The enzyme hunt is learning from the reactions that failed

For four previously untested pairings of known chemical ingredients, researchers in Graz used machine learning to rank enzymes for the bench. Each 15-enzyme shortlist contained successful catalysts. Five enzymes from the bottom of each ranking showed little or no activity.[1]

The result grew out of a much larger laboratory effort. The team examined 175 enzymes in nearly 28,000 individual reactions, collecting data over two years. Fewer than a quarter of the reactions initially gave positive results. The University of Graz highlighted the work on 12 October, following its August publication in Nature Communications.[2]

Choosing the next experiment is a substantial job in its own right. As an illustrative calculation, ten possible ingredients on one side of a reaction, ten on the other and 100 candidate enzymes create 10,000 combinations. A chemist needs a way to search that space without running every reaction.

An enzyme for the particular ingredients

The enzyme family is called imine reductases, or IREDs. Its members can help make amines through reductive amination, joining a carbonyl compound with an amine partner and reducing the intermediate. Work on the earlier IREDFisher screening tool explains why selecting a small panel can be useful: a protein’s structure helps determine whether a particular substrate can fit and react in its active site.[3]

IREDFisher used structural models, molecular docking and a scoring function to choose small enzyme panels for laboratory testing. That approach makes the candidate list more manageable.[3]

Graz adds extensive measured activity data to the search: a model can learn from combinations that were actually attempted, including those that did not produce a useful reaction.[1]

The practical ambition also concerns how much material a reaction needs. Graz’s programme sought enzymes that work with equal numbers of molecules of the two starting ingredients, reducing reliance on a large excess of one partner. Its academic researchers worked with Roche and enzyme specialist Innophore. The university reports that the later predictive campaign took four weeks, after the much longer effort to assemble the initial reaction data.[2]

The failed experiment is part of the map

Negative results have a scientific use beyond this enzyme family. In a 2016 Nature study, researchers drew on unsuccessful reactions recorded in laboratory notebooks to train a model of crystal formation. They combined reaction histories with descriptions of the ingredients to predict conditions under which templated vanadium selenites would crystallise. The experiments that failed helped describe the territory in which successful reactions could occur.[4]

A list of winners answers one question: what has worked? A collection of attempted combinations can answer a second: how does success change when the ingredients or conditions change? The absence of a useful product is informative when the experiment and its conditions have been recorded. It can help distinguish an unsuccessful combination from one nobody tried.

In Graz’s dataset, 77% of tested enzyme–substrate combinations were non-hits. The authors released the activity data in machine-readable form, including this unsuccessful majority.[1]

Three different meanings of a good result

A catalyst also needs to produce the desired molecular form. Chiral molecules have mirror-image forms that cannot be placed exactly on top of one another. The usual comparison is a pair of hands: the parts correspond, but their arrangement differs. In a synthesis, identifying the molecular connections alone therefore leaves a further question about the three-dimensional product.[5]

Enantiomeric excess measures the difference between the proportions of the two mirror-image forms. Using IUPAC’s definition, 98% excess means 99% of one form and 1% of the other: a 99-to-1 ratio, calculated from those two proportions. This describes the balance between enantiomers; it does not say that every other chemical impurity is absent.[6]

The table follows a reaction through increasingly demanding questions. Has the starting material reacted? How much desired product can be collected? Has the chemistry favoured the required handedness? An enzyme shortlist earns its value by helping the laboratory reach those answers with fewer candidates to test.

Report an error or suggest a correction