Pfizer—global biopharmaceutical leader advancing innovative medicines and vaccines. This project focused on kinetic studies for a hydrogenation reaction during process development.
Traditional reaction modeling and optimization through kinetic modeling is typically time-consuming, expensive, and resource-heavy. However, kinetic models provide scientists with key insights about reaction behavior across variable conditions. The modeling team’s time and materials were being consumed at a rate that slowed other priorities. Therefore, they evaluated SuntheticsML as a reliable approach to accurately reproduce the insights provided by the kinetic model, including the coordinates of optimal reaction conditions. The team selected this hydrogenation as a case study given that the successful kinetic model had been previously validated by the team with only 15 experiments, which would act as a benchmark for this study.
Demonstrate SuntheticsML capabilities in the prediction of optimal reaction conditions (maximized yield and minimized by products) with comparable accuracy using only 15-20 experiments.
SuntheticsML deployed its small-data modeling framework powered by Active Learning and Bayesian Optimization. The platform ingested the initial dataset, generated model-based predictions across the design space, recommended the next best experiments, and iterated—accelerating discovery while conserving runs (Upload → Predict → Select Experiments → Iterate).
● Yield: Optimized to ~97.5%, matching a highly trusted prediction that matched the optimal found by Pfizer’s calibrated kinetic model
● Impurities: Reduced in comparison to the optimal in the kinetic model optimization
● Experiments: 15 → 14 (7% reduction compared to a highly efficient kinetic model)
● Operating ranges identified: ramp rate, temperature, hold time
● Search efficiency: ML-guided suggestions converged on the optimal operating window with minimal experimentation
● Model robustness: SuntheticsML’s model proved accurate in the prediction of optimal conditions even with an added 5% error noise in the training data.
SuntheticsML further recommended optimal ranges for ramp rate, temperature, and hold period, even with additional tests for added system error and key training data points removed. Sunthetics located the ideal testing point in a shortened process that would normally have been hinderingly long through traditional methods.
SuntheticsML could let our Process Modeling group focus on other high-impact work while giving bench chemists an intuitive web platform. The model’s suggested point led directly to a ~97.5% yield—faster, and with fewer experiments than our standard kinetic modeling process.
● 7%fewer experiments to reach optimal conditions, with lower impurities
● Intuitive platform for bench chemists; lighter workload for modeling teams
● Bayesian Optimization navigates complex kinetic spaces efficiently
● Enables lean, agile biopharma R&D—especially valuable when resources are constrained or experiment costs are high
