The substrate scopes of enzymes: a general prediction model based on machine and deep learning

Abstract

For a comprehensive understanding of metabolism, it is necessary to know all potential substrates for each enzyme encoded in an organism's genome. However, for most proteins annotated as enzymes, it is unknown which primary and/or secondary reactions they catalyze, as experimental characterizations are time-consuming and costly. Machine learning predictions could provide an efficient alternative, but are hampered by a lack of information regarding enzyme non-substrates, as available training data comprises mainly positive examples. Here, we present ESP, a general machine learning model for the prediction of enzyme-substrate pairs, with an accuracy of over 90% on independent and diverse test data. This accuracy was achieved by representing enzymes through a modified transformer model with a trained, task-specific token, and by augmenting the positive training data by randomly sampling small molecules and assigning them as non-substrates. ESP can be applied successfully across widely different enzymes and a broad range of metabolites. It outperforms recently published models designed for individual, well-studied enzyme families, which use much more detailed input data. We implemented a user-friendly web server to predict the substrate scope of arbitrary enzymes, which may support not only basic science, but also the development of pharmaceuticals and bioengineering processes.

Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning

“…In our study on predicting the substrate scope of enzymes 30 , we found that prediction performance depends strongly on the sequence similarity between a target enzyme and enzymes in the training set, consistent with the widely held belief that enzymes are more likely to be functionally similar if they have more similar sequences [34]. We hence examined the performance of TurNuP for enzyme sets that differed in their maximal similarity to proteins in the training set.…”

“…In a previous project 30 , we created a fine-tuned and task-specific version of the ESM-1b model that led to improved predictions for the substrate scope of enzymes, a problem for which abundant training data exists. Such comprehensive data is required to re-train the ESM-1b model, but is not available for k cat , and we were thus unable to create a version specific to the task of predicting k cat .…”

“…However, we speculated that the ESM-1b vectors fine-tuned previously for the prediction of enzyme-substrate pairs might also improve k cat predictions. To test this hypothesis, we used our previously published model, ESP 30 , to calculate fine-tuned representations for all enzymes in our dataset. In the following, we will refer to these representations as ESM-1b ESP vectors.…”

What puzzle are you in?

“…The second example is a project that uses deep learning to predict the substrate scope of enzymes [10]. Our initial idea was to simply modify a prediction pipeline that we had developed previously for the Michaelis constants of enzymes, K m [11].…”

“…We indeed found a solution: we added another "layer" to an existing numerical representation [12] of amino acid sequences. Training this model on our enzyme data, we could teach it to produce output that was more informative for our predictions [10].…”