AI gives chemists a shorter enzyme list for new reactions

A useful enzyme is useful for a particular job. Change either starting molecule in a chemical reaction and the best catalyst may change with it. For a chemist planning the next experiment, the valuable AI output can therefore be quite modest: a defensible shortlist.

A University of Graz report on 12 October describes just such a result from work with Roche and Innophore. The team investigated 175 imine reductases, or IREDs, and used the reaction data to nominate catalysts for new starting-material combinations. The underlying Nature Communications paper appeared on 22 August.[3][1]

These enzymes help make chiral amines, nitrogen-containing compounds that can have two mirror-image forms and serve as pharmaceutical building blocks. Selecting which form a reaction produces is part of the manufacturing task.[3]

Predicting the match between catalyst and ingredients

The prospective model combined chemical descriptions of the starting molecules with learned representations of enzyme cavities—the pockets where ingredients bind—and earlier activity measurements. A predictor combining decision trees turned those inputs into rankings. For each of four new substrate pairs, the researchers ranked 100 enzymes and tested the top 15 plus the bottom five.[2]

That design asks a useful question beyond whether a model can recognise a familiar success. The two ingredients were being brought together in combinations absent from the experimental screening. Testing the bottom of each list also checked whether the ranking separated promising candidates from poor ones.

All four shortlists contained active enzymes; the bottom groups produced no product or only traces. For one pair, enzyme IR2147 converted 37% of the starting carbonyl into the target product, with more than 99% enantiomeric excess. The tests used 0.5-millilitre reactions over 24 hours at 30°C.[2]

Enantiomeric excess describes the balance between the mirror-image products. A value of 99% corresponds to 99.5% of one form and 0.5% of the other, assuming those are the two forms present. Conversion answers a separate question: how much starting material became the product. Strong selectivity and incomplete conversion can coexist.[4]

The new-pair experiments reached at most 37% conversion. Separately, five other selected reactions were run at 200 millilitres, giving isolated yields of 63–89%. Those larger batches demonstrate the library’s synthetic usefulness; they are not scale-up results for the four newly predicted combinations.[1]

A library that helps choose the next experiment

Graz reports that the four target molecules were made within four weeks after prediction. That is a reported research sequence, not a controlled estimate of an industry-wide speed-up.[3]

The deposited data includes consolidated screening results, enzyme sequences and chemical descriptors in reusable files.[5]

This creates a practical division of labour. A model can prioritise where to look; the laboratory establishes whether the reaction works and which product it makes. An imperfect but selective catalyst is already a more specific starting point for optimisation than an unranked list. The durable result is a collection of experiments that becomes useful again when the chemist changes the question.

Report an error or suggest a correction