A heuristic approach to handling missing data in biologics manufacturing databases

Abstract

The biologics sector has amassed a wealth of data in the past three decades, in line with the bioprocess development and manufacturing guidelines, and analysis of these data with precision is expected to reveal behavioural patterns in cell populations that can be used for making predictions on how future culture processes might behave. The historical bioprocessing data likely comprise experiments conducted using different cell lines, to produce different products and may be years apart; the situation causing inter-batch variability and missing data points to human- and instrument-associated technical oversights. These unavoidable complications necessitate the introduction of a pre-processing step prior to data mining. This study investigated the efficiency of mean imputation and multivariate regression for filling in the missing information in historical bio-manufacturing datasets, and evaluated their performance by symbolic regression models and Bayesian non-parametric models in subsequent data processing. Mean substitution was shown to be a simple and efficient imputation method for relatively smooth, non-dynamical datasets, and regression imputation was effective whilst maintaining the existing standard deviation and shape of the distribution in dynamical datasets with less than 30% missing data. The nature of the missing information, whether Missing Completely At Random, Missing At Random or Missing Not At Random, emerged as the key feature for selecting the imputation method.

Multivariate Data Analysis Methodology to Solve Data Challenges Related to Scale‐Up Model Validation and Missing Data on a Micro‐Bioreactor System

“…Therefore, these missing VCC time‐points compromise the entire run. Previous missing data algorithms specific to bioprocessing utilize only the available off‐line data and single point measurements of on‐line data to infer these missing values, however the novel approach taken here utilizes both high frequency on‐line and low frequency off‐line data to predict these missing values. This approach is advantageous as it exploits all available data recorded during this time period of missing data.…”

“…Therefore, they may not predict the observed nonlinear growth patterns observed in Figure B as the cells shift from exponential stage to stationary stage . Mante et al demonstrated the ability of simple polynomial, logarithmic regression, and mean imputation techniques to successfully predict missing time‐series titre data suitable for secondary analyses. Therefore, these methods can be useful.…”

Missingness analysis of manufacturing systems: A case study

“…They perform duplicate detection but most of the work concentrates on identifying relationships from the data and missingness is not really considered. Mante et al 23 applied mean substitution and regression imputation for handling missing data in biologics manufacturing databases and show it performs well and introduces no bias. However, the missingness was artificially introduced, so the focus was more about minimising the error to a known dataset rather than understanding the missingness patterns.…”

Exploring the potential of machine learning for more efficient development and production of biopharmaceuticals

“…As highlighted by Mante et al, the traditional bioprocessing data often consists of experiments conducted for different projects, products, and cell lines. This eventually leads to batch‐to‐batch variability and missing data points due to instrument or human‐associated technical faux pas 43 . Thus, bioprocessing datasets typically require data pre‐processing steps which involve transformation, normalization, and handling of the missing information 40,43 .…”