Luedeking-Piret regression for multi-step-ahead forecasting and clone selection in monoclonal antibodies biomanufacturing Authors: Penghua Wang, Deepika Verma, Yuk Chiu, John Klier, Chongle Pan Published: 2025 ABSTRACT Efficient production of monoclonal antibodies (mAb) using Chinese Hamster Ovary (CHO) cells is central to pharmaceutical biomanufacturing. The clone selection process traditionally requires lengthy 7-to-14-day assessments to evaluate performance, which extends development timelines. Here we introduce a hybrid Luedeking-Piret Regression model that integrates mechanistic insights with machine learning to more accurately predict mAb yields in fed-batch CHO cultures. Using experimental data from the early growth stages (up to day 9) of seven (n=7) distinct CHO cultures, the model performed multi-step-ahead forecasting to predict final production. The model predicted monoclonal antibody titers on day 16 with a mean percentage error of 5.85%, correctly selected higher-performing clones in 76.2% of trials from leave-two-out cross-validation and accurately forecasted daily production trajectories from day 10 to day 16. The model’s multi-step-ahead forecasting capabilities have the potential to accelerate clone selection, providing the biomanufacturing community with a computationally straightforward algorithm for predicting production yields. Introduction The biomanufacturing industry commonly uses genetically engineered Chinese Hamster Ovary (CHO) clones to produce therapeutic proteins such as monoclonal antibodies (mAbs)1. Genetic modification of CHO may generate a heterogeneous population of cells that have integrated the gene of interest in different positions in their genomes2. Clones derived from these cells have different performance metrics for commercial production, including product yield, stability, and quality3. Clone selection aims to identify the best-performing clones through multiple rounds of screening. Mid-term assessments during clone selection generally use small-scale fed-batch cultures that simulate the conditions of large-scale bioreactors4 and measure clone performance over 7–14 days of cultivation5. Mid-term assessments allow for the evaluation of how well the clone performs under production-like conditions, including growth, productivity, and metabolic profile. In comparison with short-term assessments lasting only 4–8 days6, mid-term assessments provide more detailed information on productivity, growth kinetics, maximum cell density, and clone viability7. However, mid-term assessments are more time-consuming than short-term assessments. In this study, we hypothesize that the final mAb performance metrics of clones from mid-term assessments can be predicted from their growth data acquired in earlier cell culture stages using a predictive model. If this hypothesis is validated, biomanufacturers can reduce the number of cell culture days required for mid-term assessments and still select the best-performing clones based on their projected performance. The performance of CHO clones is influenced by complex biological and environmental processes8 and can be modeled using mechanistic models that are grounded on our understanding of these underlying processes. A foundational example is the work of Luedeking and Piret9, who studied glucose to lactic acid fermentation kinetics at constant pH levels between 4.5 and 6.0. They developed a mathematical model that related the instantaneous rate of acid formation (dP/dt) to both bacterial growth rate (dN/dt) and bacterial density (N), using pH-dependent constants α and β to characterize the fermentation process. Building on this foundation, López-Meza et al.10 demonstrated that simple kinetic models, including a Luedeking-Piret-type model, could effectively describe monoclonal antibody (mAb) formation in CHO cells across various culture conditions. Hogiri et al.11 developed and validated a pH-dependent dynamic model to optimize pH-shift strategies in CHO cell cultures, successfully increasing mAb by 40% and thereby demonstrating the effectiveness of mechanistic modeling in enhancing mAb manufacturing processes. However, our theoretical understanding of many biomanufacturing processes remains insufficient to develop accurate mechanistic models. Clone performance, such as mAb yields, can also be modeled using a data-driven machine learning approach. Gangadharan et al.12 achieved improved predictive accuracy on key variables of interest in biologics manufacturing by integrating multiple machine learning models. Sokolov et al.13 demonstrated that combining partial least squares regression (PLS2) with genetic algorithms could effectively predict monoclonal antibody quality attributes from cell culture process conditions. However, developing accurate machine learning models requires a large amount of training data collected from a variety of conditions. By combining the strengths of mechanistic models and machine learning models, hybrid modeling has recently emerged as an approach to the optimization and control of bioprocesses14. Le et al.15 used early-stage bioprocess data from 243 runs and developed support vector machine and partial least squares regression models that can effectively predict final antibody and lactate levels. Yeo et al.16 employed multivariate statistical and flux balance analyses to identify nutrient limitations, such as glucose and amino acid depletion, in CHO cell culture media. They demonstrated that partial least squares regression can effectively correlate these nutrient levels with cell growth and productivity and offer a strategy to optimize media formulations for enhanced CHO cell performance. By integrating mechanistic insights with empirical machine learning techniques, hybrid models can accurately predict key performance indicators including viable cell density (VCD), titer, and glucose consumption, thereby offering a framework for the optimization and control of CHO cell bioprocesses. Most of the existing studies focused on one-step-ahead forecasting (OSAF), which used the input data up to a given day t to predict the mAb yield of the clones for the immediate next day t + 1. The prediction of mid-term performance of clones using their growth data from earlier stages requires multi-step-ahead forecasting (MSAF) to predict final mAb production on target day n (n > t). A one-step-ahead model can be used to make MSAF using an autoregressive approach, in which the predicted value from the previous time step is recursively fed back into the model as input for predicting the next time step and this process is continued until the desired number of future steps is forecasted. However, this approach has not been used in practice because forecast errors at each step can propagate and compound, leading to increased uncertainty and reduced accuracy over extended horizons. Alternatively, a model can be trained to directly predict final titer without generating predictions for each intermediate time step between t and n. Park et al.17 conducted a systematic evaluation of data-driven models for CHO cell cultures. Their study using an initial dataset 24 CHO cell culture batches across four different conditions compared five different MSAF strategies. These included recursive, direct strategies, Multiple Input Multiple Output, where a single model predicts the entire forecast horizon at once, and two hybrid strategies, such as Direct and Recursive and Direct and Multiple Input Multiple Output. By examining 122,400 distinct models combined these five multi-step strategies with various algorithms and input features they concluded that the optimal forecasting strategy is highly dependent on the specific prediction target (e.g., VCD, IgG, or metabolites). Furthermore, their work emphasized that the selection of these data-driven model elements necessitates a careful balance between computational load and the desired model accuracy and reliability. Mid-term clone selection processes are typically conducted under standardized experimental conditions to control environmental factors and ensure consistent evaluation across different clones. While modern machine learning approaches offer results for developing predictive models to accurately forecast final mAb yields, their performance is often constrained by the limited training data available. In the bioprocess, relatively few experimental runs or batches are available for analysis compared to the parameters being monitored. A challenge in developing such predictive models is the biological variability and complex nonlinear dynamics of CHO cell systems. The overall bioprocess is characterized by distinct phases, including an exponential growth phase followed by stationary and death phases. This nonlinearity makes direct linear regression across the entire culture duration ineffective. However, following the temperature shift designed to halt growth and maximize productivity, the mAb production phase itself often exhibits near-linear accumulation dynamics. In this study, we developed a hybrid model based on a Luedeking-Piret-type model using key process variables including VCD, growth rates, and mAb concentration. We derived a mechanistic mathematical relationship between the early mAb production phase (from temperature-shifted day t to t + k, where k is the number of steps ahead of t) and the target-day mAb production. This relationship was used to develop a Luedeking-Piret Regression (LPR) model for the prediction of end-of-cycle mAb production yields. Then we demonstrated the LPR performance on seven distinct clones using their growth data up to day 9 to predict the end-of-cycle productivity on day 16. The MSAF strategy was further extended to predict the mAb production on days 10, 11, 13, 14, and 15 using the input data from the first nine days. The LPR models predicted the entire production trajectory for a range of target days using earlier-stage input data. The MSAF results across an extended time horizon demonstrated the robustness and accuracy of LPR modeling. The accurate prediction results potentially supported the feasibility of accelerated clone selection based on shortened mid-assessments. Methods Cell cultures and process variables In this study, six distinct clones (i.e., clone 1 to clone 6) and a bulk pool were evaluated for production of a mAb protein. The clones were selected from a bulk pool. Bulk Pool was single cell seeded into 96-well plates and tested for clonality. Clones were monitored for productivity and the top 50% of clones were moved forward at each scale-up step. The clones in this study are among the top 12–24 clones. No additional engineering or genetic modification was performed. These clones were subjected to identical feeding strategies and experimental conditions in fed-batch 250-ml bioreactors according to the industrial standards18. Each clone was grown in duplicate cultures. All 14 cultures were grown for 16 days under the same conditions over this period except that the growth temperature was shifted on day 5 from a high temperature to a low temperature. The temperature shift was implemented to transition the CHO cells from a high-growth phase to a high-protein-production phase19. The key process variables analyzed included viable cell density (VCD, vc mL−1), %VC, product titer (Prod Titer, mg L−1), glutamine (Gln, mM), glutamate (Glu, mM), glucose (Gluc, g L−1), Lactate (Lac, g L−1), ammonium (NH4+, mM), sodium (Na+, mM), potassium (K+, mM), and osmolality (Osmo, mOsm kg−1). These variables were measured at various time points throughout the cell culture process. Daily sampling and feeding protocols were followed, but only the sampling data from the cell culture process, which measured a more comprehensive set of variables were utilized for this study. Kinetic analysis Precise monitoring and regulation of cellular growth are key for optimizing production efficiency and maintaining consistent product quality. The daily specific growth rate, rx, is mathematically expressed as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{x}=\frac{\ln ({{{{{\rm{VCD}}}}}}_{i})-\ln ({{{{{\rm{VCD}}}}}}_{i-1})}{\Delta {t}_{i}}\times 24$$\end{document}rx=ln(VCDi)−ln(VCDi−1)Δti×24 VCDi denotes the viable cell density at the ith time point, and Δti represents the time interval between consecutive measurements. The multiplication by 24 adjusts the rate to a daily growth rate, providing a standardized metric for assessing cell culture dynamics. In cellular metabolism, the ratios of glutamine to glutamate (Gln/Glu)20 and glucose to lactate21 serve as indicators of a cell’s energy balance and biosynthetic capacity. A higher Gln/Glu ratio reflects a prioritization of anabolic processes and efficient energy production, suggesting that cells are actively engaging in protein and nucleotide synthesis essential for growth and maintenance22. Similarly, a higher glucose to lactate ratio indicates a reliance on oxidative phosphorylation, signifying effective glucose utilization with minimal lactate production23. Conversely, lower ratios in both Gln/Glu and glucose to lactate imply a metabolic shift towards increased glycolytic activity and rapid proliferation, often associated with the Warburg effect and metabolic stress24. Supplementary Table S1 details metabolite production ratios and other ratio definitions. Missing value imputation Since each clone had duplicated runs, missing values in one run were imputed using the corresponding values from the other run. For example, one of the Bulk Pool runs was missing the product titer on day 6. In this case, the productivity of the duplicated run was calculated and then the calculated productivity × VCD was used to impute the missing product titer. The productivity is defined in Supplementary Table S1. For variables such as Gln and Glu, the values from the duplicated run were used directly. Three methods were evaluated to impute the missing Osmo value for day 7: linear interpolation, multiple imputation, and a Bayesian linear model. The results of these evaluations are shown in Supplementary Fig. S1. As osmolality exhibited a linear trend during this phase of the cell culture, we selected linear interpolation to estimate the final value. For day 12, Gln, Glu, and Prod Titer values were missing. Consequently, the day 12 measurements were excluded from the model building process. Modeling of mAb production using the Luedeking-Piret regression model (LPR) A Luedeking-Piret-type model was employed to model the production of monoclonal antibodies (mAbs). The model is described by the following differential equation:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{d[P]}{dt}={r}_{p}=\alpha {r}_{x}+\beta [X]$$\end{document}d[P]dt=rp=αrx+β[X]Where:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[P] = \,{{\mbox{concentration of mAbs at time}}}\,t,\\ {r}_{p} = \,{{\mbox{rate of mAb accumulation within the reaction vessel}}}\,,\\ {[X]} = \,{{\mbox{cell concentration at time}}}\,t,\\ {r}_{x} = \,{{\mbox{rate of cell growth}}}\,,\\ \alpha ,\beta = \,{{\mbox{Luedeking-Piret constants}}}\,.$$\end{document}[P]=concentration of mAbs at timet,rp=rate of mAb accumulation within the reaction vessel,[X]=cell concentration at timet,rx=rate of cell growth,α,β=Luedeking-Piret constants.The Luedeking-Piret equation for current time t can be expressed as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{t}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t$$\end{document}Pt=Pt−1+αt−1rxt−1+βt−1Xt−1Δtwhere:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \hspace{0.5em}{P}_{t}\,\,{\mbox{is the antibody concentration at the current time step}}\,\,t,\\ \hspace{0.5em}{P}_{t-1}\,\,{\mbox{is the antibody concentration at the previous time step}}\,\,t-1,\\ \hspace{0.5em}{\alpha }_{t-1}\,\,{\mbox{and}}\,\,{\beta }_{t-1}\,\,{\mbox{are the Luedeking-Piret constants at time step}}\,\,t-1,\\ \hspace{0.5em}\Delta t\,\,{\mbox{is the time step used in the numerical integration.}}$$\end{document}Ptis the antibody concentration at the current time stept,Pt−1is the antibody concentration at the previous time stept−1,αt−1andβt−1are the Luedeking-Piret constants at time stept−1,Δtis the time step used in the numerical integration.For the next time step, t + 1, the expression is:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{t+1}={P}_{t}+\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t$$\end{document}Pt+1=Pt+αtrxt+βtXtΔtwhere:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{t+1}\,\,{\mbox{is the antibody concentration at culture stage}}\,\,t+1,\\ {P}_{t}\,\,{\mbox{is the antibody concentration at culture stage}}\,\,t,\\ {\alpha }_{t}\,\,{\mbox{and}}\,\,{\beta }_{t}\,\,{\mbox{are the Luedeking-Piret constants at time step}}\,\,t,\\ \Delta t\,\,{\mbox{is the time step used in the numerical integration}}\,.$$\end{document}Pt+1is the antibody concentration at culture staget+1,Ptis the antibody concentration at culture staget,αtandβtare the Luedeking-Piret constants at time stept,Δtis the time step used in the numerical integration.Substituting Pt into the equation for Pt+1:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{t+1}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t+\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t$$\end{document}Pt+1=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+αtrxt+βtXtΔt\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\vdots$$\end{document}⋮Let Pn denote the antibody concentration at final step n, then:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{n}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t$$\end{document}Pn=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+⋯+αn−1rxn−1+βn−1Xn−1ΔtNow, let us denote:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t,$$\end{document}c=αtrxt+βtXtΔt+⋯+αn−1rxn−1+βn−1Xn−1Δt,then:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{n}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t+c$$\end{document}Pn=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+cAlternatively, let:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=\left({\alpha }_{t+1}{r}_{{x}_{t+1}}+{\beta }_{t+1}{X}_{t+1}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t,$$\end{document}c=αt+1rxt+1+βt+1Xt+1Δt+⋯+αn−1rxn−1+βn−1Xn−1Δt,then:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{n}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t+\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t+c$$\end{document}Pn=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+αtrxt+βtXtΔt+cAlternatively, let:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=\left({\alpha }_{t+2}{r}_{{x}_{t+2}}+{\beta }_{t+2}{X}_{t+2}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t,$$\end{document}c=αt+2rxt+2+βt+2Xt+2Δt+⋯+αn−1rxn−1+βn−1Xn−1Δt,then:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{n}={P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t+\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t+\left({\alpha }_{t+1}{r}_{{x}_{t+1}}+{\beta }_{t+1}{X}_{t+1}\right)\Delta t+c$$\end{document}Pn=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+αtrxt+βtXtΔt+αt+1rxt+1+βt+1Xt+1Δt+c\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\vdots$$\end{document}⋮Let k denote the number of steps ahead of t, where t < k < n − t, let:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=\left({\alpha }_{t+k+1}{r}_{{x}_{t+k+1}}+{\beta }_{t+k+1}{X}_{t+k+1}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t,$$\end{document}c=αt+k+1rxt+k+1+βt+k+1Xt+k+1Δt+⋯+αn−1rxn−1+βn−1Xn−1Δt,then:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{n}= {P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t\\ +\left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t\\ +\cdots \\ +\left({\alpha }_{t+k}{r}_{{x}_{t+k}}+{\beta }_{t+k}{X}_{t+k}\right)\Delta t\\ +c$$\end{document}Pn=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+αtrxt+βtXtΔt+⋯+αt+krxt+k+βt+kXt+kΔt+c In general, let k denote the number of steps ahead of t, Pq is the antibody concentration at culture stage time q, where t < k < q < n, let:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=\left({\alpha }_{t+k+1}{r}_{{x}_{t+k+1}}+{\beta }_{t+k+1}{X}_{t+k+1}\right)\Delta t+\cdots +\left({\alpha }_{n-1}{r}_{{x}_{n-1}}+{\beta }_{n-1}{X}_{n-1}\right)\Delta t,$$\end{document}c=αt+k+1rxt+k+1+βt+k+1Xt+k+1Δt+⋯+αn−1rxn−1+βn−1Xn−1Δt,then:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{q} = \,{P}_{t-1}+\left({\alpha }_{t-1}{r}_{{x}_{t-1}}+{\beta }_{t-1}{X}_{t-1}\right)\Delta t\\ + \left({\alpha }_{t}{r}_{{x}_{t}}+{\beta }_{t}{X}_{t}\right)\Delta t\\ + \cdots \\ + \left({\alpha }_{t+k}{r}_{{x}_{t+k}}+{\beta }_{t+k}{X}_{t+k}\right)\Delta t\\ + c$$\end{document}Pq=Pt−1+αt−1rxt−1+βt−1Xt−1Δt+αtrxt+βtXtΔt+⋯+αt+krxt+k+βt+kXt+kΔt+c This deduction indicates that the final titer Pn (Eq. 14) is influenced by prior antibody production, cell concentration, and the rate of cell growth. More broadly, Eq. (16) generalizes this relationship for any future time point Pq. Consequently, a predictive model can be constructed to forecast mAb production at any future time step (e.g., day 16) by utilizing sequence interval data from earlier stages, ranging from t to t + k, of the production cycle. The Luedeking–Piret model provides a mathematical framework for both one-step-ahead forecasting (OSAF) and multi-step-ahead forecasting (MSAF) of mAb production by leveraging product titer, VCD, and growth rate from earlier stages of the production cycle. Partial least squares regression (PLSR) is a dimensionality reduction technique that creates new variables (called latent variables or components) from the original input variables. Unlike standard regression that directly relates inputs to outputs, PLSR finds linear combinations of input variables that best explain both the input patterns and the target variable simultaneously. This approach is useful in bioprocessing where variables are often highly correlated and the number of measured variables approaches or exceeds the number of experimental runs. By creating these latent variables, PLSR effectively captures the most important relationships while reducing noise and avoiding overfitting that commonly occurs with limited bioprocess datasets. When applied to Luedeking-Piret-derived features (titerPt, VCDXt, growth raterx,t across days t to n), PLSR generates a latent variable structure constrained by and reflective of fundamental bioprocess mechanisms. The parameter of the [P]t, α and β in the Luedeking-Piret model is estimated using PLSR, the mathematical expression is:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{P}}_{q}=\,{\mbox{PLSR}}\,\left({[P]}_{t},\,{[X]}_{t},\,{r}_{{x}_{t}},\,{[X]}_{t+1},\,{r}_{{x}_{t+1}},\cdots \,,\,{[X]}_{t+k},\,{r}_{{x}_{t+k}}\right)$$\end{document}P^q=PLSR[P]t,[X]t,rxt,[X]t+1,rxt+1,⋯,[X]t+k,rxt+k We term this mechanistically-informed approach the Luedeking–Piret Regression (LPR) model. Within the LPR framework, the PLSR component is trained by fitting a PLS model to the specific input features derived from the Luedeking-Piret equations e.g., the historical series of mAb concentration ([P]t), viable cell density ([X]t), and specific growth rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{{x}_{t}}$$\end{document}rxt) against the target mAb concentration at a future time point (Pq). Purely data-driven machine learning models approaches for bioprocess prediction operate as black boxes learning patterns directly from raw process variables without incorporating biological understanding. These models require extensive feature engineering and large datasets to identify relevant relationships between potential input variables and the target output. In contrast, the LPR approach (Eqs. 2–16) leverages the Luedeking-Piret equation to pre-define the most biologically relevant features: product titer history, cell density, and growth rates. This mechanistic foundation reduces the feature space from potentially many variables to just three key parameters. Fig. 1 illustrates the LPR model’s architectures compared with the purely data driven PLSR and Machine Learning methods.Fig. 1Comparison of modeling approaches for monoclonal antibody (mAb) production prediction.a The Luedeking–Piret equation provides a mechanistic foundation relating product formation to cell growth and density. b The Partial Least Squares Regression (PLSR) approach integrates mechanistic insights with PLSR for parameter estimation using biologically relevant features. c Purely data-driven machine learning approaches operate as black boxes requiring extensive feature engineering and large datasets. Comparison of modeling approaches for monoclonal antibody (mAb) production prediction. a The Luedeking–Piret equation provides a mechanistic foundation relating product formation to cell growth and density. b The Partial Least Squares Regression (PLSR) approach integrates mechanistic insights with PLSR for parameter estimation using biologically relevant features. c Purely data-driven machine learning approaches operate as black boxes requiring extensive feature engineering and large datasets. To demonstrate that the recurrence relation (Eq. 5) can be extended to k days ahead, the LPR model performs OSAF using the initial features from day t − 1 and day t as input variables to predict the product titer on day t + 1 (Eq. 7). For instance, to predict mAb production on day 9, the model considers only the data from days 7 and 8. In scenarios where data for certain days are missing, such as the absence of mAb data for day 12, the model adjusts accordingly. For example, to predict mAb production on day 13, the LPR model utilizes data from days 10 and 11, and for day 14, it uses data from days 11 and 13. Following the temperature shift on day 5, the predicted model uses only the feature from day 5 to predict mAb production on day 6. For MSAF, various combinations of sequence feature days were explored to determine the most accurate predictors of end-of-cycle mAb production. Feature day combinations were constructed by selecting an initial feature day 5, 6, 7 and pairing it with up-to next subsequent 2 days (Equations 10, 12, and 14) to predict mAb yields on day 16. For initial day t = 5 include day [5], day [5, 6], and day [5, 6, 7]; for initial day t = 6 include day [6], day [6, 7], and day [6, 7, 8]; and for initial t = 7 include day [7], day [7, 8], and day [7, 8, 9], a total of 9 predefined features day combinations. The initial day t features encompass mAb concentration, VCD, and the rate of cell growth at time t, while the selected subsequent day corresponds to specific variables such as VCD and the rate of cell growth at the respective time points. For instance, the combination [t5, t6, t7] included mAb concentration from day 5, and VCD and the rate of cell growth from days 5, 6, and 7. To predict the mAb concentration on day 16 using the LPR model as follows:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{P}}_{16}=\,{\mbox{LPR}}\,\left({[P]}_{5},\,{[X]}_{5},\,{r}_{{x}_{5}},\,{[X]}_{6},\,{r}_{{x}_{6}},\,{[X]}_{7},\,{r}_{{x}_{7}}\right)$$\end{document}P^16=LPR[P]5,[X]5,rx5,[X]6,rx6,[X]7,rx7The maximum predicted value or the mean predicted value of test clone from the nine predefined feature day combinations was utilized to assess the accuracy of the model’s performance. Equation 18 provides a formulation that enables the prediction of antibody concentration, Pq, at any future time point q by aggregating contributions from previous time steps up to t + k. The nine predefined feature day combinations were utilized to predict mAb concentrations on days 10, 11, 13, 14, and 15. Model development The data were organized in a three-dimensional matrix, where the rows corresponded to individual experimental runs and specific time points at which measurements were taken, and the columns to different process variables. For this study, this three-dimensional matrix was unfolded25 into a series of two-dimensional matrices, with each row representing an individual experimental run and each column representing individual process variables at each time point. In this study, the temperature shift initiated on day 5 from a temperature for cell growth to a lower temperature for more protein production resulted in an increase in productivity. A predictive model using various machine learning methods was developed to estimate the product titer at a target day of t + 1, using the historical process information available from day 5 until day t. A leave-one-out method was employed to test the model’s accuracy. Each clone type was marked as the unseen clone in turn and iteratively tested across all clone types. The model’s performance was assessed by calculating the absolute and percentage errors between the predicted and actual values for the validation set. The percentage error and mean percentage error (MPE) were calculated to assess the relative accuracy of the prediction compared to the actual value. These are defined as:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mbox{Percentage Error}}}_{i}=\frac{| {\hat{y}}_{i}-{y}_{i}| }{{y}_{i}}\times 100 \%$$\end{document}Percentage Errori=∣y^i−yi∣yi×100%20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mbox{MPE}}}_{i}=\frac{1}{n}\frac{| {\hat{y}}_{i}-{y}_{i}| }{{y}_{i}}\times 100 \%$$\end{document}MPEi=1n∣y^i−yi∣yi×100% Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{y}}_{i}$$\end{document}y^i is the predicted value for observation i, and yi represents the actual value for observation i. To simulate the real time clone selection process, a leave-two-out method was used to evaluate the model’s accuracy for clone selection. In each trial, two of the clones were held out and the remaining 5 clones were used to build an LPR model. For each such training set of 5 clones, the LPR model was constructed using nine predefined mechanistically-derived input features combinations of early-stage product titer, VCD, and growth rates. The LPR model was determined the optimal number of PLS components using a 5-fold cross-validation, all performed solely on that 5-clone training set. The resulting optimized LPR model was then used to predict the final mAb titer (day 16) of the two held-out clones. In each leave-two-out trial, the trained LPR model predicted final mAb titers for both held-out clones. Clone ranking was determined by comparing these predicted values. This ranking was then compared against the actual ranking based on measured final titers. A correct rank indicates that the model successfully identified which of the two clones achieved higher actual production. The clone selection accuracy was calculated as the percentage of correct clone selections out of all 21 trials. Model comparison for predicting mAb production The LPR OSAF model strategy was compared with several machine learning models, including random forest, gradient boosting (GB), support vector regression (SVR), and partial least squares (PLS), to predict next-day mAb production. Two feature input strategies were employed: a consistent input features approach and an expanded features approach that incorporated additional processing and ratio variables. The consistent input features included all mAb concentration, VCD, and the cell growth rate. The expanded features included all key process variables and ratios defined in Supplementary Table S1. Optuna was systematically employed to perform hyperparameter optimization for the Random Forest, GB, and SVR models. For the Random Forest model, the optimization targeted the number of estimators within the range of 50 to 200, the maximum tree depth between 2 and 20, and the minimum number of samples required to split an internal node from 2 to 10. For the GB model, Optuna fine-tuned the number of estimators from 50 to 200, the learning rate using a log-uniform distribution between 0.01 and 0.2, and the maximum depth of the individual regression estimators ranging from 2 to 7. For the SVR model, the hyperparameters optimized included the regularization parameter (C) spanning from 0.1 to 10 and the epsilon parameter between 0.1 and 0.5, both sampled using a log-uniform distribution. Mutual information regression was applied to rank the features based on their dependency with the target variable, and the top 15 features were used to build the PLS model for predicting the next time step (t + 1). Five specific feature day combinations included single day [5], and cumulative days [5, 6], [5, 6, 7], [5, 6, 7, 8], and [5, 6, 7, 8, 9] were used for the best model of OSAF to predict day 16 mAb to compare with LPR MSAF. Statistics and reproducibility Seven distinct Chinese Hamster Ovary (CHO) cell clones (clone 1 through clone 6 and bulk pool) were evaluated. Each clone was cultured in duplicate bioreactors (n = 2 per clone), resulting in 14 independent fed-batch cultures over a 16-day period. Replicates were defined as independent biological replicates consisting of separate bioreactor runs initiated from the same clone but cultured independently under identical feeding and environmental conditions. All process variables were measured at daily points throughout the culture period. Statistical analyses included leave-one-out and leave-two-out cross-validation for model evaluation, one-sided paired t-test for comparing model performance (n = 7, α = 0.05), and permutation testing (1000 iterations) for validation robustness. Prediction accuracy was quantified using mean percentage error (MPE) between predicted and actual values. No data points were excluded from the analysis. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Results Cell culture performance Viable cell density (VCD) (Fig. 2a) and the percentage of viable cells (%VC) (Fig. 2b) are parameters for assessing the health and proliferation of cell cultures. Throughout the culture period, VCD exhibited a characteristic growth curve, encompassing distinct lag, exponential, stationary, and death phases. Initially, all clones entered a lag phase, marked by a gradual increase in VCD as the cells adapted to the new environment.Fig. 2Daily key performance parameters of Chinese Hamster Ovary (CHO) clones over 16 days.Colored circles represent individual daily measurements; dashed lines connect measurements from the same clone and replicate run. Each color represents a different clone (n = 7 clones; n = 2 biological replicates per clone). a Viable cell density (VCD, cells mL−1) over time, showing initial exponential growth phase followed by stationary phase during cultivation. b Percentage of viable cells (%VC), indicating cell culture health and viability throughout the cultivation period. c Glucose to Lactate ratio, capturing the metabolic shift from oxidative phosphorylation to glycolytic activity. d Glutamine to Glutamate ratio (Gln/Glu), reflecting amino acid metabolism and nitrogen balance during cell growth. e Product Titer (mg L−1), illustrating monoclonal antibody (mAb) concentration accumulation over time. f Productivity (mg cells−1), measuring the efficiency of mAb production per cell. Temperature shift from the high-growth to the high-production phase occurred on day 5 for all cultures. Daily key performance parameters of Chinese Hamster Ovary (CHO) clones over 16 days. Colored circles represent individual daily measurements; dashed lines connect measurements from the same clone and replicate run. Each color represents a different clone (n = 7 clones; n = 2 biological replicates per clone). a Viable cell density (VCD, cells mL−1) over time, showing initial exponential growth phase followed by stationary phase during cultivation. b Percentage of viable cells (%VC), indicating cell culture health and viability throughout the cultivation period. c Glucose to Lactate ratio, capturing the metabolic shift from oxidative phosphorylation to glycolytic activity. d Glutamine to Glutamate ratio (Gln/Glu), reflecting amino acid metabolism and nitrogen balance during cell growth. e Product Titer (mg L−1), illustrating monoclonal antibody (mAb) concentration accumulation over time. f Productivity (mg cells−1), measuring the efficiency of mAb production per cell. Temperature shift from the high-growth to the high-production phase occurred on day 5 for all cultures. The glucose to lactate ratio (Fig. 2c) is high in the initial stage, indicating reliance on oxidative phosphorylation and an aerobic metabolic state. Concurrently, the glutamine to glutamate ratio (Gln/Glu) (Fig. 2d) suggested that cell growth was stabilizing on day 3. Starting on day 4, minor fluctuations were observed, characterized by a slight shift toward increased glycolytic activity and consistent nitrogen metabolism. These metabolic ratios serve as indicators of a cell’s energy balance and biosynthetic capacity (Supplementary Table S1). A rapid exponential growth phase was observed, with clone 4 achieving the maximum VCD value of 23.3 × 106 cells mL−1 on day 7. Similarly, clone 1 and bulk pool reached peak VCD values of 21.2 × 106 cells mL−1 on day 7 and 20.6 × 106 cells mL−1 on day 8, respectively. The VCD trends revealed some fluctuations but generally followed a pattern of growth followed by a decline due to the temperature shift. In the early growth stage, %VC remained high, indicating favorable conditions for cell viability. Notably, clone 1, clone 3, and clone 6 exhibited %VC values approaching 100%, reflecting high viability. As cultures progressed, a gradual decline in %VC was observed. This reduction in %VC coincided with the plateau and subsequent decline in VCD for some clones, such as clone 3 and clone 5. Throughout the culture period, the production titer (Fig. 2e) exhibited an initial increase, reaching a peak around the mid-point of the culture duration, followed by a plateau or decline in some clones. Notably, clone 1 and clone 6 achieved the highest production titers, demonstrating higher productivity compared to other clones. Productivity (Fig. 2f) began to increase starting on day 5 because of temperature shift initiated on day 5. These findings underscore the importance of selecting high-producing clones like clone 1 for large-scale production. Optimizing culture conditions can further enhance productivity and sustainability. The variations in product titer among clones highlight the need for thorough screening during bioprocess development. One-step-ahead forecasting (OSAF) of mAb yields The OSAF models were constructed for target days from days 6 to 16 and their performances were compared using the mean percentage errors (MPE) on these target days (Fig. 3). The PLS and LPR models generated more accurate predictions than the other models on the majority of the target days. In particular, the LPR model outperformed all the other models for days 15 and 16. Conversely, the GB model exhibited greater variability in its prediction accuracy among different target days. The SVR model yielded consistently higher MPE, particularly in the early target days. Overall, the PLS and LPR model offered substantially higher prediction accuracy than others for the last two target days.Fig. 3Comparison of one-step-ahead forecasting (OSAF) models using product titer, viable cell density (VCD), and growth rate as input.Grouped bars represent mean percentage error (MPE) calculated across n = 7 Chinese Hamster Ovary (CHO) clones (n = 2 biological replicates per clone) for each model at different target days (days 6–16). Each model is represented by a distinct color and hatch pattern: gradient boosting (blue bars with diagonal hatching), random forest (orange bars with reverse diagonal hatching), support vector regression (SVR, green bars with vertical hatching), partial least squares (PLS, red bars with horizontal hatching), and Luedeking-Piret Regression (LPR, purple bars with cross hatching). Models were trained to predict monoclonal antibody (mAb) levels using daily measurements from preceding days. LPR achieved the lowest overall MPE of 5.29% across all target days. Comparison of one-step-ahead forecasting (OSAF) models using product titer, viable cell density (VCD), and growth rate as input. Grouped bars represent mean percentage error (MPE) calculated across n = 7 Chinese Hamster Ovary (CHO) clones (n = 2 biological replicates per clone) for each model at different target days (days 6–16). Each model is represented by a distinct color and hatch pattern: gradient boosting (blue bars with diagonal hatching), random forest (orange bars with reverse diagonal hatching), support vector regression (SVR, green bars with vertical hatching), partial least squares (PLS, red bars with horizontal hatching), and Luedeking-Piret Regression (LPR, purple bars with cross hatching). Models were trained to predict monoclonal antibody (mAb) levels using daily measurements from preceding days. LPR achieved the lowest overall MPE of 5.29% across all target days. We hypothesized that the machine learning models could benefit from using more key process variables and ratios as input features. To test this hypothesis, all the key process variables and ratios defined in Supplementary Table S1 were used to construct OSAF machine learning models (i.e., GB, Random Forest, SVR, and PLS). The accuracy of the production predictions from these machine learning models on the target days was compared with that of LPR in Fig. 4. The PLS model, which incorporated a broader range of input features, demonstrated higher predictive accuracy relative to the other machine learning models. The inclusion of additional variables, such as processing and ratio variables, resulted in improved performance for PLS, as evidenced by its consistently low average mean percentage error of 2.92% across all target days. The LPR model achieved the second lowest, an average percentage error of 5.29%. The other machine learning models failed to benefit from considering a large number of key process variables. The performance boost of the PLS model may have stemmed from the feature selection capability of PLS. Additionally, deep learning approaches using long short-term memory (LSTM) neural networks were evaluated with both the three mechanistically-derived variables and all available process variables (Supplementary Table S2). However, LSTM models exhibited higher prediction errors compared to LPR and PLS, likely due to the limited training data available for complex neural network architectures.Fig. 4Performance of machine learning models with an expanded feature set.Grouped bars represent mean percentage error (MPE) calculated across n = 7 Chinese Hamster Ovary (CHO) clones (n = 2 biological replicates per clone) for each model at different target days (days 6–16). Each model is represented by a distinct color and hatch pattern: gradient boosting-ALL (blue bars with diagonal hatching), random forest-ALL (orange bars with reverse diagonal hatching), support vector regression-ALL (SVR-ALL, green bars with vertical hatching), partial least squares-ALL (PLS-ALL, red bars with horizontal hatching), and Luedeking-Piret Regression (LPR, purple bars with cross hatching). The -ALL suffix indicates models trained with all key process variables and metabolic ratios, in contrast to Fig. 3 models that used only three input features (product titer, viable cell density (VCD), and growth rate). LPR uses mechanistically-derived features regardless of input strategy. Incorporation of additional features improved PLS performance, achieving an MPE of 2.92% across all target days. Performance of machine learning models with an expanded feature set. Grouped bars represent mean percentage error (MPE) calculated across n = 7 Chinese Hamster Ovary (CHO) clones (n = 2 biological replicates per clone) for each model at different target days (days 6–16). Each model is represented by a distinct color and hatch pattern: gradient boosting-ALL (blue bars with diagonal hatching), random forest-ALL (orange bars with reverse diagonal hatching), support vector regression-ALL (SVR-ALL, green bars with vertical hatching), partial least squares-ALL (PLS-ALL, red bars with horizontal hatching), and Luedeking-Piret Regression (LPR, purple bars with cross hatching). The -ALL suffix indicates models trained with all key process variables and metabolic ratios, in contrast to Fig. 3 models that used only three input features (product titer, viable cell density (VCD), and growth rate). LPR uses mechanistically-derived features regardless of input strategy. Incorporation of additional features improved PLS performance, achieving an MPE of 2.92% across all target days. Comparative analysis of LPR and PLS models for multi-step ahead forecasting of mAb yields Table 1 presented a comprehensive analysis of predictive performance using leave-one-out cross validation. Overall, the LPR-based predictions demonstrated relatively lower MPE values than the PLS-based predictions. The LPR Mean predictions resulted in an average MPE of approximately 6.86%, and the LPR Max predictions yielded an average MPE of 5.85%. In contrast, the PLS Mean and PLS Max predictions produced higher average MPE values of 11.47% and 14.88%, respectively. When comparing performance across individual clones, some clones (clones 2 for instance) showed relatively small deviations between actual and predicted titers for cross all models. Conversely, predictions for clone 3 was less accurate. A one-sided paired t-test was conducted to compare the MPE of two modeling approaches (LPR Max and PLS Mean, lower MPE of each model). LPR Max MPE is significantly lower than PLS Mean MPE (t(6) = −2.01, p < 0.05). This result indicates that the LPR model outperformed the PLS model in predicting mAb production titers. We further validated the model’s statistical robustness using a permutation test (n = 1000). The test confirmed that our model’s actual mean percentage error of 5.85% is statistically significant (p = 0.001) when compared against models trained on randomly shuffled data, which had a mean error of 11.53%.Table 1Comparison of LPR and PLS models for mAb yield forecasting: mean and maximum predicted values with percentage errors across clonesCloneActual titerLPRPLSMeanMaxMeanMaxPred.aMPEPred.MPEPred.MPEPred.MPEClone 14.974.4711.73%5.184.22%4.4612.11%5.194.32%Clone 24.354.057.17%4.360.27%3.988.74%4.370.35%Clone 33.764.2814.11%4.4117.42%4.1711.09%4.2914.24%Clone 44.204.203.75%4.322.95%4.216.10%4.548.07%Clone 54.354.382.75%4.657.13%3.7713.31%3.8511.49%Clone 64.724.584.98%4.805.03%5.6622.96%7.1551.67%Bulk pool4.124.053.53%4.283.90%4.105.96%4.6913.99%Average6.86%5.85%11.47%14.88%aPred. = Predicted titer. Table 1Comparison of LPR and PLS models for mAb yield forecasting: mean and maximum predicted values with percentage errors across clonesCloneActual titerLPRPLSMeanMaxMeanMaxPred.aMPEPred.MPEPred.MPEPred.MPEClone 14.974.4711.73%5.184.22%4.4612.11%5.194.32%Clone 24.354.057.17%4.360.27%3.988.74%4.370.35%Clone 33.764.2814.11%4.4117.42%4.1711.09%4.2914.24%Clone 44.204.203.75%4.322.95%4.216.10%4.548.07%Clone 54.354.382.75%4.657.13%3.7713.31%3.8511.49%Clone 64.724.584.98%4.805.03%5.6622.96%7.1551.67%Bulk pool4.124.053.53%4.283.90%4.105.96%4.6913.99%Average6.86%5.85%11.47%14.88%aPred. = Predicted titer. Comparison of LPR and PLS models for mAb yield forecasting: mean and maximum predicted values with percentage errors across clones aPred. = Predicted titer. A leave-two-out cross-validation approach was used to evaluate the performance of the LPR model using the maximum predicted value metric (Table 2). This analysis encompassed 21 distinct held-out sets, derived from all possible pairwise combinations of seven unique cell line clones (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\begin{array}{c}7\\ 2\end{array}\right)$$\end{document}72). For each validation pair, both the actual and predicted production titers of the two clones were calculated. The predicted production titers for the same clone varied between different validation pairs because they were outputs from distinct models that were trained on different sets of clones. The correct rank column indicates whether the model accurately ranked the clones based on their predicted versus actual titers, with a value of TRUE signifying correct alignment. The mean prediction error (MPE) was calculated to quantify the discrepancy between predicted and actual values for each clone, serving as a measure of prediction accuracy. This leave-two-out cross-validation serves as a direct assessment of the model’s robustness against the inherent biological variability among clones. The LPR model successfully ranked 16 of the 21 clone pairs correctly, achieving an accuracy of approximately 76.2%. To contextualize the LPR model’s performance, we benchmarked it against a naive baseline. This simple model predicts the final ranking by selecting whichever clone has the higher production titer on Day 9. In the same leave-two-out validation, this naive approach achieved 66.7% accuracy (14 of 21 correct trials, Supplementary Table S3), confirming that the LPR model’s 76.2% accuracy offers a more robust predictive capability for clone selection. Incorrect rankings occurred in two instances where similar actual titers made rankings susceptible to minor prediction errors. Three cases involved overestimations or underestimations in the predicted values, which led to reversed rankings. These incorrect rankings can be attributed to out-of-distribution generalization errors in a small, heterogeneous dataset. The held-out clones in these cases may have distinct production profiles that were poorly represented by those in the limited training set, causing the model to extrapolate inaccurately.Table 2Leave-two-out clone selection results with LPRHeld-out clonesActual valuesPredicted valuesMPE (%)Correct rank(“Clone 1”, “Clone 2”)(4.975, 4.350)(5.091, 4.379)2.330%TRUE(“Clone 1”, “Clone 3”)(4.975, 3.760)(4.774, 4.365)6.369%TRUE(“Clone 1”, “Clone 4”)(4.975, 4.200)(4.642, 3.977)8.250%TRUE(“Clone 1”, “Clone 5”)(4.975, 4.345)(4.590, 4.464)8.907%TRUE(“Clone 1”, “Clone 6”)(4.975, 4.720)(4.466, 4.740)5.170%FALSE(“Clone 2”, “Clone 3”)(4.350, 3.760)(4.513, 4.392)4.444%TRUE(“Clone 2”, “Clone 4”)(4.350, 4.200)(4.474, 4.055)4.497%TRUE(“Clone 2”, “Clone 5”)(4.350, 4.345)(4.438, 4.527)4.195%FALSE(“Clone 2”, “Clone 6”)(4.350, 4.720)(4.420, 4.727)5.288%TRUE(“Clone 3”, “Clone 4”)(3.760, 4.200)(4.421, 4.205)17.745%FALSE(“Clone 3”, “Clone 5”)(3.760, 4.345)(4.422, 4.622)6.384%TRUE(“Clone 3”, “Clone 6”)(3.760, 4.720)(4.409, 4.695)4.707%TRUE(“Clone 4”, “Clone 5”)(4.200, 4.345)(4.246, 4.627)6.508%TRUE(“Clone 4”, “Clone 6”)(4.200, 4.720)(4.305, 4.829)6.269%TRUE(“Clone 5“, “Clone 6”)(4.345, 4.720)(4.608, 4.858)6.198%TRUE(“Bulk pool”, “Clone 1”)(4.115, 4.975)(4.311, 4.479)11.762%TRUE(“Bulk pool”, “Clone 2”)(4.115, 4.350)(4.251, 4.387)2.873%TRUE(“Bulk pool”, “Clone 3”)(4.115, 3.760)(4.279, 4.413)17.579%FALSE(“Bulk pool”, “Clone 4”)(4.115, 4.200)(4.379, 4.365)6.410%FALSE(“Bulk pool”, “Clone 5”)(4.115, 4.345)(4.358, 4.572)5.402%TRUE(“Bulk pool”, “Clone 6”)(4.115, 4.720)(4.329, 4.858)6.142%TRUE Table 2Leave-two-out clone selection results with LPRHeld-out clonesActual valuesPredicted valuesMPE (%)Correct rank(“Clone 1”, “Clone 2”)(4.975, 4.350)(5.091, 4.379)2.330%TRUE(“Clone 1”, “Clone 3”)(4.975, 3.760)(4.774, 4.365)6.369%TRUE(“Clone 1”, “Clone 4”)(4.975, 4.200)(4.642, 3.977)8.250%TRUE(“Clone 1”, “Clone 5”)(4.975, 4.345)(4.590, 4.464)8.907%TRUE(“Clone 1”, “Clone 6”)(4.975, 4.720)(4.466, 4.740)5.170%FALSE(“Clone 2”, “Clone 3”)(4.350, 3.760)(4.513, 4.392)4.444%TRUE(“Clone 2”, “Clone 4”)(4.350, 4.200)(4.474, 4.055)4.497%TRUE(“Clone 2”, “Clone 5”)(4.350, 4.345)(4.438, 4.527)4.195%FALSE(“Clone 2”, “Clone 6”)(4.350, 4.720)(4.420, 4.727)5.288%TRUE(“Clone 3”, “Clone 4”)(3.760, 4.200)(4.421, 4.205)17.745%FALSE(“Clone 3”, “Clone 5”)(3.760, 4.345)(4.422, 4.622)6.384%TRUE(“Clone 3”, “Clone 6”)(3.760, 4.720)(4.409, 4.695)4.707%TRUE(“Clone 4”, “Clone 5”)(4.200, 4.345)(4.246, 4.627)6.508%TRUE(“Clone 4”, “Clone 6”)(4.200, 4.720)(4.305, 4.829)6.269%TRUE(“Clone 5“, “Clone 6”)(4.345, 4.720)(4.608, 4.858)6.198%TRUE(“Bulk pool”, “Clone 1”)(4.115, 4.975)(4.311, 4.479)11.762%TRUE(“Bulk pool”, “Clone 2”)(4.115, 4.350)(4.251, 4.387)2.873%TRUE(“Bulk pool”, “Clone 3”)(4.115, 3.760)(4.279, 4.413)17.579%FALSE(“Bulk pool”, “Clone 4”)(4.115, 4.200)(4.379, 4.365)6.410%FALSE(“Bulk pool”, “Clone 5”)(4.115, 4.345)(4.358, 4.572)5.402%TRUE(“Bulk pool”, “Clone 6”)(4.115, 4.720)(4.329, 4.858)6.142%TRUE Leave-two-out clone selection results with LPR The LPR model was further extended to predict mAb production on days 10, 11, 13, 14, and 15 using the same nine predefined feature day combinations. Fig. 5 displays the predictive performance of the model evaluated across all clones for the nine predefined featured day combinations, including both short-term (days 10–11) and long-term (days 13–16) prediction values. Supplementary Fig. S2 displays mean and max prediction accuracy. Overall, the mean prediction accuracy outperformed the max prediction accuracy, except for the day 16 prediction. The overall error margins difference range from 0.29 to 2.33%, indicating relatively consistent performance across the two different methods. Predictions for days 10 and 11 of clone 1–4 were accurate, with consistently low error percentages for both mean predicted value and max predicted value, however, clones 5, 6 and bulk pool exhibiting higher prediction errors. Bulk Pool and Clone 4 exhibited the most consistent predictions across most of target days, maintaining error percentages below 5% and accurately capturing production trajectories without anomalies. Clone 1 demonstrated accuracy for early prediction days; however, anomalies emerged on days 14 and 15, where the max mean predicted error percentage exceeded 8%. Clone 2 showed stable performance with moderate error percentages across all target days, although a slight increase in errors was observed on day 14. In contrast, clone 3 experienced a notable decline in prediction accuracy on days 14–16, with max mean predicted error percentages surpassing 8%. Clones 5 and 6 presented moderately accurate predictions with minor declines in accuracy for later target days, but no notable anomalies were detected. Overall, the model demonstrated short-term prediction capabilities but encountered limitations in day 14 and day 15 accuracy, particularly for clones 1 and 3. Notably, bulk pool and clone 4 outperformed other clones.Fig. 5Multi-step-ahead forecasting (MSAF) predictions vs. actual product titers across seven Chinese Hamster Ovary (CHO) clones.a–g show prediction performance for Clone 1, Clone 2, Clone 3, Clone 4, Clone 5, Clone 6, and Bulk Pool, respectively. In each: gray shaded area represents the predicted range (minimum to maximum) from nine feature day combinations; gray circular markers show individual predicted values from each feature day combination; black solid line with circular markers indicates the mean predicted value across all combinations; red dashed line with circular markers shows the actual measured product titer values. Predictions target days 10, 11, 13, 14, 15, and 16 using input data from days 5–9. The Luedeking-Piret regression (LPR) model generates predictions using product titer, viable cell density (VCD), and growth rate from early culture stages (n = 2 replicates per clone). Multi-step-ahead forecasting (MSAF) predictions vs. actual product titers across seven Chinese Hamster Ovary (CHO) clones. a–g show prediction performance for Clone 1, Clone 2, Clone 3, Clone 4, Clone 5, Clone 6, and Bulk Pool, respectively. In each: gray shaded area represents the predicted range (minimum to maximum) from nine feature day combinations; gray circular markers show individual predicted values from each feature day combination; black solid line with circular markers indicates the mean predicted value across all combinations; red dashed line with circular markers shows the actual measured product titer values. Predictions target days 10, 11, 13, 14, 15, and 16 using input data from days 5–9. The Luedeking-Piret regression (LPR) model generates predictions using product titer, viable cell density (VCD), and growth rate from early culture stages (n = 2 replicates per clone). Discussion Developing a biomanufacturing model entailed selecting salient key process variables as input features, defining a functional form with equations that mathematically relates these inputs to the output, and optimizing the parameters within these equations. Without mechanistic insights, machine learning approaches must infer all three components solely from data, necessitating extensive and diverse training datasets to achieve generalization. In this study, we introduced the Luedeking-Piret regression (LPR) model, a hybrid model that combined mechanistic understanding from the Luedeking-Piret equation with partial least squares (PLS) regression to predict monoclonal antibody (mAb) yields in Chinese Hamster Ovary (CHO) cell cultures. The Luedeking-Piret equation, initially formulated for microbial fermentation processes, describes product formation as comprising both growth-associated (linked to growth rate) and non-growth-associated (related to cell concentration) components. Building on these mechanistic principles, we adapted the approach to CHO cell cultivation. Our derivation defined LPR’s functional form that related the mAb titer at any future time point to the cumulative effects of earlier culture stages. Parameter estimation within the LPR model was conducted using PLS regression. Anchoring the model in mechanistic understanding reduced model complexity, obviated the need for data-driven feature selection, and maintained predictive accuracy with limited training data. This hybrid structure is particularly suited to address the challenge of inter-clone heterogeneity observed in the clone selection process, where different clones exhibit unique growth and production profiles. The LPR model demonstrated performance in clone selection by accurately predicting final mAb yields using early-stage data. Employing multi-step-ahead forecasting (MSAF), the model effectively identified high-performing clones, expediting the selection process. A challenge in developing any predictive bioprocess model is accounting for the inherent variability between different cell clones. Our results demonstrate that the LPR model’s performance is affected by these inter-clone differences. Through leave-two-out cross-validation, the LPR model achieved an accuracy of 76.2%, ranking clones in 16 out of 21 trials. While the MPE of 5.85% demonstrates accuracy, the metric for the industrial application of clone selection is the ability to correctly rank clone performance. Our model’s 76.2% accuracy in correctly ranking clone pairs provides a direct and meaningful measure of its practical utility, enabling more rapid and confident identification of high-performing clones. This simulated real-world clone selection demonstrates that the LPR framework can generalize across different clone types. This predictive capability offers advantages in biomanufacturing development by enabling early identification of high-performing clones. By focusing experimental efforts on the most promising clones, the LPR model can substantially reduce the number of required experimental days and associated costs, ultimately speeding up the biomanufacturing development timeline. Our study focused primarily on multi-step-ahead forecasting (MSAF), rather than on OSAF as in many previous studies13,14. MSAF provides a broader predictive horizon and offers comprehensive production trajectory insights. By extending the original Luedeking-Piret equation, we derived LPR’s functional form (equations 16–18) for forecasting arbitrary q-steps ahead. The LPR model achieved an average Mean Percentage Error of 5.85% when predicting end-of-cycle mAb yields on day 16 using early-stage data up to day 9. We applied this approach to multiple intermediate target days (10, 11, 13, 14, and 15) in addition to day 16. The model leveraged nine predefined feature day combinations to generate nine predictions at each intermediate point, creating a detailed, nearly continuous forecast of mAb production throughout the cell culture process. As illustrated in Fig. 5, these multiple predictions provided a band of the potential production trajectories for each clone. The spread of these predictions defined an empirical prediction envelope that captured the model’s sensitivity to input specification. In contrast to stochastic models, where uncertainty increases as forecasts extend further into the future due to the accumulation of unpredictable factors, the LPR model’s different feature day combinations lead to relatively steady uncertainty as shown in Supplementary Fig. S2. The LPR model was primarily developed for the mid-term stage of clone selection, specifically during the high protein expression phase. At this stage, candidate clones have typically undergone initial screening from a larger pool, resulting in a population with similar growth and production characteristics. In the present study, this model was applied to a limited set of clones operating under a single, consistent biomanufacturing setting. The LPR model is fundamentally derived from the Luedeking-Piret equation, with parameter estimation subsequently performed using partial least squares (PLS) regression, a technique extensively employed in biomanufacturing. Conceptually, the LPR model’s theoretical framework supports its applicability to a broader spectrum of biomanufacturing scenarios. Future work is needed to expand training data to include a greater number of clones and diverse experimental conditions, including varied feeding strategies and environmental parameters. Additionally, future work should focus on integrating omics-based features to provide deeper mechanistic insights. While the high cost of time-series omics analysis currently limits its practical application, integrating this data into the LPR framework as the technologies become more accessible will be a step in further enhancing the model’s predictive power. These advancements will further enhance the predictive accuracy and generalizability of the LPR model, ensuring its applicability across a wider range of biomanufacturing scenarios. By addressing these issues, the LPR model will become a tool in the optimization and acceleration of biomanufacturing processes. Conclusions The integration of mechanistic modeling with machine learning through the Luedeking-Piret Regression (LPR) model provides a framework for predicting monoclonal antibody yields in CHO cell cultures. This hybrid approach integrates mechanistic biological kinetics with multivariate regression to improve model robustness while mitigating common machine learning challenges such as parameter overfitting and high-dimensional feature selection preprocessing. The implementation of multi-step-ahead forecasting (MSAF), as opposed to one-step-ahead forecasting (OSAF), enables a broader predictive scope and streamlined clone selection protocols. Our application of the LPR model demonstrates its potential in biomanufacturing, offering opportunities to accelerate development timelines, optimize resource utilization, and enhance the production of high-quality biopharmaceuticals. For example, reducing a 14-16 day mid-assessment to 9 days is expected to yield an approximate 20% cost reduction in full-time equivalent and consumables. More importantly, this time saving allows for an increase in operational capacity from approximately 18 runs per year to 26 runs per year. As this modeling approach continues to evolve, future refinements and extensions will further expand its utility across diverse bioprocess optimization scenarios. Supplementary information Supplementary Information Description of Additional Supplementary Files Reporting Summary Supplementary Information Description of Additional Supplementary Files Reporting Summary Source data Source Data Source Data