A Deep Dive into the Maverick Raman Model

One, Preface:

Past 20 In recent years, Raman spectroscopy has made significant strides in pharmaceutical applications. Polymorph analysis is a leading and unique capability that Raman offers in drug analysis within analytical laboratories, as is the Raman confocal microscopy function for analyzing particles, matrices, and surfaces.

From 2010 Since the end of the decade, the application of handheld Raman systems in the pharmaceutical industry has surged. These instruments are equipped with dedicated operating systems for: GMP Excipients in the environment and API Qualitative analysis, solid dosage form identification, and anti-counterfeiting analysis have now become de facto highly efficient. GMP Raw material incoming inspection standards.

Bioprocess monitoring is a field in which spectral platforms are particularly well-suited. As early as... 20 Century 90 By the end of the decade, near-infrared and mid-infrared spectroscopy systems had already been studied for applications in monitoring metabolites during biological processes; however, water’s strong absorption of infrared light severely limited the range of wavelengths that could be used for absorption measurements. quantity the optical path, thereby leading to detection The background noise is too high. Raman spectroscopy benefits from the relatively weak scattering cross section of water; therefore, research into this application of Raman spectroscopy has been ongoing since the beginning of this century. No It's quite remarkable. Raman technology in optical sensing The sample surface also offers considerable flexibility, regardless of whether plastic is used. material , glass and Other minerals cause very little interference when used as sampling contact surfaces.

Early research on Raman-based biological processes focused primarily on cellular metabolites in various biological systems, and this application continues to expand rapidly as interest grows. Many researchers have also published studies exploring the potential for assessing key product quality attributes, such as post-translational modifications and protein aggregation.

According to Google Scholar the data, in the past 10 Year, and Raman + BioProcess The number of relevant citations is growing exponentially (Figure: 1 ), to 2023 Year, the number of citations will exceed 4000 Next.

II. Challenges of Traditional Empirical Models:

The analysis of Raman data in complex biological systems requires computational assistance. As... Ryder As commented upon, multiple chemical stoichiometries can be employed in this work. quantity Learning and versatility quantity Tool Regarding critical process parameters and critical quality attributes. quantity Attribute ( CPP and CQA ) modeling, the vast majority of literature employs partial least squares. (PLS) Return. Please It is a large class of creep. quantity / It is one of the regularization-based empirical linear calibration methods. The reason it has become overwhelmingly dominant in chemical applications is largely due to historical and commercial factors; however, compared with other methods, it does not necessarily perform better. Nevertheless, all empirical methods do share one common advantage: almost... No Need to know in detail Cell culture environment at the cellular level and the physicochemical principles of analytical instruments.

However, these empirical calibration methods are used to process biological process data. Go Modeling has some weight. The big challenge, as follows:

1 Non-stationarity (Nonstationarity) and homogeneity of variance (Homoscedasticity) In mathematics and statistics, “ Stationarity ” is a term that means each piece of data (in this study, spectral data) is drawn from a random distribution with fixed statistical properties. In most commercial software, Please The empirical method applies only to reason In theory, it is accurate, and... is using “ Stable ” Data in Go Optimized. This means that each biological reaction process It must be transported in the same way. Go and the chemical substances have a consistent phase with each other. Relevance. It also refers to the measurement in the instrument. quantity The variance remains constant across time and channels. Homogeneity of variance). This is not the case for Raman spectroscopy (or near-infrared or mid-infrared spectral absorption), especially in biological processes, when large quantity Biology quantity (Biomass) May lead to biological reaction processes. Go Middle or When the fluorescence differences between different batches are very significant, this leads to noticeable fluctuations in data noise. quantity The difference in level 。

2 Extreme covariance quantity According to the definition, many substances interact with each other during biological reaction processes. Extreme time correlation. Widely used empirical methods aim to leverage these empirical time correlations; however, such correlation-based approaches are highly prone to generating nonspecific associations, thereby reducing prediction accuracy and generalizability.

3 — Exchangeability and Cross-Validation: Related to the two points mentioned above, cross-validation is typically performed as a quasi-validation assessment of empirical models in data modeling tasks. To ensure that the results of cross-validation are valid and representative, the data must be... “ Interchangeable ” ; but due to extreme covariance quantity For this reason, biological process data are often severely... This principle has been violated.

4 Trial-and-error method: Most of these empirical methods involve variation. quantity Select Pre-processing reason a series of options for normalization and correction methods. The recommended method is: “ Give it a try and see. What What seems to work. ” , because usually there isn't any. What What? reason On Dependence According to the guidance for choosing this method, and No Yes Another method.

5 — Quality factor: Related to the above-mentioned content, the primary metric reported in most commercial software is: “RMSEC/RMSECV/RMSEP” ：[ Calibration / Cross-validation / Prediction ] Root Mean Square Error ] Pharmacopoeial analytical standards typically require evaluation of selectivity, linearity, precision, limit of detection, and sensitivity. Go Estimate; but No Fortunately, the empirical modeling approach. No Can directly estimate these. Quality factor. Users can enter. Go Experimental work to evaluate These values—though quite challenging—typically require customized programming. / Analysis.

6 Spectrometer Variations: When developing empirical models, the individual characteristics of a single spectrometer and non- reason The halo effect can also become a developer. Covariate. When More Replace the spectrometer or More Change Laser / When using detectors, it is often necessary to correct for variability. quantity The model ensures alignment with Xinguang. Individual correlations in spectrometers often require the use of various mathematical methods to perform. Go This kind of “ Calibration migration ”。

7 Regulatory Challenges: The black-box nature of empirical modeling approaches necessitates extensive empirical validation efforts to demonstrate their sensitivity, selectivity, linearity, and stability. Regulatory guidance documents (such as ICHQ 1410.3 ) provides some general guidelines, but they do not... No It's special. Don't be explicit, either. No Thus, the mathematical foundation of these methods For theoretical basis.

Given these challenges, it is no doubt that the development and deployment of robust Raman-based methods have consistently posed particularly tricky hurdles in bioprocess applications. Many efforts have already been made. Force To overcome some of these obstacles. Design for deliberate disruption. Experiment Can be used for attempting “ Break ” Intrinsic extreme covariance quantity And expand the scope of empirical data available for modeling. 。

Different studies have reported the use of Please And various pre-positions reason The method was successfully constructed. “ General ” Model, and report on the compatibility of specific platform methods. reason Success; But these jobs usually involve 25–30 Biological reaction experiments conducted more than once require significant time, manpower, and resources; and furthermore... No including the subsequent experimental deployment and maintenance costs. These literature results are consistent with... The content and ideas of the industry seminar reports are basically consistent.

Three, Maverick The brand-new model:

Our goal is to address the technical challenges involved in introducing Raman spectroscopy into the monitoring of biological reaction processes. We focus on mammalian organisms. CHO and HEK293 Starting with cell lines, these cell lines are widely used for proteins. ( Monoclonal antibody ) and the production of viral vectors, and can also be used for scale-up production.

Based solely on pure empirical modeling / Calibration makes it difficult to avoid the challenges mentioned above. Hybrid models are receiving increasing attention in the fields of biology and bioprocess engineering. To date, these approaches have largely integrated knowledge of fundamental biological mechanisms, chemical engineering principles, and computational fluid dynamics. Force study and other fields of knowledge, as well as utilizing some empirical measurements quantity or observational data, to enhance understanding of living things The process of material reaction reason solution . In the model More An excessive number of fixed elements limits empirical optimization, thereby reducing overfitting. / The risk of local minima, and guiding the overall model to achieve interpretable and consistently stable approximations. Utilizing first principles. reason or build Predicting complex outcomes based on block information is sometimes referred to as a completely new approach—such as entirely novel protein structure modeling—which is what we use to describe it. Maverick Terms for algorithm principles 。

MAVERICK The brand-new model originates from 1970 The study of multivariate calibration that began in the early years. (MVC) the probability framework, for example Morgan Early studies on people. It is related to the graph. 2 Contrast the commonly used multivariate calibration models in chemistry.

There are some reference errors. (e) In the case of experience MVC The method is based on the observed spectral data. X ( X~ ) And paired reference data (y) An approximate value to estimate the predicted variable. quantity b ； b The calculation itself is straightforward. The challenges mentioned above 1-7 Mainly manifested in each field. X Regarding the approximation of ‘, what experiments should be conducted, on what hardware, which parameters should be set, and in terms of computation... b How should it have been corrected earlier? / Process the raw data and examine how the final model performs under the conditions it was actually intended for.

X The approximation is crucial for controlling the risk of overfitting in empirical methods, and in practice, there are many, many, many... No Same X (X~) The possibility “ Approximation ”。 Please (Partial Least Squares) is one of many modeling methods and is widely used in numerous commercial software packages. In creating X ( X~ In the process, wavelength ranges are often eliminated or other linear or nonlinear transformations are applied. The excessive number of ‘approximation’ step options available for modeling is a significant secondary source of overfitting; therefore, it sometimes becomes necessary to evaluate hundreds or even thousands of such options, thereby wasting a large amount of generalized degrees of freedom.

In contrast, MAVERICK the brand-new model No Using any empirically observed X or y Data. Instead, it uses graphs. 2 Chinese terms (some static and some dynamic) over time t For proactive testing quantity System created below “ Optimal Linear Predictor ” Although the core of this model is probabilistic, several of its key parameters can be directly derived from first-principles based on optics, electronics, and multivariate statistics. Since these effects are dynamic in Raman systems, several model options for observing biological reaction processes are also dynamic—this is hardly surprising.

Parameters in the formula K, Ψ The representative can observe the possible chemistry of the Raman spectrum. / Biochemistry Contributors “ Main parameters ” as well as the associated probability density function, from which concentration estimates are generated. One might wonder how to account for all possible scenarios in the formula. Although in biological reaction processes, chemical / Number of biochemical substances quantity There are likely thousands of types. However, the sensitivity of Raman spectroscopy means that people actually only need to consider 0.01g/L The above are the main ingredients. In mammalian culture media, more than... 0.01g/L Yes, we have obtained data on hundreds of commonly used substances as well as additives (such as surfactants and defoamers). Using such a large number of parameter data to deconvolve the observed Raman spectra is typically an ill-posed problem; however, by employing a novel model—a fully self-regulating solution—we can generate concentration estimates with low variance.

The remaining conditions depend both on the equipment and on the time. F is from each MAVERICK The system employs an optimal filter function derived from multidimensional factory-characteristic parameters and dynamically adapts in real time to changing sample and system conditions. In Raman systems, many significant errors originate from the design of the optical system and electronic components. MAVERICK Its internal system model enables it to estimate in real time. ∑t the covariance of measurement errors. Correspondingly, the system model also allows And Adaptability, such as changing indoor lighting, temperature, and turbidity conditions. Finally, due to the temporal nature of biological reaction processes, t System status and time t-1 is related to the state; therefore, the inertial model includes both environmental and autoregressive components. Lambda ).

Quality Factor

Several important properties of this estimation model have already been discussed, such as the mean squared error of prediction ( MSEP The analytical solution of ).

As mentioned above, one consistency challenge in empirical model development is the opacity of model attributes. There are few studies that provide standard analytical metrics—such as sensitivity and selectivity—validated by citations from the literature on Raman applications in biological processes to support the models derived. LOD because the literature definition of multivariate models is quite complex. Compliant with IUPAC The defined sensitivity and selectivity factors can be directly estimated using a novel model based on the procedures described in the literature. Finally, other model diagnostics—such as in-plane and out-of-plane consistency—can also be inferred, similar to... Hoteling or leverage statistics and F Parameters:

Four, Rapid model calibration:

MAVERICK systemic MAVERICK The method alleviates the user’s heavy modeling burden but does not completely free them from all forms of... “ Calibration ” . Due to MAVERICK The system is designed for plug-and-play operation between the measurement module, optical path module, and probe. Therefore, before starting the analysis of biological reaction processes, a preparatory step is required to verify the suitability of the quantitative system. This is a... 3 Step by step, by MAVERICK the software in HUB On-screen guidance:

1. Immerse the Raman probe. “LOW” In the standard solution, press ‘ GO And wait for about 4 minutes;

2. Immerse the Raman probe. “HIGH” In the standard solution, press ‘ GO And wait for about 4 minutes;

3. Insert the Raman probe into the reactor and sterilize it together with the reactor.

Step 1+2 Check MAVERICK+ Do some of the probe’s parameters meet the specifications for the new model, and regarding... MAVERICK Quickly calibrate standard products using the new model output generated by a specific combination of the measurement model, optical path module, and probe. This parameter also enables automatic audit trails for probes equipped with serial numbers and chips. MAVERICK Also supports single-point. “ Real-time ” Calibration, which helps eliminate discrepancies between offline analytical instruments and... MAVERICK Data bias between them.

Five, Measured Case:

Figure 3 It shows that, compared to some common offline biochemical analyzers (enzyme-membrane method), the use of... MAVERICK In CHO and HEK293 Process analysis data.

Figure 4 It showcases some background diagnostic information provided by the new model. This information is derived from... CHO Extracted during the cultivation process, which was carried out in a laboratory with large windows. In the image above, the estimated... RMSE ( g/L The small fluctuations observed in the data are fully consistent with expectations—the new model is tracking the fundamental background noise variations throughout the entire day-and-night cycle, influencing... ∑t The same effect is also spreading to the glucose selectivity shown in the figure below, which plots glucose against the preceding... 20 Selectivity of other cell culture medium components: As ambient light levels increase, the new model adjusts and adapts—despite changes in ambient light—to maintain selectivity. Glutathione is shown as a green curve; although it happens to be selectively responsive to glucose in this biological process. “ Minimum ” of the species, but as y As shown by the axis, glucose selectivity remains excellent. >0.99 ).

In the later stages of biological processes, cells / An increase in protein concentration can induce moderate to severe autofluorescence, which, as is well known, poses significant challenges for empirically calibrated models. The superior performance of the new model reflects this effect and can be observed. RMSE a slow upward trend; however, thanks to the new model’s continuous tracking and compensation for increasing background noise, this effect is handled quite perfectly as evidenced by the fluorescence in the measurement error model.

Six, Maverick Limitations and Opportunities of the New Model

The key advantages of the new model—namely, transparency and avoidance of the pitfalls inherent in empirically derived models—can also be viewed as its key limitations. As mentioned above, if the optically active components of a biological process have not been identified in advance, the results reported by the new model are prone to bias. The extent of this data bias largely depends on the optical activity of the “unknown” substances: low micrograms. / Trace metal elements at the elevated level will not have any impact, because... a ) They are optically inactive, b ) The concentration is too low to be detected by Raman spectroscopy in solution. Typically, only... 0.01g/L Only covalently bonded organic substances within the above-mentioned range are considered relevant.

The brand-new model also cannot support the so-called... “ Indirect sensor ” —That is, there is no direct spectral effect (such as pH ), and virtual parameters can also be inferred from empirical observational data. Without spectral effects incorporated in the formula, it would be impossible to use the new model. For those interested in modeling indirect sensors or extending predictive models, they can choose to: MAVERICK The full-spectrum export, which can be achieved through OPC UA Real-time access is also available as a merged data file at the end of the measurement session.

There are even more opportunities to take advantage of. Psi and K the hybrid modeling approach. Currently, a single Psi It seems sufficient for biological processes in mammals, but we are exploring more diverse adaptive mechanisms. Psi Culture medium system (e.g., non- CHO or HEK293 mammalian cells, avian cells, insect cells, and so forth). Alternatively, if specific formulation components that are clearly absent are identified in the data, then— K dynamically constrain the motion. For example, by L1 Regularization methods for models. We note that dynamic system models (such as so-called digital twins) can also be directly connected to entirely new models for continuous updates of time-series data.

Seven, Postscript:

As we move on to other analytes and other cells, / During the media validation process, we have the opportunity to continue expanding. MAVERICK the parameters. Furthermore, as the process transitions from early process development to pilot and production scales, the flexibility of the new model can help improve scalability across different scales. / Process stability of geometric structures. Related to... MAVERICK For numerous examples of performance that has been validated in various culture medium systems, please refer to: 908Devices The app instructions on the website, available at: www.908devices.com 。