length and at the end evaluated by an external company. There is an opportunity to evaluate these experiments in-house during the crystallization process itself. This group was testing special equipment for in-line process monitoring called PVM that captures images of the material during the experiment.

The supplied software was not able to provide reliable measurements for their data. Further investigation has shown that even commonly available image analysis methods are not sufficient.

We tried to tackle this problem and tried to make progress with a deep learning-based strategy consisting of two key elements: Cycle-consistent Generative Adversarial Network (CycleGAN) to generate credible image-label pairs and a Mask-RCNN model that uses the generated data for particle detection. While our methodology streamlines the creation of training data and shows improvements in particle detection accuracy compared to previous methods by eliminating the need for manual labeling, the results are still a work in progress and not yet ready for practical applications.