Advanced Data Analytics
Advanced data analytics provide the keys to unlock the potential of data and drive the digital transformation in biopharma, across therapeutics discovery, process development and manufacturing. We utilize state-of-the-art methods and algorithms through a hybrid approach, combining AI/deep learning, multivariate statistics, and mechanistic modeling. Our aim to simplify progress drives us to exploit the latest technology developments through open and strategic collaborations with industry partners and academic leaders. All this to accelerate the development of new and better therapies and more affordable medicine to enable — better health for more people.
“Our global team of research scientists think and act as pioneers to pursue the greatest challenges with an open collaborative mindset.”
Prof. Dr. Johan Trygg, Head of Advanced Data Analytics, Corporate Research
Digital representations of biological systems enable digital design of manufacturing systems and virtual evaluation of operating scenarios. Various feed strategies, feed composition, timing and magnitude of pH and temperature shift and other processing options may be explored through what-if type simulation. With the integration of optimization methods, simulation tools can be used to identify target operating conditions. For example, temperature, pH, and feed settings may be found to maximize productivity while at the same time satisfy quality constraints.
Hybrid state-space models of cell culture processes can enhance existing measured data with additional soft-sensor metrics of unmeasured parameters such as accumulation of toxic or inhibitory biomaterials, accumulation of lysed cells, metabolic metrics including specific productivity, growth rate, death rate and many other interesting characteristics. These new metrics are important for the detection of metabolic variability influencing production rate and quality. Sartorius teams of Data Analytics and Umetrics have a long history in multivariate process monitoring. The integration of biologically relevant metrics improves an already industry-leading monitoring platform.
The biopharmaceutical industry is moving towards intensified and continuous manufacturing systems. These next-generation modes of operation offer increased production capacity and present new challenges. In particular, maintaining high productivity and quality for extended durations, 60 to 90 days, for example, requires automation that respond to process and quality variations. Model predictive control strategies derived from hybrid state-space cell culture models are a particular focus with the ultimate objective of providing autonomous control.
We are using systems biology to describe, from first principles, how observed phenotypic behavior arises from the complex molecular dynamics inside of biological systems. By reconstructing the systems piece-by-piece we can facilitate understanding of individual molecular events, as well as the context of these events within the larger integrated system. The ability to simulate phenotypic behavior a priori will lead to a paradigm shift for cell line and process development as well as for biomanufacturing process monitoring and control.
Dynamic System Parameter Identification
Dynamic systems such as cellular growth in a fermenter can be studied using differential equations. These equations have parameters that need to be identified and a level of confidence of this estimation adequately quantified. Indeed, the digital twins built from these identified models must take into account dispersion due to uncertainty on the data used to fit these models and other stochastic behavior.
Multiblock Data Analysis
Due to the diversity of measurement in the biopharmaceutical industry, it is common to face different tables of data related to the same experiment. The integration of these tables is not a trivial task. The difficulty of the exercise is even more significant when these tables are related to time series, and when missing data are present.
Error Propagation in Machine Learning
The number of machine learning tools available is growing every day. However, for most of them, the information on prediction confidence is overlooked. Nevertheless, this information can be critical, especially if the prediction is part of the output of another machine learning workflow or data analytics method.
How to know what your deep learning model does not know
At the core of deep learning are deep neural network models that, despite their flexibility and usefulness, lack common sense and cannot understand when they should not be used. We have developed a novel algorithm to identify the limit of deep neural network knowledge, providing a puzzle piece for their use in safety-critical systems.
Deep learning for label-free cell segmentation in live-cell imaging
In live-cell imaging, segmenting single cells allow in-depth studies of biological heterogeneity but doing it label-free is incredibly challenging. By building a vast dataset manually validated by experts and using the latest advances in deep learning, we achieve unprecedented segmentation performance and help extract more biological knowledge from live-cell images.