Use cases range from theoretical to operational to simply learning; a few of them are outlined here:
- Control Strategy Robustness
- Machine Learning
- Meta-Developmental (Platform) Optimization
- Equipment Selection
- Teaching and Learning
- Macroscopic Modeling
Control Strategy Robustness
A control strategy that works in theory may fall apart when it meets an actual process. A PID that is stable in offline tests may actually be unstable when faced with cells in exponential growth; a glucose feed that doesn’t account for compounding errors in assays may result in glucose concentration reaching 0 between offline samples; a perfusion feed strategy that assumes constant growth may wildly over-(or under)-shoot osmo levels.
In each of these cases an error could have been highlighted by simulating the process along with its interaction with the control strategy without ever having to run a set of experiments that are doomed to fail. Even in the cases where a platform would work with a specific cell line, CC-Sim’s ability to automate testing the entire process with a diverse range of cell lines allows for an assessment of how robust the platform really is. While CC-Sim will not catch all errors that one may encounter when developing a process, any screening will still pay for itself several times over.
Machine Learning
Machine Learning is notorious for requiring large amounts of data. In the modern age this is becoming increasingly feasible, as all data collected is organized and stored for future use; however, there are any number of reasons why we may not presently have the data we need.
CC-Sim was originally designed to provide an alternative: a large amount of data generation that trades off accuracy for low-cost. A Machine Learning technique that performs well on the simulated data is more likely to perform well on real data, because it must be able to handle the similarities:
- Errors in measurements
- Time-dimensioned data with inconsistent sample times
- Hidden correlations between data (E.G. cell diameter -> cell volume -> productivity / consumption)
- Cell lines that are each unique, yet correlated
- Variable experimental setups
Data from CC-Sim is great for all kinds of machine learning, but may be of particular interest to those aiming to develop reinforcement learning (RL) techniques. In reinforcement learning, the ML algorithm must have the freedom to set up a process as it wants; it must therefor be designing and running its own experiments, rather than simply using past data. Current RL uses enormous quantities of runs, and so far has almost exclusively been limited to digital processes and easily automatable data generation, such as games and simulations.
In process development, the cost of experiments is high, and any RL technique employed would not make efficient use of those experiments and would be cost prohibitive. A RL technique applied specifically to process development could be developed and tested with CC-Sim. Techniques developed in this manner are ideal because they can be developed with nearly no experimental cost before being applied to real processes.
Meta-Developmental (Platform) Optimization
Process development is largely about optimizing a process; yet the optimizing a process is itself a process. It is important to be able to optimize with as few experiments as possible while also testing the broadest range of conditions to be able to ensure a robust understanding of the process.
A platform is almost the meta-process for process development; it includes the initial setup to be run with along with the steps of changing the process in response to results. Such a platform could be informed by CC-Sim, for example by running different numbers of experiments on different conditions and noting whether any gaps in process knowledge or failure to optimize occur.
Equipment Selection
How does one choose whether to switch to a different brand of pH probe which has half the amount of drift yet will cost an additional $50k over 5 years? It can be incredibly difficult to quantify how errors compound to affect not only increased process failure but also a decreased efficiency in process optimization.
CC-Sim includes all of the error parameters in a process, and can be run using different error parameters in order to calculate what the marginal effect of each parameter is on the process. It becomes possible to point to reduced-drift pH probes and say that adopting them has results indicating 0.5% higher titers, allowing a cost-benefit analysis to be run.
Teaching and Learning
If you’re like me, you learn much better by actually being able to experiment with something and see the results in near real-time. Trying various things and seeing the consequences of those actions lead to a more intuitive process understanding. CC-Sim makes it possible to both teach and learn about process development by allowing the generation of data without needing access to a lab and hundreds (or thousands?) of hours of work.
Teachers could generate data to give examples of whatever changes they are talking about. Students could have practice optimizing a process with limited (realistic) access to data.
Macroscopic Modeling
Say you hypothesize that cell growth can be modeled as cells that are inactive, active and dividing, and active and growing – with cells switching between these three states. That might explain why cells have a growth delay, and how during different phases cells might preferentially increase cell density or cell volume. Using CC-Sim, this model could be coded into how the cells work, and predictions made for how the cells would behave.
Data generated by CC-Sim with these changes could be compared to past data or tested in a planned experiment. Such modeling allows for increased process understanding or even the design of novel systems that take advantage of predicted cell behavior.