Machine learning-driven optimization of culture conditions and media components to mitigate charge heterogeneity in monoclonal antibody production: current advances and future perspectives

ABSTRACT

Charge heterogeneity in monoclonal antibodies (mAbs), caused by post-translational modifications, remains a substantial obstacle to ensuring consistent, stable, and effective therapeutics. Conventional optimization techniques, such as one-factor-at-a-time and design of experiments, often fail to capture the complex, nonlinear interactions between culture parameters (e.g. pH, temperature, duration) and medium components (e.g. glucose, metal ions, amino acids). This review highlights machine learning (ML) as a powerful approach for modeling these relationships and forecasting charge variant profiles in CHO cell-based mAb process development. We summarize supervised learning and regression methods used to link process conditions with charge heterogeneity and present case studies showing ML’s role in reducing acidic and basic variants. We also discuss challenges related to data quality, model interpretability, scalability, and regulatory compliance. Finally, we propose a roadmap for adaptive, ML-driven optimization strategies for bioprocess development, aligned with Quality-by-Design principles.

Introduction

Monoclonal antibodies (mAbs) are invaluable biopharmaceuticals, being essential for treating complex conditions such as cancer, autoimmune disorders, and infectious disease.¹ MAbs stand out as exceptional candidates for precision therapies and large-scale production. Unlike small-molecule drugs, they offer increased specificity, fewer off-target effects, reduced risk of drug interactions, and superior therapeutic efficacy, making them a cornerstone of modern medicine.^2,3 Over half of recent biopharmaceutical approvals are attributed to mAbs.⁴ Their adoption has considerably advanced innovation and strengthened the global market, driven by engineering breakthroughs, rising demand for biosimilars, and the increasing prevalence of chronic diseases.^5,6 While mAbs have revolutionized modern medicine, maintaining their consistent quality remains a major challenge because of the complex manufacturing processes and product variations that can affect their safety and effectiveness.^7,8

One of the biggest challenges in maintaining consistent quality of mAbs is controlling charge variants, which are a critical quality attribute (CQA) that can greatly affect their stability, bioavailability, pharmacokinetics, efficacy, and safety.^9,10 Heterogeneity in charge variants mainly arises from post-translational modifications (PTMs), such as deamidation, glycosylation, and oxidation, forming both acidic and basic charge variants.¹¹ These charge variants can disrupt the structural integrity and biological activity of mAbs, influencing their tissue distribution, pharmacokinetics, and overall effectiveness as therapeutic agents.^10,12 Charge variation is among the most common sources of heterogeneity in recombinant mAbs, presenting substantial challenges for ensuring batch-to-batch consistency and streamlining downstream processing.^9,13 Given its critical impact on drug safety and efficacy, regulatory agencies like the Food and Drug Administration (FDA) and European Medicines Agency (EMA) have established strict guidelines to closely monitor and strategically control charge variation during process development and optimization of mAbs.^14,15

Several key factors influence the quality of recombinant proteins in biopharmaceutical manufacturing, including the choice of cell lines, the culture media, and the optimization of key process conditions.¹⁶ Among these factors, culture conditions and medium components are critical, as they directly impact cellular metabolism.¹⁷ Culture conditions, such as pH,¹⁸ temperature,¹⁹ ambient light,²⁰ and cell culture duration,²¹ are critical in driving non-enzymatic reactions that influence cellular metabolism and PTMs.¹⁶ Modifying these parameters can either amplify or diminish unwanted changes, directly affecting the distribution of charge variants and the overall quality of the final product.⁹ The composition of the culture medium, including feed components and specific chemical supplements, also plays a pivotal role in modulating CQAs, such as glycosylation patterns, aggregation, and charge heterogeneity.^22–24 Effective management of these factors throughout the bioreactor process is essential for minimizing the impact of accumulated byproducts and ensuring consistent product quality.^25,26

In recent years, several approaches have been used to optimize culture parameters. This includes traditional methods like one factor at a time (OFAT) and design of experiments (DOE), more advanced techniques such as multivariate data analysis (MVDA) and Systems Biology (SB), and cutting-edge tools like machine learning (ML).^1,27 Although traditional methods like OFAT and DOE have been essential for optimizing culture parameters, they often fail to address the complex interactions between culture parameters and cellular processes.²⁸ Advanced techniques such as MVDA and SB have provided invaluable insights into these complexities, but, they still encounter challenges, including limited scalability and the inability to directly optimize specific properties, such as charge variants.^29,30 Conversely, ML utilizes large datasets and advanced algorithms to uncover hidden patterns and predict the optimal conditions for protein production, even when the underlying mechanisms are not fully understood.¹ By analyzing factors such as pH, temperature, cell culture duration, and nutrient composition, ML provides important insights into the intricate connections between culture parameters and CQAs, including charge variants.^21,31,32 ML’s innovative capabilities can compensate for shortcomings of traditional methods and enable a new era of precision and efficiency in bioprocess optimization.

Here, we examine the crucial role of culture conditions and medium components in influencing charge heterogeneity, which is an important quality attribute in mAb bioprocess optimization. We investigate the causes and effects of charge variants and emphasize the transformative potential of ML in optimizing culture parameters and improving CQAs. Through case studies, we demonstrate how ML effectively links culture parameters to charge variants and outline a proposed workflow for ML-driven optimization in industrial settings. Finally, we discuss the challenges and opportunities of implementing ML in scalable, efficient, and high-quality biopharmaceutical processes, providing a roadmap for future innovations.

Charge heterogeneity in MAbs

Charge heterogeneity in mAbs refers to various molecular variants within the same product batch, exhibiting distinct net surface charges. These variants are typically categorized into three primary groups: main species, acidic species, and basic species.¹⁰ The main species is the dominant form of a mAb, typically appearing as the primary peak in chromatographic analysis.³³ Although regarded as the desired product, it often contains common modifications, such as converting N-terminal glutamine to pyroglutamate, removing C-terminal lysine from the heavy chain, and adding neutral oligosaccharides to the standard asparagine site within the CH2 domain.³⁴ Acidic variants, which possess a more negative charge than the main species, frequently arise from PTMs such as the deamidation of asparagine residues, widely regarded as the primary driver of these variants. Acidic species also stem from sialylation of glycans, glycation, or the presence of terminal acidic groups.^35–37 In contrast, basic variants have a higher positive charge than the main species due to incomplete removal of C-terminal lysine, along with modifications such as C-terminal amidation, formation of succinimide, N-terminal glutamate conversion, or other structural changes.³³ The proportions of these charge variants can change based on the production process, cell line, and storage conditions.⁹

The charge heterogeneity seen in mAbs results mainly from various PTMs, both enzymatic and non-enzymatic, that occur during biomanufacturing and storage, in addition to inherent contributions from the host cell line.^38–40 Acidic variants often arise when asparagine residues undergo deamidation, a process that converts them into aspartic acid or isoaspartic acid, adding negative charges to the molecule. This process is accelerated under elevated pH and temperature.^13,33 Other important factors that contribute to acidic species include the enzymatic addition of sialic acids to N-glycans and the non-enzymatic glycation of lysine residues, both of which alter the net charge by increasing negative charges or masking positive ones.^15,22 Additionally, reactive carbonyl species like methylglyoxal can modify arginine residues, which further decreases their basicity.⁴¹ In contrast, basic variants usually arise when C-terminal lysine residues are not fully removed or when N-terminal glutamine fails to cyclize into pyroglutamate, both of which result in the retention of positive charges.^40,42 Other modifications, such as succinimide formation, which temporarily neutralizes negative charges, and enzymatic C-terminal amidation, can also help increase base specificity.^15,38 Furthermore, oxidative changes to methionine and tryptophan residues, along with variations in disulfide bond patterns, can lead to conformational changes that subtly affect chromatographic behavior, even if they do not directly modify the molecule’s overall charge.^9,33,43 Cell line-specific factors, including glycosylation profiles, enzyme activity levels, and metabolic state, play a crucial role in influencing the distribution of charge variants.^16,31 A comprehensive overview of these mechanisms and their implications is presented in Table 1.

Table 1.

Characteristics and impact of charge variants in mAbs.

Feature	Acidic	Main	Basic	Ref.
Net Charge	More negative/Lower pI	Expected net charge/pI for the target molecule profile	More positive/Higher pI	Xie et al.44
Chromatographic Profile	Elutes earlier in Cation Exchange Chromatography (CEX)/Migrates towards anode in cIEF	A principal peak in CEX/cIEF; Represents the target elution/migration time	Elutes later in CEX/Migrates towards cathode in cIEF	Beck et al.9 Xie et al.44
Representative PTMs	Deamidation (Asn → Asp/isoAsp) - Sialylation (N-glycans) - Glycation - Trp Oxidation (sometimes) - Arg Modification (e.g., by methylglyoxal)	N-terminal Pyroglutamate Formation - Complete C-terminal Lys Removal - Core/Neutral N-glycosylation (e.g., G0F, G1F)	Incomplete C-term Lys Removal - Incomplete N-term PyroGlu Formation - Succinimide Formation - C-terminal Amidation - Asp Isomerization to isoAsp (sometimes) - Met Oxidation	Brorson and Jia38 Du et al.33 Gangwar et al.32
Key Origin Drivers	High Temperature; High pH (esp. for deamidation); Long culture duration; High glucose/sugar; Oxidative stress (DO, light, certain metals); Nutrient depletion; Cell stress/lysis	Optimal biosynthesis, Efficient enzymatic processing, Controlled process conditions	Suboptimal enzymatic processing (esp. carboxypeptidase, glutaminyl cyclase), Low pH (esp. isoAsp/succinimide), Specific metal ion levels (e.g., Zn/Cu affecting Cp), Cell line-specific enzyme levels	Gangwar et al.29 Ha et al.32 Sissolak et al.36 Weng et al.21
Analytical Detection Methods	CEX, cIEF/iCIEF, LC-MS, Peptide Mapping	CEX, cIEF/iCIEF, LC-MS, Peptide Mapping	CEX, cIEF/iCIEF, LC-MS, Peptide Mapping	Gupta et al.35 Horvath et al.45 Wagner-Rousset et al.15
Stability Profile	Often linked to degradation pathways (deamidation, oxidation); Potential for increased aggregation/fragmentation propensity	Represents target stability profile; Generally stable within specifications	May contain degradation intermediates (succinimide, isoAsp); Can impact conformation/solubility	Hazeltine et al.46 Kaur et al.26 Luo et al.42 Zheng et al.47
Functional Implications	Altered antigen binding (if in CDR), ↓ ADCC/CDC (sialylation), altered PK, immunogenicity (sometimes)	Represents the target activity profile; Benchmark for comparison.	Altered FcRn binding/PK (C-term Lys), potentially altered potency or clearance (conformational variants), immunogenicity (sometimes, e.g., leader seq.)	Chen et al.48 Hong et al.49 Zhang et al.50 Zheng et al.51
Regulatory Perspective	CQA; Must be characterized, monitored, and controlled within justified limits.	Defined as the Reference Standard profile; CQA; Must be characterized, monitored, and controlled within justified limits	CQA; Must be characterized, monitored, and controlled within justified limits	Horvath et al.45 Rouiller et al.52 Xu et al.53

Even at low levels, charge variants can affect the therapeutic effectiveness, stability, and regulatory compliance of mAbs.¹⁰ These variants typically arise from PTMs that can affect pharmacokinetics, pharmacodynamics, and immunogenicity, collectively influencing the clinical effectiveness of mAbs.^4,36 For example, acidic variants of antibodies may decrease their ability to bind to target molecules, reducing their therapeutic effectiveness.^13,39 Conversely, basic variants might change how antibodies interact with Fc receptors, potentially shortening their half-life or causing undesirable immune reactions.⁵⁴ Even at low levels, these charge variations can affect the treatment’s effectiveness and stability, making it more challenging to maintain consistency across batches and adding complexity to the development and regulatory approval processes.^22,54,55 Given these potential impacts, regulatory bodies such as the FDA and EMA recognize charge heterogeneity as a CQA that requires careful control.^15,56 As such, addressing these variations effectively is key to achieving reliable and scalable mAb production with consistent quality.⁹

Understanding the charge variability in mAbs relies on various advanced analytical techniques.⁴⁰ Capillary isoelectric focusing (cIEF) provides a high-resolution view of charge variants based on isoelectric points, making it essential for analyzing acidic, basic, and main species in mAb formulations.^40,54 Its accuracy makes it invaluable in both development and quality control. Ion exchange chromatography (IEX), especially cation exchange chromatography (CEX), refines the separation of different charge variants by capitalizing on subtle charge differences.³³ Through carefully tuned pH and salt gradients, it can tease apart distinct molecular species with remarkable precision.⁵⁴ In addition, liquid chromatography-mass spectrometry (LC – MS) delivers a high-resolution snapshot of PTMs that affect mAb charge profiles.³⁹ Complementary techniques like electrospray ionization (ESI) and detailed peptide mapping take it a step further, uncovering deeper molecular details for a more thorough and precise analysis.^9,39 Innovative techniques such as imaging capillary IEF (icIEF) and integrated chromatography-mass spectrometry methods are advancing charge variant analysis by providing enhanced sensitivity and resolution, ultimately leading to improved stability, safety, and efficacy of mAbs.^54,55

Culture condition and media component optimization for mAbs

Optimizing culture conditions and culture medium components is a primary strategy for reducing charge heterogeneity in mAb production, as it directly affects the PTMs that determine the charge profile. Culture conditions can be categorized into physical (e.g., temperature, pH), operational (e.g., culture duration), and environmental (e.g., ambient light, dissolved gases), all of which affect the cellular metabolic state and the kinetics of PTMs.^{18,20,31,37,44} Temperature can modulate charge heterogeneity; reducing the culture temperature (for instance, to 32–34 °C) can decrease metabolic activity and slow non-enzymatic reactions like deamidation, thus lowering the formation of acidic species.^37,47 However, this drop in temperature may also affect enzymes such as carboxypeptidases, which could lead to incomplete removal of C-terminal lysine and, as a result, an increase in basic variants.⁵⁰

pH, though easier to control, plays an equally important role. Higher pH levels typically accelerate deamidation and reduce sialylation, both leading to an increase in acidic species.³¹ On the other hand, lowering the pH can encourage sialylation and enhance glutamine cyclization at the N-terminus, which in turn shifts the charge profile toward more basic species.^18,44 Besides temperature and pH, factors like culture duration also influence charge heterogeneity. For instance, in GS-CHO cells, longer culture periods (e.g., from day 5 to day 14) result in the accumulation of PTMs, which in turn increase the proportion of acidic variants.¹⁸ Environmental factors such as ambient light and dissolved gases (pCO₂ and pO₂) also influence PTMs.^20,31 For instance, light exposure can lead to the oxidation of residues such as tryptophan, which increases acidic variants.²⁰ Higher levels of pCO₂ lead to the formation of acidic species, while variations in pO₂ levels affect the production of basic compounds.³¹ Table 2 provides a comprehensive summary of studies analyzing these parameters over various Chinese hamster ovary (CHO) cell lines.

Table 2.

Key culture parameters influencing charge Heterogeneity^a.

Parameter	Cell line	Rangeb	Effectc		Ref.
			Acidic	Basic
Temperature (ºC)	CHO-K1	31, 34, 37	Affected	Increases	Sissolak et al.36
	CHO-DG44	35, 36.5	Decreases	Affected	Weng et al.21
	CHO	29, 32, 35, 37	Increases	N/A	Tang et al.37
	CHO	33, 35, 37	Decreases	N/A	Kishishita et al.57
	CHO-K1 HP & LP	33 → 37	Decreases	Increases	Zhu et al.19
	CHO-K1	30, 32, 34	Increases	Decreases	Zheng et al.47
	CHO	33, 35, 37	N/A	Increases	Luo et al.42
	CHO	32, 35, 37	Decreases	Increases	Zhang et al.50
	CHO	31 → 35	Decreases	N/A	Yang et al.58
	CHO	34 → 36.5	Increases	N/A	Chung et al.25
	CHO-K1 GS	30, 32, 33.5, 35, 36.5	Decreases	Increases	McHugh et al.59
	rCHO	30, 33, 37	Decreases	Increases	Yoon et al.60
	GS CHO1, DG44 CHO2	32, 33, 34, 35, 36.5, 37.5, 38.5	Decreases	Increases	Xu et al.61
	CHO	36, 37, 38	Increases	N/A	Horvath et al.45
pH	CHO	6.8, 7, 7.2	Increases	Affected	Brunner et al.31
	CHO-S	7.05, 7.15	Increases	Decreases	Rouiller et al.52
	GS‐CHO	6.7, 6.9, 7.1	Increases	Decreases	Lee et al.18
	CHO	5.5 → 7	Affected	N/A	Trexler‐Schmidt et al.62
	GS-CHO	6.75±0.1 → 6.95±0.1	Decreases	N/A	Xie et al.44
	CHO	6.85±0.05 → 7.15±0.05	N/A	N/A	Zheng et al.51
	CHO	7 → 7.3	Increases	N/A	Horvath et al.45
Culture duration (days)d	CHO-DG44	11, 12, 13, 14, 15	Increases	N/A	Weng et al.21
	CHO	3, 6, 9, 12, 15	Increases	N/A	Tang et al.,37
	CHO	12 vs. 14	N/A	Decreases	Luo et al.42
	GS-CHO	5, 8, 10, 12, 14	Increases	N/A	Lee et al.18
	CHO-K1 GS	3, 8, 12	Increases	N/A	Park et al.63
	CHO-K1 GS	3, 6, 9	Increases	N/A	McHugh et al.59
Ambient Light (lux)	CHO	1000, 3000, 10000	Increases	Increases	Mallaney et al.20
Feed media age (weeks)e	CHO	3 → 10	Increases	N/A	Chung et al.25
Passage number	GS-CHO	PN-10 → PN-60	Decreases	Decreases	Kaur et al.26
	CHO DG44	P5 → P33	N/A	Increases	Xu et al.53
pCO2%)	CHO	5 → 20	Affected	Increases	Brunner et al.31
pO2%)	CHO	10 → 40	N/A	Increases	Brunner et al.31
Process mode	CHO	Traditional Fed-batch vs. Concentrated Fed-batch	Decreases	N/A	Yang et al.58
Seed density (PCV%)	CHO	0.12 → 0.36	Affected	N/A	Horvath et al.45

^aAll products are mAbs, and all media used are chemically defined.

^b“→” indicates a change from one value to another; “vs.” compares two values; commas separate multiple values for the same parameter.

^cEach parameter’s impact on charge variants is classified as Increase, Decrease, Affected (without a consistent trend), or N/A (no data or negligible impact reported).

^dAll the studies included in this table for culture duration employ a fed-batch cultivation method.

^eThis parameter primarily refers to the duration of storage for prepared liquid media, which can affect stability and component integrity. Dry powder media are generally unaffected until reconstitution.

The effects of culture parameters on mAb charge heterogeneity, as summarized in Table 2, often differ across studies. These variations are due to the inherent complexity of bioprocesses and context-specific factors such as the CHO cell line or clone, the structural features of the expressed antibody, and interactions among variables like pH, temperature, and nutrient composition.^10,21,36 Process configurations (e.g., fed-batch regimes) and media-related factors (e.g., age and formulation) also significantly influence the outcomes.^25,26 Rather than indicating contradictions, this variability highlights the biological diversity of upstream systems and emphasizes the importance of case-specific, data-driven process development strategies.^9,56

Along with the culture conditions, the composition of the culture medium and feed is key in shaping the charge variant profile of mAbs. Ingredients like glucose and trace metals in the media can either promote or hinder specific PTMs, influence metabolic pathways, and ultimately shape the overall charge variant distribution.^22,23,31,36 Amino acids are essential building blocks for protein synthesis and key modulators of PTMs.⁶⁴ For instance, adequate levels of asparagine act as a substrate for deamidation and promote acidic variants.⁵² In contrast, high levels of lysine and arginine can inhibit carboxypeptidases, restrict lysine cleavage at the C-terminal, and increase basic variants.²⁴ Similarly, trace metal ions such as zinc, copper, iron, and manganese are critical cofactors for enzymes involved in a range of PTMs, such as carboxypeptidase activity, amidation, and glycosylation, while also catalyzing oxidation reactions.^22,37,65 Most importantly, their equilibrium and speciation can affect enzyme activity and charge profile. For example, excess copper (e.g., 0.004–0.4 mM) can catalyze oxidative reactions, contributing to charge heterogeneity.^21,66 Moreover, reducing sugars, mainly glucose and galactose, at high concentrations (> 15 g/L) promote the formation of acidic species via non-enzymatic glycation and influence the distribution of charge species via sialylation.^36,52,67 Furthermore, certain vitamins, supplements, hydrolyzates, sodium butyrate, and accumulated metabolic byproducts can modulate cellular metabolism, redox state (as antioxidants or prooxidants), and PTM enzymatic activities, adding further layers of complexity to the optimization of the charge profile.^{25,29,44,49,68} Table 3 summarizes key findings from studies on the effects of various parameters on mAb charge variants in different CHO expression systems.

Table 3.

Key media components influencing charge Heterogeneity^a.

Components	Cell line	Rangeb	Effectsc		Ref.
			Acidic	Basic
Glucose (g/L)	CHO-K1	2 → 15	Increases	N/A	Sissolak et al.36
	CHO	11.5–23	Increases	N/A	Quan et al.67
Copper (mM)	CHO-DG44	0.004, 0.04, 0.15, 0.4	N/A	Increases	Weng et al.21
	CHO-GS	0 → 0.0315	Decreases	N/A	Gangwar et al.22
	CHO-DXB11	0, 0.00015, 0.00025, 0.0005, 0.001, 0.002, 0.003	Decreases	Increases	Yuk et al.43
	CHO	1X → 4X	N/A	Increases	Luo et al.42
	CHO-K1	0.0004, 0.00055, 0.00075, 0.001	Affected	Increases	Kaschak et al.40
	CHO-S	0x, 1x, 3x	Affected	N/A	Rouiller et al.52
	CHO-K1	0, 0.005, 0.05, 0.1	Decreases	N/A	Chaderjian et al.66
	CHOK1SV	N/A (increased)	Decreases	Decreases	Hazeltine et al.46
Iron (mM)	CHO-GS	0.0197, 0.3008, 0.3581, 1.6240	Increases	N/A	Gangwar et al.22
	CHO/dhFr	0.01, 0.018, 0.075	Increases	N/A	Vijayasankaran et al.68
	CHO-S	0, 0.1343, 0.2686	Increases	N/A	Rouiller et al.52
	CHO DG44	0.01, 0.02, 0.03, 0.05, 0.11	Increases	Affected	Xu et al.53
	CHO	Control Medium vs. Control Medium + 0.08 mM Iron	Decreases	N/A	Chung et al.25
	CHO-GS	0.1, 0.2, 0.4	Increases	Increases	Gangwar et al.32
Zinc (mM)	CHO-DG44	0, 0.004, 0.03, 0.04, 0.07, 0.08, 0.1, 0.13, 0.15	Decreases	N/A	Weng et al.21
	CHO-GS	0.0061, 0.0138, 0.1530, 0.0306	Increases	Decreases	Gangwar et al.22
	CHO	1X → 12X	N/A	Decreases	Luo et al.42
	CHO-S	0X, 1X, 3X	Affected	N/A	Rouiller et al.52
	CHO-GS	0.060, 0.090, 0.180	Decreases	Decreases	Gangwar et al.32
Manganese (mM)	CHO-DG44	0, 0.004, 0.040	N/A	N/A	Weng et al.21
	CHO-GS	0.00091, 0, 0.00182, 0.0273	Increases	N/A	Gangwar et al.22
	CHO-S	0, 0.0002, 0.001	Affected	Affected	Rouiller et al.52
	CHOK1SV	N/A (increased)	Decreases	Decreases	Hazeltine et al.46
Magnesium (mM)	CHO-GS	0.8023, 0.8147, 0.8393, 0.8517	N/A	N/A	Gangwar et al.22
Cobalt (mM)	CHO-GS	0, 0.0017, 0.00509	Increases	N/A	Gangwar et al.22
Nickel (mM)	CHO-GS	0, 0.017, 0.0341	Increases	Increases	Gangwar et al.22
Sodium butyrate (mM)	rCHO	0, 0.01, 0.05, 0.2, 1, 2, 4	Decreases	Increases	Hong et al.49
Selenium (mM)	CHO-S	0, 0.0002, 0.001	Affected	N/A	Rouiller et al.52
Biotin (mM)	CHO-GS	0, 0.0073, 0.0094	Decreases	Increases	Gangwar et al.22
Ascorbic acid	CHO	0, 0.568, 5.678, 17.034	Increases	N/A	(Chumsae et al.41
Surfactants (%v/v, %w/v)	CHO-DUX B11, CHO-K1	Type: PS80, PS20, P188	N/A	N/A	Hossler et al.69
		Concentration: 0.0025% − 10%	Decreases	Increases	Gangwar et al.22
Choline (mM)	CHO-GS	0.2414, 0.6615, 0.873, 0.912	N/A	Affected
Carboxypeptidase B (w/w)	CHO-GS, CHO-DG44	0, 1:100, 1:10000	Increases	Decreases	Xu et al.70
	CHO	0, 1:125	Affected	Decreases	Kim et al.71
Methylglyoxal (mM)	CHO	0 → 2.8	Increases	N/A	Chumsae et al.39
Folic acid (mM)	CHO-GS	0.0062, 0.008, 0.0332, 0.0419	N/A	N/A	Gangwar et al.22
	CHO/dhFr	0, 0.2232, 0.4463	Increases	N/A	Vijayasankaran et al.68
Dextran sulfate (g/L)	CHO-K1, DG44	MW: 4000, 15000, 40000	Affected	Decreases	Hyoung Park et al.72
		Concentration: 0.1, 0.3, 0.5, 1
Hydrolyzate (g/L)	CHO-K1	0 → 0.1	Decreases	Increases	Zheng et al.47
Taurine (mM)	CHO	0 → 13.2	N/A	Decreases	Liu et al.73
Pantothenic acid (mM)	CHO-GS	0.0084, 0.0138, 0.0214, 0.0559	N/A	Decreases	Gangwar et al.22
Arginine (mM)	CHO	2 → 10	N/A	Increases	Zhang et al.24
Rutine (g/L)	CHO	0.005 → 1	Decreases	Increases	Hossler et al.55
Rosmarinic acid (g/L)	CHO	0 → 0.36	Decreases	N/A	Chung et al.25
Lysine (mM)	CHO	2 → 10	N/A	Increases	Zhang et al.24
Epigallocatechin gallate (g/L)	CHO	0.1 → 1	Decreases	Increases	Hossler et al.55
Uridine (mM)	CHO	0 → 6	Decreases	Increases	Niu et al.74
Thiamine (mM)	CHO-GS	0.0129, 0.0134, 0.0159, 0.0273	N/A	Affected	Gangwar et al.22
Galactose (g/L)	CHO	0, 8, 19.50	Increases	N/A	Quan et al.67
	CHO-S	0, 1.802, 3.603	Affected	N/A	Rouiller et al.52
β-glycerol phosphate (g/L)	CHO DG44	0 → 0.9	N/A	Increases	Xu et al.70
Spermine (mM)	CHO-S	0, 0.075, 0.2	Increases	Affected	Rouiller et al.52
Asparagine (mM)	CHO-S	0, 10, 20	Decreases	Decreases	Rouiller et al.52
Pyridoxine (mM)	CHO/dhFr	0, 0.2, 0.4	Increases	N/A	Vijayasankaran et al.68
	CHO-GS	0, 0.02, 0.0211, 0.0287	N/A	N/A	Gangwar et al.22
Riboflavin (mM)	CHO/dhFr	0, 0.0133, 0.0266	Increases	N/A	Vijayasankaran et al.68
Cysteine (mM)	CHOK1SV	1×(historical media) to 1×/1.5×(CD media)	Decreases	Decreases	Hazeltine et al.46
Medium (%)	CHO	85, 100, 115	Decreases	N/A	Horvath et al.45
	CHO-DUX B11 and CHO-K1	1X CDFM, 2X CDFM (X)	Decreases	Increases	Hossler et al.69
	CHO	BM-1, BM-2، BM-3، BM-4	Affected	N/A	Tang et al.37
Cobalamin (mM)	CHO/dhFr	0, 0.0177, 0.0354	Increases	N/A	Vijayasankaran et al.68
Tryptophan (mM)	CHOK1SV	1×(historical media) to 0.7×/0.4×(CD media)	Decreases	Decreases	Hazeltine et al.46

^aAll products are mAbs, and all media used are chemically defined.

^b“→” indicates a change from one value to another; “vs.” compares two values; commas separate multiple values for the same parameter.

^cEach parameter’s impact on charge variants is classified as Increase, Decrease, Affected (without a consistent trend), or N/A (no data or negligible impact reported).

Historically, managing charge heterogeneity by optimizing culture conditions and media components has predominantly relied on experimental and statistical methods.^21,57 These conventional techniques, including the DoE, statistical modeling, OFAT, and process controls, have formed the backbone of optimization efforts to manage charge heterogeneity.^29,44 OFAT refers to an approach in which only one variable is changed while others are held constant, making it easy to implement but limited in detecting interactions. DoE, in contrast, enables the simultaneous investigation of multiple variables and their interactions using structured experimental designs.⁷⁵ Early efforts often used a simple OFAT approach, although its inability to assess complex parameter interactions posed limitations.⁷⁶ As revealed in studies such as Wang et al.²¹ and Horvath et al.,⁴⁵ DOE enables the simultaneous evaluation of multiple factors, such as temperature, pH, and zinc levels, to identify optimal conditions and interactions, but it is still limited by the number of testable variables and the complexity of nonlinear biological systems.^77,78

Furthermore, statistical modeling techniques such as MVDA and partial least squares are widely used to predict parameter effects and analyze complex relationships, but their reliance on large datasets and linear assumptions limits their accuracy in capturing nonlinear biological responses and PTM dynamics underlying charge heterogeneity.^{20,25,27,31,79} Nevertheless, studies such as Chung et al. have demonstrated the use of PLS-based MVDA to correlate oxidative stress markers with acidic charge variants, providing actionable insights for media optimization and feed strategy adjustment.²⁵ MVDA encompasses statistical methods such as principal component analysis (PCA) and partial least squares that handle multiple input variables simultaneously to uncover patterns and correlations, often serving as exploratory or dimensionality-reduction tools.⁸⁰ SB combines omics data with computational modeling to optimize culture media and better understand cellular behavior.¹ Despite its potential, the complexity, data demands, and inability to directly optimize specific traits, like charge heterogeneity, of SB have limited its application to charge variant optimization in mAb bioprocess development.^81,82 Consequently, while traditional process controls (e.g., online pH and DO monitoring) help stabilize culture conditions during development, they are insufficient for resolving complex nonlinear interactions driving charge heterogeneity, thus highlighting the need for advanced modeling approaches like ML.^16,29,56 Figure 1 offers a comparative overview of the capabilities of OFAT, DoE, MVDA, and ML across key dimensions.

Figure 1.

Comparative evaluation of OFAT, DoE, MVDA, and ML capabilities for bioprocess optimization. This radar chart compares four approaches, OFAT, DoE, MVDA, and ML, across six key dimensions relevant to CHO cell culture optimization for charge heterogeneity: process integration, predictive accuracy, result interpretability, industrial scalability (higher values preferred), and modeling complexity, data requirements (lower values preferred). Process integration refers to how well a method fits into the full development pipeline, including monitoring and control, while industrial scalability reflects its applicability from lab to production scale. ML (yellow) shows strong performance in accuracy and integration, though it demands more data and modeling complexity. SB offers deep biological insights but has limited direct use in charge variant prediction. The assessment is qualitative, based on published experimental and industrial studies.

Machine learning in bioprocessing

To address the challenges of optimizing mAb production in complex cell culture systems, where traditional methods often fail to capture nonlinear interactions and high-dimensional data, ML offers advanced tools for developing predictive models.^83,84 As a subset of artificial intelligence, ML focuses on building algorithms that learn patterns from data, enabling them to make predictions or decisions without being explicitly programmed.¹ These ML approaches, including supervised, unsupervised, and reinforcement learning, offer a robust toolkit for assaying complex bioprocessing data.^85,86 Among these, supervised learning has the greatest application in biopharmaceuticals through the use of labeled datasets, as it directly links measured culture inputs, such as temperature, pH, and nutrient concentrations, to quantitative outputs such as titer or ratio of load types.^83,84,87 Supervised learning algorithms are typically categorized into two main types based on the nature of the output. For regression tasks, algorithms like support vector regression, random forest, gradient boosting, and lasso/ridge regression are used to predict continuous outcomes, such as titer levels or percentages of charge variants.¹ In contrast, classification tasks use methods like decision trees, support vector machines, and ensemble models, which combine the predictions of multiple individual models to improve overall performance and robustness (e.g., random forests and gradient boosting, which can also be applied to regression tasks, are specific types of ensemble models) to categorize samples into distinct qualitative classes or variant profiles.^86,88 Both methods effectively model the nonlinear, multivariate relationships typical in mAb production systems.⁸⁴

Beyond supervised learning, unsupervised learning is crucial for identifying hidden patterns in complex bioprocessing datasets, particularly when labeled data are limited or incomplete.^85,89 Techniques such as PCA and clustering algorithms, including k-means and hierarchical clustering, identify metabolic bottlenecks, group similar culture conditions, and reduce data dimensionality for better interpretability.^56,86,90 PCA, while a core technique in MVDA is also widely used in ML workflows as a preprocessing step to enhance model interpretability and manage high-dimensional inputs.⁵⁶ For instance, PCA has helped reveal correlations between components of the culture medium and glycosylation profiles, facilitating the optimization of the culture medium without predefined outcomes.¹⁸ In parallel, reinforcement learning offers an adaptive, feedback-based approach that is ideal for dynamic optimization approaches in bioprocess development.⁸⁴ It optimizes feeding strategies and environmental conditions to maintain CQAs, such as charge heterogeneity, during bioreactor runs.⁸⁸ Although it is still relatively new in biomanufacturing, reinforcement learning holds great promise for making complex, time-sensitive decisions, especially when used alongside simulation tools to address the challenges of limited experimental data.^1,86 These ML approaches provide essential tools for developing predictive models and optimizing inputs to achieve product quality.^1,83 They are used to build models for critical quality attributes and to optimize culture conditions and media components.⁹¹

The development of predictive models for CQAs, which play a pivotal role in supporting quality by design (QbD) and process analytical technology (PAT) frameworks,^84,88,92 has been strongly influenced by the application of ML. The use of proxy measurements that correlate with desired CQAs, permits effective assessment for QbD during process development,^90,93 as charge heterogeneity in mAbs heavily depends on earlier process factors.⁵⁶ Supervised learning models use labeled datasets to connect critical process parameters, such as temperature, pH, and dissolved oxygen, with critical material attributes, such as media composition^28,85,94 This relationship helps predict the outcomes of CQAs, which can be continuous (e.g., percentage of acidic species) or categorical (e.g., compliant vs. non-compliant profiles).⁸⁹ This approach enables early detection of deviations and proactive quality management within process development using the principles of QbD and PAT.⁹⁰

Current applications of ML in optimizing cell culture conditions and medium components are proving invaluble in upstream bioprocessing for mAb process production.^1,95,96 The complexity, high dimensionality, and dynamic nature of upstream cell‑culture bioprocesses, along with the increasing availability of rich process data, particularly benefit from more advanced ML approaches.^97,98 A notable application is their use in predicting key performance indicators and critical process parameters.^83,97 For instance, artificial neural networks and hybrid deep learning architectures, such as convolutional neural networks (CNN) and long short-term memory (LSTM), have been successfully applied to model and optimize CHO cell culture processes, demonstrating improved productivity predictions.^96,97 Furthermore, ML can be used in media development and optimization. This involves applying multivariate analysis and active learning strategies, such as gradient boosting decision trees, to identify influential components and optimize their concentrations to enhance cell culture performance.^83,99,100 These models can also predict the dynamic concentrations of essential media components, such as amino acids, and interpret the effects of complex raw materials, such as yeast extract, on protein expression.^101,102

ML is not only limited to static optimization; it also encompasses dynamic optimization of cultivation parameters and establishes quantitative links between upstream conditions and product CQAs.^95,98,103 Advanced techniques like deep reinforcement learning are being studied for dynamic adaptive optimization of multiple process parameters during development, allowing for dynamic adjustments during cultivation to maintain optimal conditions and balance objectives such as productivity and product quality.⁹⁵ By combining process data with hybrid models that blend mechanistic insights and ML, these techniques improve prediction accuracy for CQAs like glycosylation and metabolic indicators, revealing how dynamic tweaks to culture and media conditions impact product quality.^103,104 Importantly, ML models, often supported by PAT tools such as Raman spectroscopy, are increasingly capable of mapping upstream variables to CQAs, including glycosylation and, notably, charge heterogeneity.^97,98,105 In this context, PAT tools serve as high-content sensors that facilitate data-driven modeling and upstream optimization during development, distinct from their typical use in real-time process control.¹⁰⁶ ML enables proactive optimization and fine-tuning of upstream conditions in the upstream process by correlating culture conditions and growth medium components with charge variants.^88,97 This approach empowers data-driven strategies that converge with QbD principles and improve management over critical quality characteristics such as glycosylation and charge heterogeneity.^56,99

Integrating machine learning for charge heterogeneity mitigation

ML can help one understand, predict, and ultimately optimize charge heterogeneity in mAbs during upstream processing.¹⁰⁷ Accordingly, ML, applied to charge heterogeneity, establishes quantitative relationships between upstream process parameters, medium composition, and resulting charge variant profiles by identifying complex patterns in large-scale bioprocess datasets.⁵⁶ ML models are uniquely suited to this task, as they learn patterns from comprehensive datasets that pair controlled upstream variables (such as pH, temperature, feeding strategies, and media component concentrations) with high-resolution charge variant analysis data.^37,45,95,107 By training supervised learning algorithms on this data, including regression models, classification models, and tree-based methods, the model learns the complex and often non-linear interplay between these factors and outcomes such as the percentage of acidic/basic species or the classification of acceptable profiles.^29,108 This capability is vital because parameters such as pH and temperature are well-known to affect charge-shifting PTMs such as deamidation, and ML can model the combined and interacting effects of the entire upstream environment.¹⁰⁹ Building on previous efforts in predicting PTMs like glycosylation, applying similar ML frameworks to charge variant data enables the identification of key upstream drivers of charge heterogeneity and supports targeted process optimization.^1,110

Data-driven studies in CHO cells demonstrate the ability to quantitatively and clearly link specific upstream inputs such as medium composition and culture parameters to charge variant outcomes. For instance, Gangwar et al. introduced an ML framework for CHO-GS cells cultured in a chemically defined basal medium to predict CQAs and optimize the culture medium. Their approach used 15 different regression models, combined with feature selection techniques, to analyze the impact of metal ions (as inputs) on acidic and basic charge variants (as outputs) (Table 4). This work identified iron and zinc as input variables most strongly influencing charge heterogeneity, and by fine-tuning their concentrations, the researchers achieved a charge variant profile closely resembling the reference innovative product. Compared to the conventional DOE approach, which reported an R² of 0.85, the ML-based optimization demonstrated superior predictive accuracy with an R² of 0.95, highlighting the enhanced performance of ML models in predicting charge variant outcomes and optimizing medium composition.²⁹

Table 4.

Machine learning/Data-driven optimization of culture parameters influencing charge variant.

Cell line	Parameter	Media	Product	Algorithm	Charge Variant	Ref.
CHO-GS	Cobalt, Nickel, Magnesium, Manganese, Zinc, Iron, Copper	CD CHO®	IgG1 (Trastuzumab)	Random forest regression, linear regression, lasso regression, decision tree regression, extra tree regression, ridge regression, lasso least regression, Bayesian ridge, catboost regression, Huber regression, extreme gradient boosting, gradient boosting regression, elastic net, support vector regression, and k-nearest regression	Iron reduces acidic variants; Zinc increases basic variants.	(Gangwar et al.,29)
CHO-GS	Glucose, Lactate, pH, Dissolved Oxygen, Temperature, Ammonia	CD CHO®	IgG1	CNN, LSTM, Bayesian Optimization, Random Forest, K-nearest Neighbor	39.5% reduction in total acidic and basic variant levels	(Lu et al.,97)
CHO-S and CHO-GS	pH, temperature, initial seeding density	Serum-free commercial medium	IgG1, IgG4, (IgG+IgM)	Statistical Modeling, PCA, Transfer Learning, Active Learning	Higher pH increases Acidic Variants and decreases Basic Variants; Temperature shifts also inversely affect charge variants.	(Kumar et al.,111)

Extending beyond medium composition, Lu et al. showcased how ML can effectively enhance the entire CHO cell culture process. They used a deep learning model – a combination of CNN-LSTM networks – to predict and fine-tune key process parameters such as nutrient levels, temperature, and pH (inputs). Applying this framework, they achieved significant improvements in product-related outputs, including a 28.1% increase in titer, a 19.1% increase in overall process yield, and a 39.5% reduction in charge variant percentage, demonstrating the comprehensive upstream process improvement enabled by ML. The model achieved an R² of 0.956 and an RMSE of 0.082, demonstrating a substantial improvement in predictive accuracy compared to traditional models, which generally yield lower R² values and higher error metrics.⁹⁷ Furthermore, Kumar et al. presented a data-driven approach to optimize key physical process inputs – specifically pH, temperature shift, and initial seeding density – for achieving the desired product quality outputs, such as N-glycan and charge variant profiles, in mAb biosimilar production. Using multivariate statistical modeling across multiple CHO cell lines, they examined how variations in these input parameters affected the resulting CQAs. Their findings revealed consistent relationships among quality characteristics. By applying transfer learning and active learning from prior projects, they effectively reduced the experimental workload while ensuring that target output attributes were successfully achieved.¹¹¹

Implementing an ML-based strategy to mitigate charge heterogeneity in mAb production typically follows a structured workflow that integrates data collection, model development, and process optimization.^1,21 The first step involves data collection, where high-quality datasets from upstream bioprocesses, including CHO cell culture conditions and media compositions, are collected alongside analytical outputs such as charge variant profiles.^112,113 Subsequently, this step involves the preparation of the gathered raw data through preprocessing techniques. This critical step ensures data quality and consistency by cleaning the dataset, handling missing or noisy values, normalizing variables, and applying feature engineering or selection to extract the most relevant inputs affecting charge heterogeneity.^{1,107,114,115} After preprocessing, appropriate ML models are selected and trained. Based on the specific goal, such as predicting the levels of acidic or basic species using regression models or classification models to classify heterogeneous profiles, appropriate algorithms are selected to achieve high prediction accuracy (e.g., R² > 0.95) and interoperability.^29,111,116 Figure 2 illustrates the typical stages involved, from initial data handling to operational implementation.

Figure 2.

Machine learning-driven workflow for monoclonal antibody charge heterogeneity mitigation.

After training, the models undergo rigorous validation and evaluation using independent datasets distinct from the training dataset.¹⁰⁷ This process assesses their predictive accuracy, robustness, and generalizability across different process conditions and scales, often using cross-validation techniques and real-time bioprocess data, as demonstrated by Kumar et al.^110,111,115 Following this stage, model interpretation can be a valuable step to gain insight into how specific upstream parameters, such as pH or media components such as iron and zinc, affect charge heterogeneity, enhancing process understanding and guiding further optimization.^18,52,107 Once model performance is validated, it is integrated into the bioproduction workflow for prediction and optimization purposes.¹ The model can simulate different parameter combinations to predict charge profiles and identify optimal culture conditions and media compositions that minimize undesirable variants while meeting other production objectives.^117,118 Optimal culture conditions are determined by generating candidate conditions, which are input into the trained ML model to predict charge profiles.²⁸ The conditions yielding the best outcomes, such as minimal charge heterogeneity, are selected as optimal, enabling precise adjustments that enhance mAb production efficiency.²⁹ Finally, continuous monitoring and improvement, incorporating active learning principles and feedback loops, ensure the long-term relevance of the model, where new process data is used to retrain and periodically update the model, aligning with Bioprocess 4.0 principles as a forward-looking framework for bioprocess optimization and product consistency.^28,119,120 This iterative workflow enables proactive optimization of charge heterogeneity profiles, increasing yield and quality in mAb bioprocesses.¹²¹

Although promising, scaling and validating ML models to industrial bioreactors poses challenges. Scalability refers to the ability of models developed at laboratory scales to accurately predict charge variant profiles in large-scale production bioreactors.¹²² This transition is challenging due to scale-dependent differences in mixing, mass transfer, and changing process dynamics.^93,123 These factors can all impact cellular metabolism and the kinetics of PTMs.¹ In addition, models must be sufficiently robust to handle the increasing data volume and variability that arise from industrial operations.¹²⁴ Validation is equally critical and requires rigorous evaluation using independent manufacturing datasets to demonstrate accuracy, robustness, and reliability across batches, raw materials lots, and operating sites.^84,125 Validation strategies include k-fold cross-validation, use of dedicated test sets, and prospective studies to ensure that models meet stringent regulatory standards for biopharmaceutical applications.^93,126 Ultimately, successfully scaling and validating ML models to reliably predict complex CQAs such as charge variants in industrial biomanufacturing is not straightforward.^56,111 It requires a comprehensive strategy that effectively brings together robust data infrastructures, advanced computational and modeling approaches, and comprehensive data management strategies.^124,127 Most importantly, the development of industry standards and clear regulatory frameworks will be important.⁸⁸

Conclusion: challenges, opportunities, and future directions

MAbs are the dominant modern biopharmaceuticals, yet maintaining consistent product quality, especially in the case of charge heterogeneity, remains a major challenge. We showed that mAb charge heterogeneity primarily arises from a diverse set of PTMs that occur during the complex biomanufacturing process. These modifications are greatly impacted by culture conditions and media components. Traditional bioprocess optimization methods, such as OFAT, DoE, statistical modeling, and routine process controls, have been essential for optimizing cell growth and titer, and for gaining a basic process understanding. These methods, however, have inherent limitations in accurately optimizing the complex nature of the charge variants. This arises from the high dimensionality, non-linear relationships, and complex interactions between factors governing PTMs, along with the need for dynamic process management. ML has emerged as a powerful, data-driven approach that can overcome these traditional limitations by modeling complex biological dynamics and making accurate predictions. Notably, ML techniques, such as deep learning and hybrid modeling approaches integrated with PAT, are showing increasing effectiveness in quantitatively linking upstream process parameters and media components to product CQAs, including glycosylation and, in particular, charge variant distribution. This predictive capability, combined with ML-based optimization and advanced optimization frameworks, provides a promising path to proactively and strategically mitigate charge heterogeneity during upstream bioprocesses. This strategy also supports the integration of QbD principles into bioprocess development.

While ML holds promise for resolving charge heterogeneity in industrial mAb production, several barriers remain to its widespread adoption. A major hurdle lies in ensuring both the quality and accessibility of data. Training effective ML models requires large, diverse, and detailed datasets that capture the complex relationships between upstream conditions, medium composition, and charge variant profiles across batches and scales. Historical manufacturing data, however, are often incomplete, inconsistent, or lacking sufficient detail, and generating comprehensive experimental data from scratch can be time-consuming and expensive. Moreover, preparing this data, by correcting inconsistencies, handling missing entries, and managing variability, requires substantial manual effort.¹²⁸ Scalability is another concern. Models that perform well in the lab or pilot are not always effective in industrial settings due to scale-dependent factors such as mixing, mass transfer, and operational variability. Furthermore, many ML models, especially deep learning approaches, are often seen as “black boxes,” making it difficult to interpret their predictions or link them to underlying biological mechanisms.¹²⁹ Perhaps most importantly, the regulatory landscape for ML in Good Manufacturing Practice environments is still evolving. Standard guidelines on model validation, transparency, data integrity, and lifecycle management are needed to ensure trust and compliance. Without them, widespread adoption in the industry remains a challenge.

Despite challenges around data quality, model scalability, and regulatory acceptance, ML offers powerful new avenues for harnessing charge heterogeneity in mAb production. By integrating ML with real-time PAT tools, such as Raman spectroscopy, Fourier transform infrared spectroscopy, or fluorescence sensors, we can continuously feed high-resolution time-series data into predictive models. These models can actively adjust variables such as pH, temperature, or nutrient feeds to lock in target charge profiles. At the same time, automated micro‑bioreactor platforms and high-throughput analytics can generate the rich and diverse datasets that ML requires. Active learning strategies can focus experiments on the most informative conditions, reducing development time and cost compared to traditional factorial studies.¹³⁰ On the modeling side, specialized algorithms tailored to individual PTMs, such as deamidation or oxidation pathways, can be paired with explainable Artificial Intelligence tools such as SHAP, and complemented by inherently interpretable tree-based algorithms (e.g., decision trees, random forests).²⁹ The combination reveals exactly why and how specific process tweaks affect each modification. This transparency increases process understanding, supports regulatory communications, and enables more targeted cell line and media engineering. Moreover, ML can optimize multiple CQAs simultaneously, ensuring balanced control of charge variants, glycosylation, aggregation, and overall product quality.

Going forward, the computational approaches described here are being increasingly adopted to facilitate the establishment of a “smart factory” vision for bioprocessing. In the context of Bioprocessing 4.0, ML-based digital twins continuously analyze PAT streams and operational data to simulate and optimize every aspect of production, both upstream and downstream, in silico.¹¹⁹ Real-time predictive control will replace end‑point testing, automatically adjusting bioreactor settings to maintain CQAs within precise ranges, thereby reducing batch variability and the failure risk. The insights gained from this process will also lead to more rational designs of cell lines and culture media, embedding robust charge profiles from the very beginning of development. Together, these innovations promise better product stability, lower costs, faster time to market, and broader global access to life‑saving mAbs. Ultimately, this will make high-quality biologics more affordable, scalable, and responsive to emerging health needs.

Funding Statement

This work was supported by the NIHMS under Grant [R35 GM119850 (NEL)].

Disclosure statement

No potential conflict of interest was reported by the author(s).

CRediT authorship contribution statement

Hossein Kavoni: Conceptualization; investigation; writing-original draft; writing-review and editing. Iman Shahidi Pour Savizi: Investigation; writing-review and editing. Saratram Gopalakrishnan: writing-review and editing. Nathan E. Lewis: writing-review and editing. Seyed Abbas Shojaosadati: Conceptualization; project administration; supervision.