At an Intellegens webinar, Tosoh Bioscience presented a case study of the use of AI-driven Design of Experiments (DOE) to identify purification conditions that maximize both product purity and recovery for an important class of therapeutics – oligonucleotides. Applying Alchemite™ machine learning to the project meant that goals were achieved after just 24 rounds of experiments, in contrast to the 64 expected from standard DOE.
Therapeutic oligonucleotides, including antisense oligonucleotides (ASOs) and aptamers, are short, synthetic strands of DNA or RNA used for the treatment of cancer, genetic disorders, and rare diseases. Ensuring the quality and safety of these molecules requires the efficient removal of closely related impurities. Oligonucleotide structures can be complex, making development of a robust purification process challenging and resource-intensive.
Anion exchange chromatography is frequently used as an initial purification step for oligonucleotides. However, identifying optimal process conditions, including resin selection, buffer composition, pH, elution salt type, flow rate, and temperature, often requires extensive experimental screening and significant material consumption. At an Intellegens webinar, Dr Elena Kumm, Application Scientist at Tosoh Bioscience, global leaders in liquid chromatography, shared experience of applying machine learning to accelerate identification of the best purification conditions for single stranded modified and unmodified 20-mer oligonucleotides.
The project used a range of AEX Q-type resins in the same StillPak 5 column. With five different process conditions to optimize, fully exploring the experimental design space using standard DOE approaches would normally require around 64 experiments.
Instead, the team first used the Alchemite™ software to randomly select ten experiments. Four key outputs were measured – purity, recovery, concentration of elution buffer, and fraction volume. Maximizing purity and recovery was the target for the project. The Alchemite™ software then recommended what ten experiments to do next. This time, the recommendation was not random but used the Alchemite™ machine learning method to learn from the results of the first round of experiments and decide what tests would be best to do next, given the target outputs. This approach, with iterative rounds of experimentation learning from previous rounds, is called Adaptive Design of Experiments.
In Adaptive DOE, further rounds of experiment are continued until the targets are met. This typically happens much sooner than would be the case with standard DOE methods, and that was the case in this project. The desired purity and recovery was achieved during the third round of experiments, after just 24 tests, a 60%+ improvement on standard DOE.
Elena Kumm also pointed out the value of Alchemite™ in giving you predictions for the outputs of experiments before they are run. This means you can prioritize experiments and even tweak inputs before running to try to hit targets even faster. In this project, predictions were shown to be a good match to final experimental results.
Not only did the team find optimal conditions for their purification processes, they were able to conclude that use of Alchemite™ could lower dependence on resin type when performing Anion exchange chromatography, since it makes it much easier to optimize conditions for a specific resin.
Elena Kumm concluded with a comment on the project process: “Alchemite was new to us. We had no experience of this software, but Intellegens was always there for us. While we were running these experiments we always had meetings in between where we had questions… they were always there, helping and supporting us. So don’t be scared if it’s new software – you will get support!”
