Rossmann-toolbox: a deep learning-based protocol for the prediction and design of cofactor specificity in Rossmann-fold proteins

Abstract

The Rossmann fold enzymes are involved in essential biochemical pathways such as nucleotide and amino acid metabolism. Their functioning relies on interaction with cofactors, small nucleoside-based compounds specifically recognized by a conserved βαβ motif shared by all Rossmann fold proteins. While Rossmann methyltransferases and enzymes involved in the polyamine synthesis recognize only a single cofactor type, the S-Adenosylmethionine (SAM), the oxidoreductases, depending on the family, bind nicotinamide (NAD, NADP) or flavin-based (FAD) cofactors. In this study, we show that despite its short length, the βαβ motif unambiguously defines the specificity towards the cofactor. Following this observation, we trained two complementary deep learning models for the prediction of the cofactor specificity based on the features of the βαβ motif. The first utilizes contextualized sequence embeddings, whereas the second relies on structures represented as graphs. A benchmark on two test sets, one containing βαβ motifs bearing no resemblance to those of the training set, and the other comprising 38 cases of the experimentally confirmed redesign of the cofactor specificity from NAD to NADP and vice versa, revealed nearly-perfect performance (~95% accuracy) of the two methods. Finally, by combining the two approaches, we built a pipeline for the design of cofactor-switching mutations. Both prediction methods can be accessed via the webserver https://lbs.cent.uw.edu.pl/rossmann-toolbox and are available as a Python package at https://github.com/labstructbioinf/rossmann-toolbox.

Predicting enzyme substrate chemical structure with protein language models

“…In addition to catalyzing a wide variety of redox reactions, some SDRs can also act as epimerases and lyases 43 . While the task of predicting redox cofactor preference from sequence is well-established in the literature [46][47][48] , we reasoned it would serve as an initial, baseline task to benchmark our approach.…”

Learning deep representations of enzyme thermal adaptation

“…As the field of ML interpretation expands in both theory and software availability, different avenues of feature identification may be pursued besides our black-box perturbation approach. A promising example is the integrated gradients method 55 (available currently only in the PyTorch-based package captum), as illustrated in the recent study by Kaminki et al 56 Given these recapitulations of known primary and secondary protein structure determinants of thermostability by DeepET's features, which were learned by the model from sequence alone, and the observed shift in model-relevant domains between mesophiles and thermophiles, the use of the DeepET is a promising avenue towards elucidating the physical mechanisms that convey enzymes resistance to extreme temperatures. Future work will therefore focus on further interpreting DeepET and its learned representations both using in silico analyses, and in a biological context, to deepen our understanding of enzyme thermal adaptation.…”

“…As the field of ML interpretation expands in both theory and software availability, different avenues of feature identification may be pursued besides our black‐box perturbation approach. A promising example is the integrated gradients method 55 (available currently only in the PyTorch‐based package captum ), as illustrated in the recent study by Kaminki et al 56 …”

RCsearcher: Reaction Center Identification in Retrosynthesis via Deep Q-Learning

“…G2G (Shi et al, 2020) and RetroXpert (Yan et al, 2020) regards the bond that appears in the product but not in the reactants as the reaction center. They use R-GCN (Schlichtkrull et al, 2018) and EGAT (Kamiński et al, 2022) respectively to predict the probability of each bond as a single reaction center. GRAPHRETRO (Somnath et al, 2021) also considers the atom whose number of attached hydrogens changes in the product as a single reaction center.…”

“…To the best of our knowledge, the only existing method that can simultaneously address single and multiple reaction center identification task appear in RetroXpert (Yan et al, 2020). It uses one shared EGAT (Kamiński et al, 2022) to output two tensors (Fig. 3).…”