Real-Time Titer, Glucose and CQA Prediction in Biologics: How AI Soft Sensors Work

For the basics, read the practical guide to soft sensors. To understand the models behind the estimates, see how hybrid AI uses biology and physics. For implementation, explore the BioMedAna Soft Sensors solution and the platform buyer’s guide.

What If You Could Predict Titer, Glucose and Product Quality Before the Next Lab Result?

In biologics manufacturing, some of the most important measurements are not available continuously.

A bioreactor may provide real-time information about pH, dissolved oxygen, temperature, agitation, and gas flow. Yet measurements such as product titer, glucose concentration, and critical quality attributes (CQAs) often depend on periodic laboratory analysis.

This creates a gap between what manufacturing teams can observe and what they need to understand.

Is product titer increasing as expected? Is glucose being consumed faster than anticipated? Could changing process conditions indicate an emerging quality risk?

AI-powered soft sensors help answer these questions by estimating difficult-to-measure variables using available process data, scientific models, and machine learning.

For Process Development, MSAT, Manufacturing, and Quality teams, these predictions can provide earlier insight into process performance and support more informed decisions.

This article explores three important applications: real-time titer estimation, glucose prediction, and CQA risk monitoring.

What Are AI Soft Sensors in Biologics Manufacturing?

An AI soft sensor is a software-based predictive model that estimates a process variable using available measurements and relationships learned from historical data or represented through scientific models.

Unlike a physical sensor, a soft sensor does not necessarily measure the target variable directly.

For example, a titer soft sensor may estimate product concentration using cell growth, feeding history, process measurements, and analytical reference data.

A glucose soft sensor may use Raman spectroscopy and process measurements to estimate nutrient concentration between laboratory samples.

A quality-related soft sensor may estimate selected attributes or flag conditions associated with quality variability.

Soft sensors can use statistical methods, machine learning, mechanistic modeling, or hybrid approaches combining scientific knowledge and AI.

Their purpose is to make useful process information available sooner—not to replace scientific judgment or automatically eliminate laboratory testing.

The practical guide to soft sensors in biopharma covers upstream and downstream applications, and the hybrid AI guide explains how biology and physics inform prediction.

Real-Time Product Titer Prediction: Understanding Productivity Earlier

Why is titer prediction important?

Product titer measures the concentration of a desired biological product, such as a monoclonal antibody, in a manufacturing process.

It is an important indicator of productivity and is frequently used to assess process performance.

However, titer is typically measured through offline analytical methods, creating intervals when current product concentration is not directly known.

How does an AI titer soft sensor work?

A titer soft sensor estimates product concentration using available information such as:

Viable cell density and growth history.

Feeding strategy and cumulative feed.

Glucose and lactate measurements.

Temperature, pH, and dissolved oxygen.

Cultivation time and historical titer results.

Depending on the application, the model may use statistical relationships, machine learning, biological kinetics, or hybrid modeling.

Predictions can update as new measurements become available, subject to data availability and the model's execution requirements.

Example: Titer prediction during mAb manufacturing

Consider a fed-batch monoclonal antibody process.

At Day 10, the latest offline titer result is 2.6 g/L.

A soft sensor estimates the evolving titer between laboratory measurements based on the available process trajectory.

As the run progresses, scientists can compare the estimated product concentration with historical successful batches.

If the predicted trajectory begins diverging from the expected behavior, the team can investigate potential differences in growth, metabolism, or process conditions.

Business value: Earlier visibility into productivity trends, more informed batch comparisons, and additional evidence for harvest planning.

A titer estimate is not a substitute for an analytical measurement. Its usefulness depends on demonstrated accuracy and reliability for the intended process.

Explore hybrid CHO titer and glucose prediction and monoclonal antibody process intelligence.

Real-Time Glucose Prediction: Monitoring Metabolism Between Samples

Why is glucose monitoring important in cell culture?

Glucose is an important nutrient in mammalian cell culture and microbial fermentation.

Its availability influences cellular metabolism and process performance.

Glucose is frequently measured through offline sampling, although online analytical technologies are also available.

Between measurements, glucose concentration may change as cells consume nutrients and additional feed is introduced.

How does a glucose soft sensor work?

A glucose soft sensor estimates concentration using suitable information such as Raman spectroscopy, process measurements, feeding history, and biological relationships.

Raman-based approaches use spectral measurements and calibrated statistical models to estimate glucose concentration.

Other approaches may incorporate nutrient-consumption kinetics and process data.

Example: Detecting unexpected glucose consumption

Imagine a CHO cell culture process operating under an established feeding strategy.

The most recent laboratory measurement indicates glucose is within the expected range.

However, the soft sensor subsequently estimates that glucose is declining more quickly than in comparable historical batches.

This observation gives Process Development or Manufacturing teams an opportunity to review current nutrient trends, compare feeding and metabolic behavior, and determine whether additional confirmation or investigation is warranted.

The prediction does not automatically justify changing the feed rate.

Business value: Better nutrient visibility, earlier investigation of metabolic changes, and more informed feeding strategy evaluation.

What does scientific research show?

A 2025 study published in Pharmaceutics investigated Raman-based monitoring of nutrients, metabolites, IgG titer, and cell density in 10-liter CHO cell cultures.

The researchers reported a root mean squared error of prediction of 0.51 g/L for glucose and 0.12 g/L for IgG titer on their evaluation data.

These findings demonstrate the potential of Raman-based soft sensing but are specific to the study's process, instrumentation, and model evaluation.

They are not universal accuracy benchmarks or BioMedAna performance claims.

CQA Prediction: Identifying Emerging Product Quality Risks

What are critical quality attributes?

Critical quality attributes are physical, chemical, biological, or microbiological properties that must be controlled within appropriate limits to help ensure product quality.

For monoclonal antibodies, relevant attributes may include glycosylation profiles, charge variants, aggregation, and other product-specific characteristics.

These attributes can influence product safety, efficacy, stability, or manufacturability.

Unlike routine bioreactor measurements, many CQAs require specialized analytical methods.

Can AI soft sensors predict CQAs?

Yes, selected CQAs may be estimated using process-specific predictive models when adequate analytical measurements and scientifically relevant process data are available.

However, CQA prediction is often more challenging than estimating glucose or titer.

Product quality can be affected by complex interactions involving cell metabolism, raw materials, process conditions, product structure, and downstream operations.

An AI model must demonstrate that its predictions are reliable for the specific attribute and operating conditions.

Example: Monitoring glycosylation-related risk

Consider an mAb manufacturing process in which historical evidence suggests that certain combinations of culture conditions and metabolic behavior are associated with differences in glycosylation.

A quality-related soft sensor may use these process relationships to estimate a selected glycosylation-related attribute or generate an indicator of increased quality risk.

If the current batch begins following a trajectory associated with unfavorable historical outcomes, MSAT and Quality teams can review the evidence and determine whether additional analytical assessment is needed.

Business value: Earlier awareness of potential quality variability, better-informed investigations, and stronger process understanding.

Importantly, a predicted CQA or quality-risk indicator is not equivalent to a measured product-quality result and should not automatically replace approved analytical methods or release testing.

How Do Titer, Glucose and CQA Predictions Work Together?

These three applications provide complementary information about manufacturing performance.

Consider a batch showing lower-than-expected titer.

The titer soft sensor identifies a difference in productivity trajectory.

The glucose soft sensor provides additional information about nutrient consumption and metabolic behavior.

A quality-related model may help assess whether the observed operating conditions are associated with particular CQA risks.

Together, these estimates help scientists form a more complete understanding of the process.

The individual predictions remain distinct, and an apparent association does not establish causality.

What Does an AI Soft Sensor Dashboard Show?

For end users, the value of soft sensing depends on how clearly predictions are presented.

A useful dashboard should show:

Estimated values and trends.

Latest measured laboratory results.

Measured-versus-predicted comparisons.

Uncertainty or prediction intervals where appropriately supported.

Historical or reference batch trajectories.

Model status and operating-range limitations.

For example, a titer dashboard could show a continuously updated estimated trajectory alongside the latest offline assays.

The user should be able to distinguish measured results from model estimates and understand whether the prediction remains within evaluated conditions.

This is particularly important when soft sensor outputs influence manufacturing or quality decisions.

How Accurate Are AI Soft Sensors?

There is no single accuracy level that applies to all soft sensors.

Performance depends on the predicted variable, process characteristics, model design, measurement quality, and the representativeness of the evaluation data.

A glucose soft sensor supported by informative Raman measurements may perform differently from a CQA model based on indirect process indicators.

Before operational use, manufacturers should evaluate predictions against independent reference measurements, assess performance across representative batches, and establish acceptable errors for the intended decisions.

Users should also understand what happens when required data is unavailable or process conditions change.

A trustworthy soft sensor provides not only predictions, but also evidence of when those predictions can be relied upon.

When comparing platforms, use the buyer’s guide to soft sensor and digital twin software to scope a representative pilot and define success criteria.

Can AI Soft Sensors Be Used in GMP Manufacturing?

AI soft sensors can support GMP manufacturing when their intended use, performance, risks, and lifecycle controls are appropriately addressed.

There is a meaningful difference between displaying a predicted titer as an additional monitoring indicator and using that estimate to make an automated manufacturing decision.

For higher-impact applications, manufacturers need appropriate qualification or validation, data integrity, model traceability, monitoring, change control, and human oversight.

The same principle applies to CQA predictions.

A model that identifies a quality risk may support investigation, but replacing an approved analytical method requires additional scientific, quality, and regulatory justification.

How BioMedAna Connects Titer, Glucose and Quality Prediction

BioMedAna provides a modality-native AI platform that combines process data, scientific models, digital twins, and governed analytical workflows.

Rather than treating soft sensors as isolated predictions, BioMedAna connects them to broader Process Development, MSAT, and Manufacturing activities.

Bio Hub connects process measurements, laboratory data, PAT signals, and batch context.

Bio Twin provides mechanistic, statistical, and hybrid modeling capabilities for process estimation, prediction, and simulation.

Bio Agents support batch comparisons, process investigations, prediction reviews, and source-backed scientific analysis.

Bio OS provides governance, traceability, access control, model lifecycle management, and human oversight.

For a titer prediction application, this means scientists can examine estimated productivity alongside cell growth, metabolite behavior, process conditions, and historical batch evidence.

For a CQA-related application, predictions can be reviewed with appropriate analytical evidence and scientific context.

This integrated approach connects predictive soft sensing to the decisions bioprocess professionals actually make.

Conclusion: From Delayed Measurements to Predictive Bioprocess Intelligence

Biopharmaceutical manufacturing increasingly requires timely information about biological performance, productivity, and product quality.

AI-powered soft sensors can help bridge the gap between continuously available process measurements and variables that are difficult to measure frequently.

Titer prediction provides earlier insight into productivity.

Glucose prediction improves visibility into nutrient and metabolic behavior.

CQA-related models can support earlier assessment of selected quality risks when suitable evidence is available.

Together, these capabilities can strengthen process understanding and help teams make more informed decisions.

The objective is not simply to predict more variables. It is to make reliable, scientifically grounded information available when it matters most.

Discover how BioMedAna combines biology, physics, process data, hybrid models, and AI to support titer, metabolite, and quality-related prediction in biologics manufacturing.