# Which Reaction Conditions Work on Drug-Like Molecules? Lessons from 66,000 High-Throughput Experiments

## TL;DR

This paper uses 66,000 experiments to find better chemical reaction conditions for drug-like molecules using statistical analysis.

## Contribution

A robust z-score-based statistical method is introduced to analyze high-throughput reaction data and derive optimal conditions.

## Key findings

- Optimal reaction conditions for Buchwald–Hartwig and Suzuki–Miyaura reactions differ from traditional guidelines.
- A tool and dataset are published to enable data-driven insights for chemical optimization.
- Data-driven approaches improve the efficiency of reaction optimization campaigns.

## Abstract

High-throughput experimentation
(HTE) accelerates chemical discovery
by shortening the lead times for molecule synthesis. The choice of
initial reaction conditions directly influences the outcome and length
of any reaction optimization. But human involvement in plate design
and data analysis remains a significant cost factor and is accompanied
by biases. Therefore, making the most out of past reaction outcomes
is crucial. While advances in machine learning allow us to generate
promising reaction conditions, this approach is often not suitable
because not enough relevant reaction data are available or it is of
insufficient quality. Herein we introduce a robust statistical method
using z-scores to analyze 66,000 internal HTE reactions
on complex molecules. Additionally, we publish the underlying data
as well as a tool to analyze and draw actionable conclusions from
this data set. We exemplify the power of this method for the widely
employed Buchwald–Hartwig and Suzuki–Miyaura cross-coupling
reactions. The results reveal optimal conditions that differ significantly
from literature-based guidelines. These data-driven insights provide
high-quality starting points for optimization campaigns, improving
their overall efficiency.