Trastuzumab (sold as Herceptin) is one of the most widely used cancer drugs in the world. It is a monoclonal antibody, a protein engineered to latch onto the HER2 receptor that drives some of the most aggressive breast cancers, and it has helped make HER2-positive disease one of the most treatable forms of the illness.
Antibodies like trastuzumab are too large and complex to make with chemistry. They're grown by living cells, usually Chinese Hamster Ovary (CHO) cells, engineered to secrete the drug. Those cells are cultured in a bioreactor: a sealed, stirred steel or single-use vessel that holds temperature, pH and dissolved oxygen steady while the cells multiply and produce the antibody over roughly two weeks.
Many commercial bioreactors are fed-batch, where the culture starts with a fixed volume of nutrient-rich media, and operators add concentrated feed, mostly glucose, as the cells consume it. Feed at the right time and the cells stay healthy and productive; get it wrong and a batch worth hundreds of thousands of dollars can lose yield or quality.
The hard part is that engineers and scientists are unaware of the process behaviour between samples. Glucose and cell density are measured by pulling a sample, usually once a day, and sending it to an offline analyser. Results arrive hours later, and in between the culture keeps changing. Operators end up feeding on a fixed recipe or a best guess, which is slow, labour-intensive and hard to scale. Spectroscopic PAT such as Raman or NIR reads glucose more often, but a probe reading is itself a model, spectrum in, glucose out. It tells you where glucose is, not why the cells are using it at that rate or what to feed next.
With CPI, Labman and Nicoya, we closed that loop on small-scale fed-batch bioreactors producing a trastuzumab biosimilar candidate. A hybrid model estimates glucose and viable cell density between samples, an extended Kalman filter corrects it against one robotic sample a day, and a model predictive controller sets the feed without an operator.
Cells use glucose as their primary fuel source, which makes it the most important state variable in a fed-batch run. It drives growth, viability and productivity, and it moves faster than any daily sample can follow. Get the balance wrong in either direction and the culture starts poisoning itself.
When glucose runs too high, cells burn it inefficiently. Instead of routing it fully through the mitochondria, they convert much of it to lactate (LAC) and pump it out. When glucose runs low, cells turn to glutamine as a backup fuel, and breaking glutamine down releases ammonia (AMM). Both by-products build up in the medium with nowhere to go. Lactate acidifies the culture; ammonia is toxic to the cells. Either one slows growth, cuts viability and can change the quality of the antibody being made, and both are hard to reverse once they have accumulated.
Keeping glucose inside a narrow window avoids both. A daily sample cannot do that on its own, and sampling more often is not an option, because every sample removes working volume. Raman or NIR probes close the gap but add hardware, method development and upkeep, and not every vessel or scale can take one.
To control a bioreactor you need to know what the cells are doing, not just what the sensors read. A cell growth model describes the culture as a set of linked mass balances: how fast cells grow and die, how fast they consume glucose and glutamine, and how fast they make antibody and by-products. Each balance is driven by rates that change as conditions in the vessel change.
The structure comes from biology; the hard-to-know rate terms are learned from batch data. That is what makes the model hybrid. The terms below are the ones that appear throughout this post.
Living cells per millilitre of culture, typically in millions of cells/mL. The main measure of how much "factory" is in the vessel.
The share of cells that are alive. It stays high during healthy growth and falls as nutrients run out or by-products build up.
How fast the cell population is growing per hour. It depends on glucose, glutamine and the level of lactate and ammonia.
How much oxygen the culture consumes per hour, worked out from the gas flows and dissolved oxygen the control system already logs. A live signal of metabolic activity between samples.
Glucose consumed per cell per day. Multiplied by VCD, it tells the model how fast glucose is being used up.
The concentration of antibody in the culture, in g/L. The commercial output of the run.
Antibody produced per cell per day. Titer is the result of VCD and qP over time.
By-products of glucose and glutamine metabolism. Both accumulate in the medium and slow growth.
The system runs on three streams of data, all from equipment already on or around the vessel. No new probes were added.
75 tags logged every minute, including pH, dissolved oxygen and feed volumes, streamed from CPI's ambr bioreactors through the PI Web API. This stream drives the hybrid model between samples.
One automated sample a day measures glucose and viable cell density offline. These are the ground-truth points the EKF uses to correct the model.
Antibody concentration measured from the same daily samples, used to track productivity and compare controlled and manual runs.
Three components run continuously during the batch. Together they read the process, estimate its state and act on it, using only the instrumentation already on the vessel.
Mass balances for cell growth, glucose consumption and product formation, with kinetic terms learned from batch data. It reads 75 process tags every minute, including pH, dissolved oxygen and feed volumes, and forecasts glucose and VCD four hours ahead with an uncertainty band.
An extended Kalman filter (EKF) keeps the estimate honest. Labman's robotic sampler draws one sample a day; when the result lands, the EKF compares it with the forecast and corrects the model state, so every later prediction follows the corrected trajectory.
A model predictive controller (MPC) uses the forecast to choose feed timing and amount, keeping the lower edge of the glucose band above its limit. It writes the setpoint straight to the feed pump every four hours, within limits the team sets.
The standard deviation of glucose fell from 2.25 to 0.82 g/L, so swings shrank by almost two thirds.
Glucose stayed inside the target band for 81% of the batch, up from 47% with manual feeding.
Share of the batch with no glucose left: about 20% under manual feeding, 0% under model control, because the controller fed ahead of demand.
We compared CPI's manual feeding strategy with closed-loop control in paired bioreactors. On the standard run, manual feeding produced large swings between daily feeds, while the controller made smaller feeds every four hours, held glucose in a band about half as wide, and cut those swings by about 64% (Figure 5). When we doubled the inoculum, cells consumed glucose faster than the recipe expected: the manual arm (BR21) ran out of glucose before its next feed, more than once, while the controlled arm (BR22) stayed in range throughout, because the forecast saw the higher demand coming and fed ahead of it.
Across runs, model-driven control cut glucose variability by about 64% (standard deviation 2.25 → 0.82 g/L), raised the share of the batch spent in the target band from 47% to 81%, and took glucose depletion from about 20% of the batch to zero. Closed-loop control also improved cell viability and titer compared with manual control, which left the reactor glucose-limited.
Closed-loop control held glucose in a band more than 50% tighter than manual feeding.
Steadier glucose means steadier cells. Holding feeding inside the target window keeps the culture away from the starvation dips and overflow spikes that push cells toward lactate and ammonia, the kind of consistency that matters when a process moves from development to manufacturing.
The model did not need to be perfect at the start, parameters were deliberately initialized far off to assess the model's capabilities to rapidly self correct.
Process development is the first place this pays off. With glucose held on target automatically, it stops being a confounder: teams can screen setpoints, media and clones, rerun fewer experiments, and study low-glucose strategies that are too risky to run by hand. In manufacturing, the same loop feeds ahead of inoculum and consumption swings and covers nights and weekends without an operator. None of this requires replacing existing PAT. Where Raman or NIR is already installed, its reading becomes one more input to the model, and the controller adds the decision layer on top. The customer keeps ownership of control decisions; Basetwo provides the models, controller and tooling.
Work with engineers and researchers building autonomous models for biological and chemical production.
