For granulation endpoint or cell cultures, for example
This application describes collaboration with Biogen Idec; the results have been presented at the Bioprocess International Summit, June 2007, and at IFPAC, January 2008.
Background: meeting the endpoint objective
A major issue in the pharmaceutical and life science industries is predicting the endpoint of a process. For example in the granulation step during tablet manufacture, or in cell culture in the biologics area. There are one or more objectives that are used to determine endpoint, such as:
- obtaining complete mixing
- obtaining a product that has the desired properties for downstream processing
- maximize the batch yield
- maintain undesirable values above/below certain constraints
The problem is usually that these objectives can only be quantified much later, sometimes weeks, after the batch is completed.
Current practices and what can be learned
The current practice in many industries is to use heuristics:
- an operator with experience on that process "knows" when the endpoint is reached
- quick lab tests are used to measure surrogates for the endpoint objectives
- a recipe is strictly followed
- use trigger points based on on-line measurements from the process, for example, stop the batch when the theoretical energy input reaches 300kJ.
But these may be prone to failure:
An experienced operator is a valuable resource, but problems arise if they are reassigned or go on vacation. Very often the operator's experience is embedded in the on-line process measurements, or could otherwise be easily instrumented.
Intermediate lab tests provide excellent insight into the batch progress, but more can be done by merging this data with the on-line process measurements.
A recipe based system can fail if the raw materials or process conditions differ from those normally experienced. If changes in the raw materials or process are present, then the recipe must be adjusted to obtain the expected final properties - this is the principle of feedback control. It is similar to how an experienced operator will compensate for upstream changes.
Trigger points and calculations from the raw measurements are a start: more can be done by combining various measurements by using multivariate data methods.
How ProSensus assists Biogen Idec
The current practice at Biogen Idec for a particular product is to measure viable cell density during the batch progress. This number is then used to predict the remaining batch duration and to adjust nutrient feed rates. The objective of the cell culture is to maximize yield, while maintaining certain quality variables within specified limits.
However, data was available every few seconds for pH, dissolved oxygen and a few other variables. The raw materials charged are also characterized with some laboratory measurements. And, the cell viability and cell density at one or two days into the batch are also used.
All this data is combined in a single PLS (projection to latent structures) model. The data structure is shown below - Z contains the raw material characterization, X contains the real-time trajectories (pH, etc), and the model predicts the values in Y.
The model now uses all available data (Z and X) and provides a continuous, real-time prediction of all final quality attributes (Y) as the batch nears its end. The batch is terminated (cells are harvested) when the main quality attributes show signs of deteriorating.
The figure shows the predictions for several batches. In particular, one quality metrics for batch 17 [red line] is shown to peak and then drop off slowly. The optimal termination point is at the peak.
