Fig. 2: Overview of CatPred’s deep learning architecture and prediction interface.

a The three different modalities with increasing level of detail explored for Enzyme feature learning. The Sequence-Attention (Seq-Att) module learns features of amino-acid embeddings using multi-head attention layers. The pLM module uses features extracted from a pre-trained protein Language Model (pLM). The Equivariant Graph Neural Network (E-GNN) module extracts features of 3 d structures of enzymes by employing equivariant graph neural networks on their amino-acid level graphs. b Substrate feature learning is carried out using Directed Message Passing Neural Networks (D-MPNN) that extract molecular representations by leveraging 2D atom-bond connectivity graphs. c CatPred models are trained on CatPred-DB datasets utilizing both substrate and enzyme feature learning modules with a probabilistic regression approach. The enzyme and substrate features are input to a fully connected neural network that predicts the kinetic parameters as outputs in the form of Gaussian distributions characterized by their respective means (μ) and variances (σ²). d CatPred production models are made available through the Google-Colab interface for ease of access. The inputs are the substrate SMILES and either enzyme sequence or structure along with a choice of kinetic parameter for prediction. The interface then loads the respective trained models and outputs uncertainty quantified kinetic parameters in terms of a predicted mean and standard deviation (SD).