Repligen mapped a biosimilar design space from five experiments

Repligen is a bioprocessing technology company working in the field of biologic drug production, supplying biopharma companies and CDMOs with filtration, chromatography, process analytics and fluid management. In this project it contributed the MAVEN in-line glucose and lactate measurement and the ZipChip analyzer for product quality. Since 2025 Repligen has been a strategic partner of and investor in Novasign.

Glycation rises steadily with glucose exposure, while the glycan profile suffers when glucose runs short, and the temperature shift between the growth and the production phase acts on growth, productivity and glycosylation at the same time.

In a biosimilar development, reference data from the originator define the acceptable range of each attribute in advance, and the process has to meet all of them consistently. The answer is not a single optimum but a design space in the sense of ICH Q8(R2), the range of glucose setpoints and post-shift temperatures within which every attribute stays inside its range. The direction of each effect is known, its exact boundaries for a given molecule and process are not, so they have to be measured across the whole parameter space rather than one factor at a time.

That takes three conditions. Glucose has to be a controlled process parameter rather than a fluctuating outcome, both quality attributes have to be measured throughout the cultivation, and the data have to be described by a model that carries the structure of the process. The project tested whether, under these conditions, five experiments are enough to map the design space.

The project drew on five data sources, each with its own format and time base: the bioreactor, offline viable cell density, offline titer, MAVEN for glucose and lactate, and ZipChip CE-MS for the glycan profile and glycation. In Novasign Studio, a dataset reader read each source, sort and filter nodes prepared it, and merge nodes combined the streams into one dataset aligned on process time.

The whole preprocessing workflow can then be packaged into a single node, so the team working with the model sees one preprocessing step and the complexity behind it runs without anyone adjusting its settings.

Temperature is recorded continuously by the bioreactor, but glucose is usually measured offline once or twice a day and swings between bolus feeds. MAVEN closes that gap. It measures glucose and lactate in-line without sampling and adjusts the feed rate automatically, so glucose becomes a controlled process parameter held at its setpoint. ZipChip CE-MS measured the glycan profile and glycation in the same samples, and Novasign Studio brought both data streams together in one hybrid model.

Five bench-scale fed-batch cultivations of the NISTCHO cell line ran for about 14 days and were evaluated up to 320 h, varying two factors. MAVEN feedback control held glucose at 1, 3 or 5 g/L, and because the basal and feed media contained no glucose, the controlled feed was its only source. Every run started at 37 °C, and at about 160 h the temperature was either held or shifted to 33.5 or 30 °C. The five experiments were designed in Novasign Studio, the four corners of a two-level full factorial and its centre point.

Viable cell density was measured at-line and titer offline. Glycation and two glycoforms were quantified by ZipChip CE-MS at five to eight time points per run during the second half of the cultivation. The charge variants were measured too and showed no systematic dependence on either factor.

The model was trained on one data table holding the five runs, with the run, process time, reactor volume, the two process parameters and the measured quality attributes. In Novasign Studio the analysis is a chain of nodes. The Dataset Reader imports the table, the Reactor Model trains the hybrid model, the Model Selector picks the most suitable model, the Reactor Process Simulator simulates the parameter space and the Simulation Explorer shows the result. When further runs are added, the same chain runs again without being rebuilt.

The temperature after the shift and the glucose setpoint are the process parameters. The model has three parts that run in sequence at every time step. A mass balance describes the dilution of cells and product from first principles, a small neural network computes the specific rates from the current state and the process parameters, and a kinetics step integrates those rates to the next time point. Only the network is learned from data, which is why five runs are enough to train it.

Because the best training settings are not known in advance, a tuner trained the model 45 times over combinations of network seed, data split and optimizer step size, each time holding one run out for validation. The best trial held out the centre point, so a model trained on the four corners was tested in the middle of the space. Its final checkpoint reached a validation RMSE of 0.95, against 2.84 before training, with a training R² of 0.996, and training and validation loss fell in parallel throughout.

In the digital twin a process parameter can be declared as a variable with a range, which turns a single recipe into a grid of conditions simulated in silico. Here 15 post-shift temperatures from 30 to 37 °C and 9 glucose setpoints from 1 to 5 g/L give 135 conditions. Each was simulated from the measured starting point of all five runs, 675 simulations up to 320 h.

The quality criteria are then applied to that grid as filters, each with an acceptance range and the time at which it is evaluated, and every condition that fails one of them drops out. What remains is the design space, colored by predicted titer, so the most productive condition inside it can be read directly. The limits here are examples. In a program the product specification, for instance derived from originator data, goes into the same panel.

With these example limits, at 320 h, G0F/G1F between 10.3 and 29.2% and glycation below 19%, shown from a predicted titer of 1.1 g/L, the acceptable conditions form a wedge at low glucose and high post-shift temperature. At 1 g/L a moderate temperature shift is still acceptable, at 2.5 g/L only the condition without a shift meets every criterion, and above 2.5 g/L no condition in the grid does. Set as two independent ranges, the post-shift temperature has to lie between 35 and 37 °C and glucose between 1 and 1.5 g/L. Inside the region the highest predicted titer is at 1 g/L without a temperature shift, so here the conditions of acceptable quality are also the most productive ones. The diagonal edge of the region is the interaction between glucose and temperature, which a one-factor-at-a-time approach would not capture. The design space finder reads each condition’s titer as the mean of all five of its runs, which widens the same wedge by a few conditions between 1.5 and 2.5 g/L.

The result is an operating region that shows where both attributes meet their limits and where they compete, which optimizing one attribute after the other cannot provide. The whole workflow, from the combined online and offline data to the design space, runs in Novasign Studio, and every step stays traceable for the documentation a regulatory submission needs.

Changes after approval are expensive, because each one needs a comparability exercise between the product before and after the change under ICH Q5E, generally with commercial-scale data, and this applies to biosimilars too. The design space therefore has to be known and verified before filing. Verification runs at its boundaries, first at laboratory scale and then at production scale, are added as new data and refine the same model.