Automating regression analysis of heteroscedastic data in non-linear detectors using an integrated CDS platform

Importance of the topic

This work addresses a common analytical challenge: quantitative calibration with detectors that produce inherently non-linear and heteroscedastic signals, specifically charged aerosol detectors (CAD). Objective, statistically rigorous calibration and weighting selection are critical for reliable quantitation of lipid nanoparticle (LNP) components used in biopharmaceutical development. Automating these decisions within a chromatography data system (CDS) improves reproducibility, regulatory traceability, and throughput in method development and routine analysis.

Goals and overview of the study

The study proposes and verifies a predictive modeling workflow that:

• derives intrinsic CAD response from a single reference calibration acquired at a standard detector condition (PFV = 1.00),
• uses an inverse power-law transform (1/p or PFV tuning where y^(1/p)=x) to identify an effective linear operating range,
• simulates detector response at alternative PFV settings and evaluates linearity and residuals, and
• formalizes objective selection of regression weighting (unweighted, 1/x, 1/x^2) using residual diagnostics and confidence criteria, implemented within Empower CDS.

Methodology

The workflow uses a predictive modeling framework combining inverse power-law linearization and weighted least-squares regression. Key statistical measures include slope, intercept, weighted residual sum of squares (WRSS), weighted total sum of squares (WTSS), standardized residuals, root-mean-square (RMS) error, and coefficient of determination (R²). Monte Carlo simulation and forward simulations based on PFV = 1.00 calibration data were used to predict detector behavior at alternative PFV settings and to define PFV percentile limits for linearity. Residual distributions across concentration (log scale) guided selection of appropriate weighting schemes to compensate for heteroscedastic variance.

Instrumentation used

Experimental verification employed a Waters analytical LC setup with Empower CDS (Empower 3.9.0):

• UPLC column: Waters GTxResolve Lipid Phenyl-Hexyl+ RP, MaxPeak Premier, 1.6 µm, 230 Å, 2.1 × 50 mm, operated at 50 °C,
• LC conditions: 3 µL injections, 0.400 mL/min flow, A = 0.1% formic acid in water, B = 50:50 MeOH:MeCN with 0.1% formic acid, 6-min gradient,
• Detector: Charged aerosol detector (CAD) on ACQUITY Premier System; CAD settings included sampling rate ~10 Hz, normal time constant, evaporation temperature ~35 °C, operating voltage ~20 V,
• Data analysis: Empower CDS and Microsoft Excel for simulations and diagnostics.

Experimental design and samples

The analytical model was tested on representative LNP components: ionizable lipid (SM102, ~50%), phospholipid DSPC (~10%), cholesterol (~38%), and PEGylated lipid (DMG-PEG 2000, ~2%). Stock solutions (5 mg/mL in methanol) were diluted in 90:10 water:methanol to produce calibration ranges spanning roughly 0.0004–0.249 mg/mL (covering multiple orders of magnitude). Calibration data were acquired at PFV = 1.00 and experimentally verified at PFV = 1.30 in select cases.

Main results and discussion

The predictive model successfully characterized intrinsic CAD response and predicted optimal PFV settings from a single reference calibration. Key findings:

• Inverse power transformation (y^(1/p)=x) linearized CAD response effectively; adjusting PFV provided a mathematically justified way to define an effective linear operating range without re-acquiring multiple calibrations.
• Simulations of detector behavior at alternative PFVs (e.g., 1.20–1.37 ranges) were consistent with experimental data; experimental verification at PFV = 1.30 matched simulations with relative deviations generally within ±20% for all LNP species tested.
• Residual analysis showed clear heteroscedasticity when using unweighted regression (excess variance at low concentrations) and over-compression with 1/x² weighting. The 1/x weighting produced the most uniform standardized residual distribution across the dynamic range in many cases and was selected as the optimal weighting for PFV = 1.30.
• Goodness-of-fit metrics (R² values) remained high (typically ~0.99), but residual trends and WRSS diagnostics were necessary to select weighting that minimized bias across concentrations.

Benefits and practical applications of the method

The proposed workflow provides practical advantages for analytical laboratories:

• Reduced experimental burden: using a single reference calibration to predict optimal detector settings avoids repeated empirical calibrations at different detector parameters.
• Objective weighting selection: residual diagnostics standardize weighting decisions (unbiased, data-driven selection among unweighted, 1/x, 1/x²), improving accuracy at low concentrations.
• Regulatory readiness: embedding the workflow in Empower CDS preserves traceability, auditability, and reproducibility required for regulated environments and reduces reliance on external spreadsheet procedures.
• Scalability: the CDS-integrated approach facilitates deployment across instruments and labs with consistent reporting and minimized analyst subjectivity.

Future trends and applications

Potential extensions and future developments from this proof-of-concept include:

• Applying the predictive linearization approach to other non-linear detectors (e.g., ELSD) and detector families with known power-law behavior.
• Integrating more advanced statistical tools (e.g., Bayesian methods, adaptive Monte Carlo, or machine learning) to refine PFV selection and uncertainty estimation in real time.
• Automated, closed-loop method optimization where CDS suggests PFV and weighting and then verifies with a small confirmatory dataset.
• Integration with laboratory information management systems (LIMS) and electronic notebooks for streamlined method transfer and regulatory submissions.

Conclusion

This study demonstrates a statistically rigorous, CDS-integrated workflow to automate regression analysis for heteroscedastic, non-linear detector responses like CAD. By formalizing inverse power-law modeling, simulation of alternative PFV settings, and residual-based weighting selection within Empower CDS, the approach reduces analyst subjectivity, accelerates method development, and supports regulatory-compliant documentation. Experimental verification confirmed the predictive model's ability to estimate PFV-dependent linearity and recommend appropriate weighting, improving quantitative robustness across diverse LNP analytes.

References

1. Fekete S, Doneanu C, Addepalli B, et al. Challenges and emerging trends in liquid chromatography-based analyses of mRNA pharmaceuticals. Journal of Pharmaceutical and Biomedical Analysis. 2023;224:115174.
2. DeLaney K, Han D, Birdsall R, Yu Y. Optimized ELSD Workflow for Improved Detection of Lipid Nanoparticle Components. Waters Application Note 720007740. 2022.
3. Han D, Birdsall R, Simeone J, Fogwill M, Yu Y. Comparing ELSD and CAD Performance on Polysorbate Quantification in Infliximab Drug Products. Waters Application Note 720007501. 2022.
4. International Council for Harmonisation (ICH). ICH Q2(R2): Validation of Analytical Procedures. 2023.