Yoneda Optimize - Benchmarks

OPTIMIZE REACTIONS 3x FASTER

Introduction

We developed Yoneda Optimize, a software tool that helps chemists optimize their reactions faster and cheaper. Yoneda Optimize consistently gets within 3% of the best possible yield, and helps chemists optimize reactions in days instead of weeks.

Direct Arylation

Motivated by the synthesis of the JAK2 inhibitor BMS-911543, the Doyle group[1] used the following Pd-catalyzed arylation of imidazole to benchmark performance of computer guided (Bayesian) optimizations against human performance.

Although there are over 1700 possible reaction conditions, Bayesian optimization can repeatedly reach 100% yield using only 40 experiments, that is much faster than chemists who often need 50 - 60 reactions.

Maximum yields achieved using Yoneda Optimize, dashed lines are individual runs, average is in bold. See similar figures for Bayesian optimization and human performance published in 10.1038/s41586-021-03213-y.

Bayesian optimization is an iterative and lengthy process. If each batch has only 5 reactions as published by Doyle[1], finding the optimal conditions would take 8 batches or 2 weeks.

Yoneda Optimize uses only 2 or 3 batches - first 20 experiments are designed to understand the reactions space, and only one or two additional iterations are needed to find the optimal conditions. This results in 3× speedup!

Experiment
Design		 Reactions Time Spent
in Lab		 Achieved Yield
(averaged)
Yoneda
Optimize		 20+10 3 days 98% (N=30)
20+10+10 4 days 100% (N=30)
Simple BO 6×5 1.5 weeks 98% (N=30)
8×5 2 weeks 100% (N=30)
Human
Expert		 6×5 1.5 weeks 93% (N=50)
8×5 2 weeks 93% (N=50)

Suzuki cross-coupling

In Suzuki cross-coupling, it is often hard to decide whether to use more expensive aryl iodides and bromides, or cheaper but less reactive aryl chlorides, or in some cases more synthetically accessible triflates.

The performance of Yoneda Optimize for Suzuki cross-coupling was benchmarked on dataset of nearly 4000 reactions published by Pfizer[2].

Yoneda Optimize lets you find nearly optimal yield within two or three screenings, so that you can decide which substrate is most suitable for your synthesis.

Electrophile Yield after specified reaction batches (average, N = 30) Best possible yield
After 2 batches
(20+10 reactions)		 After 3 batches
(20+10+10 reactions)
ArOTf 97% 97% 99%
ArCl 88% 91% 94%
ArBr 95% 96% 98%
ArI 99% 100% 100%
Average (N=30) achieved yield after two or three optimization iterations.

Buchwald-Hartwig amination

Predicting suitable reactants is especially hard if chemists have limited experience with them. For example, the following dataset published by the Doyle group[3] performs Buchwald-Hartwig amination with 22 possible isoxazole additives.

Depending on the nature of the aryl halide, it might be hard to find good reaction conditions, and sometimes achieving high yield is not even possible (within the parameter space). Yoneda Optimize efficiently guides the reaction screening to get within a few percent of the best possible yield after only 2 or 3 batches.

Aryl Halide Yield after specified reaction batches (average, N = 30) Best possible yield
After 2 batches
(20+10 reactions)		 After 3 batches
(20+10+10 reactions)
o-pyridine Cl 90% 91% 91%
Br 94% 96% 97%
I 98% 99% 100%
m-pyridine Cl 64% 67% 69%
Br 93% 95% 99%
I 94% 95% 98%
p-CF3-benzene Cl 40% 43% 45%
Br 49% 52% 54%
I 53% 55% 56%
p-ethylbenzene Cl 14% 14% 16%
Br 83% 84% 87%
I 85% 86% 86%
p-anisole Cl 42% 44% 44%
Br 66% 67% 68%
I 64% 68% 68%
Average (N=30) achieved yield after two or three optimization iterations.

Structures of Ligands and Additives

References

  1. Benjamin J. Shields, Jason M. Stevens, Jun Li, Marvin Parasram, Farhan Damani, Jesus I. Martinez Alvarado, Jacob M. Janey, Ryan P. Adams, Abigail G. Doyle (2021). "Bayesian reaction optimization as a tool for chemical synthesis". Nature, 590(7844), 89–96. DOI: 10.1038/s41586-021-03213-y.
  2. Damith Perera, Joseph W. Tucker, Shalini Brahmbhatt, Christopher J. Helal, Ashley Chong, William Farrell, Paul Richardson, Neal W. Sach (2018). "A platform for automated nanomole-scale reaction screening and micromole-scale synthesis in flow". Science, 359(6374), 429–434. DOI: 10.1126/science.aap9112.
  3. Derek T. Ahneman, Jesús G. Estrada, Shishi Lin, Spencer D. Dreher, Abigail G. Doyle (2018). "Predicting reaction performance in C–N cross-coupling using machine learning". Science, 360(6385), 186–190. DOI: 10.1126/science.aar5169.