This is an ambitious endpoint and should not be thought of as simple or quick to achieve. PATs goal is to help pharmaceutical companies achieve the same level of understanding and control over the final product already enjoyed by the chemical, petroleum, polymer and food industries.
The only caveat is that life-saving drugs are not potato chips. Predicting the salt level of chips is not on par with predicting the blood level of a drug. Saying that a process resembles previous processes does not guarantee product equivalence without some other safeguards in place.
Why NIR is a Major Actor in PAT
I will attempt to summarize the reasons succinctly. NIR, by its nature and history of application, became the first analytical technique to require computer algorithms. Used in the food and agriculture industries for years, it was later discovered by other industries: chemical, polymer, paper, textile and, yes, even pharmaceutical.
Normally, NIR is applied to mixtures of materials (including natural products), usually in the diffuse reflection mode. The signal generated contains both chemical and physical information; the physical usually dominates the spectrum. Because of this, higher math is needed to extract the chemical information. Principal Component Analysis (PCA) and Partial Least Squares (PLS) are two commonly used algorithms.
What should occur, if chemometric principles are applied correctly, is that through appropriate design of experiments (DoE) and selection rules, process samples are used to build NIR equations. The analyst must use samples with as many different attributes and from as many manufacturing conditions as one expects to find in manufacturing and clinical settings. Because we cannot envision every possible variation in the process (differences in particle size, new raw material sources, levels of hydration, new/ rebuilt process equipment, polymorphic changes, blending distributions, etc.) or every possible clinical ramification of the designed dosage form, we need to build in a change mechanism the control space when we construct the equation.
To avoid having to go through the currently required change control process every time a process hiccup or suspect sample arises, manufacturers can address changes as they occur using process control strategies. Thus, when the equation gives a suspicious prediction, we have a mechanism for dealing effectively with the change in real time. That is, we design the process and the algorithm to anticipate gaps and to correct predictions in the face of potential process changes.
Despite having to defend the term predict to management, that is what Near Infrared does: predict, not assay. The legal definition of assay is simply that, in an assay, a method compares an analyte directly with a standard material. For instance, HPLC might use a USP reference to compare with the substance in a tablet. We inject known concentrations of the reference sample into the HPLC before and after unknown samples. The identity and concentration of the unknown are calculated from a simple Beers law plot at a set wavelength. Titrations, both for assay and Karl Fisher for water, polarography, gas chromatography, and any other assay technique that uses a reference sample is, by definition, an assay.
When a standard reference tablet, capsule, or an intermediate material is not available (essentially in every pharmaceutical process), an equation is built from historical samples. The instrument itself becomes the reference. We measure the wavelength accuracy, light level, noise, linearity and a few other parameters to ensure the instrument is performing to manufacturers specifications. We call the analysis, based on previous samples and a well-tuned instrument, a prediction.
The simple reason that process tests such as NIR are the backbone of PAT is speed and ease of use. Most assays are both slow and destructive. Both attributes are drawbacks for large samples and a fast process. What PAT seeks is large numbers of rapid results that indicate where the process is and where it is going.
Chemometric analysis of monitoring instruments provides this control. Methods such as NIR, FT-IR, thermal effusivity, acoustics, Raman, rapid LC, pH, temperature, flow rates, etc. give a signal used in following the process. Based on previous lots and their associated historical data, adjustments bring the batch into compliance. Making sure each step gives the same signals is one way to assure that each lot of material will produce the same product each time.
Thus, with reasonable assurance, we can adjust a process to meet certain parameters and expect a product that resembles past lots. These monitors are a godsend to production managers. When PAT reaches its true efficiency, there will no longer be OOS (out of specification) investigations, simply because no lots will ever be OOS.
While the process itself is of great value to control, it is important to keep the purpose of the process in mind: to deliver the proper medicine, in the proper dosage form, to patients. Whether we are making simple aspirin for headaches or immune suppressants for liver transplants, we are making medicine, not potato chips! Two key features are part and parcel of the finished product: it must contain the correct amount of (pure) drug substance AND it must deliver it in the proper fashion.
The correct amount of API (active pharmaceutical ingredient) is critical. In fact, most of our release testing focuses on the assay of the API. Another important test measures how quickly it releases from its matrix. While we design immediate-release products to explode in the stomach and release the drug all at once, controlled-release forms are more complicated. These are designed to deliver a controlled and metered amount of the active to a specific part of the GI tract over a specified amount of time.
After an innovator produces a new drug and establishes the release rate for efficacy, generic manufacturers will attempt to mimic this pattern with their products. Why make them the same? They mimic the innovator profile to ensure the delivery (hence bioequivalence) of their drug substance. Currently, one of the best ways to measure this release rate in the QC lab is dissolution.
Just what is dissolution and why does it matter? Dissolution is placing a single, solid dosage form (sdf) in a controlled (usually aqueous) environment and, via various measurements, observing the rate at which it releases into the solvent. In one configuration, a paddle agitates the solvent with the sdf at the bottom of the vessel, allowing the drug to dissolve in a gentle, reproducible manner. Another method is to place the sdf in a small basket, which is spun to effect dissolution.
During product development, studies are performed to correlate this dissolution rate with blood levels in clinical trials. This in vivo/in vitro correlation shows that dissolution reasonably measures what the sdf does in a patient. Does the blood level reach a therapeutic, ineffective or toxic level?
On the other hand, this test, when used as a QC tool, indicates whether a batch of product, placed on stability testing, will release in a manner similar to the clinical studies. It is merely a prediction, not truly an assay.
In itself, thats not a bad thing. Surely companies cannot perform thousands of clinical trials to measure blood levels of all drug lots and all stability time points. If it were even physically possible, cost and time factors would place the retail cost of drugs beyond most peoples reach. So, we are left with a QC tool called dissolution. Considering we have not been producing bad products, it must work.
However, recall that dissolution is a prediction with a tenuous correlation to reality (i.e. bioavailability). NIR is used to predict a method that is, in itself, a prediction. Having used NIR to predict dissolution (under controlled, limited circumstances) since the late 1980s, I have found that an amazing number of parameters affect the rate of release: particle size (API, excipients), moisture (amount, location), hardness of the tablet, cluster sizes of drug, etc., etc.
Much work is being done, using wonderful techniques such as chemical imaging, to elucidate which factors affect release rates and why. However, we are in the Kitty Hawk stage as far as 100% predictability of dissolution using NIR. When I read reports of a 13-factor PLS equation being used to predict dissolution times, I have great trepidation. In some cases, we are approaching astrology, not science.
My conclusions are simple: NIR is a tool. Like any other tool, it has limitations, one of which is lack of a direct reference material. We have a long way to go before we can consider abandoning final testing completely. The major improvements in PAT will be: months of warehouse time obviated, tons of materials NOT destroyed, and eventually, better products produced. In the meantime, we need to keep repeating: Drugs are not potato chips. After whittling the process time from months to days, how much can a few hours of lab time cost?