Bespark*bio reached the same decisions with 8 runs instead of 22

Bespark*bio is a biotech contract research and development organization of about ten people, a CRO. It develops and optimizes bioprocesses for companies from a wide range of fields, from raw material suppliers to biotech startups. The campaign on this page optimized one of those processes, an upstream viral-vector production process.

The goal was to optimize the process, for a higher virus titer, fewer impurities or a shorter process time, and to get there with fewer bioreactor runs. Each run takes about two weeks of operation and a panel of assays, so every run saved is material, time and analytics not spent.

Three process parameters were in question, temperature, pH and the multiplicity of infection (MOI). The virus titer, the viable cell density (VCD) and two impurities, double-stranded DNA and host-cell protein (HCP), were measured throughout each run so that a model could predict them over time.

Bespark*bio had originally run a design of experiments (DoE) of 22 bioreactor runs in blocks of four parallel 1 L bioreactors, over 30 to 37 °C, pH 7.0 to 7.6 and an MOI from 0.0005 to 0.005. Temperature had the strongest influence on the titer, with the highest titers between about 30 and 34 °C. The titer was far less sensitive to pH and MOI, so both can be run over a wide range. The original design gave these answers with 22 runs, and the question was whether a reduced design would give the same.

A hybrid model was trained on 8 of the 22 runs, 64% fewer experiments, spread across the design and chosen without looking at a single titer or impurity value. Tested on the 14 runs it never saw, it predicted the virus titer with an R² of 0.92. It also made the same decisions as the model trained on all 22 runs. In both, temperature is the parameter that explains most of the variation in the titer, and the optimum lies in the low 30s °C.

One selection could be a lucky draw, so eight different selections were tested, four of them random. All eight found temperature to be the dominant factor, and seven of eight predicted the optimum in the same place, between 30 and 32.5 °C. One selection had never seen a run at the optimum temperature and found it anyway, because the mass balances in the hybrid model describe how the process responds, and the data only has to determine the rates.

In this process the virus is released together with the impurities. Harvesting at the titer peak adds the last few per cent of titer but also a large share of the impurities, which the downstream process then has to remove.

For every condition, the hybrid model shows when the titer reaches 95% of its maximum, which opens the harvest window, and how much impurity each further gain in titer brings with it. The longer one waits, the more impurity each gain brings. Temperature also sets the timing. Running at 32.5 °C instead of 31 °C lowers the peak titer by less than 0.1 log and opens the window about a day earlier.

The model trained on 8 runs then proposed three new runs, at conditions it predicted would lower the impurities and shorten the process at the same or a higher titer. They were carried out in a later campaign, and the reduced model predicted them as well as the model trained on all 22 runs. In the run shown, harvesting at 118 h instead of at the titer peak at 140 h kept about 90% of the peak titer at half the host-cell protein.

Bespark*bio estimates a reduced design of 8 runs instead of 22 at about €40,000 of raw material and about 12 weeks of development saved. The model built from those eight gives the same process decisions as the full design and a harvest point that keeps the titer and halves the host-cell protein.