Large datasets and machine learning models fail to capture extremophile enzyme melting and optimum temperatures

Abstract

Organisms and their enzymes adapt to environmental temperatures, such that thermophilic enzymes exhibit high melting and optimum temperatures while psychrophilic enzymes exhibit low values for both. It has been proposed that the gap between an enzymes optimum temperature and its melting temperature, the temperature gap, is characteristically large in psychrophiles, implying that the loss of activity above the optimum is decoupled from global protein stability. The evidence for this relies on a small number of characterised enzymes, leaving the prevalence of large temperature gaps amongst psychrophiles unknown. We asked whether the machine-learning predictors and large datasets now available could test this at scale. We find that they cannot: predictors of melting and optimum temperature fail systematically at the thermal extremes, assigning the majority of thermophilic enzymes with optimum temperatures that exceed their melting temperatures, which is biophysically implausible, and consistently underpredict the stability of (hyper)thermophiles. This stems from training data that is both error-laden, as we demonstrate for widely used optimum-temperature records, and overwhelmingly biased toward mesophiles, which regresses predictions for cold and heat adapted enzymes toward mesophilic values. Consequently, current computational tools cannot establish how prevalent the psychrophilic temperature gap is. We argue that proteome-scale measurement of extremophile enzyme thermal behaviour, integrated as curated training data, is required to determine whether trends from small studies extend across the diversity of life.