RetKcat: A novel neural network for Kcat prediction

How to start

Requirements

freeze.yml and requirements.txt were provided in the folder "RetKcat", you can install the packages by running the following command in the folder you want to install RetKcat. The model that was trained on an RTX 3090 costs an average of 19 Gib memory and also 16 Gib on an RTX 4090 while predicting is available on an RTX 3070Ti with 6Gib on average via pre-trained model state provided by us. A large amount of evidence shows that this model will have a better performance with a higher hidden dimension, which is 64 in our work due to the limit of GPU memory.

Begin

first of all, run process.py in the child folder bin

We recommend conda to construct an environment (may need to install some packages which were provided by others)

conda env create -f freeze.yml -n RetKcat
conda activate RetKcat

or try pip

pip install -r requirements.txt

try to download the code with git or zip.

git clone https://github.com/CPU-CHINA/RetKcat.git

and run the code in the folder RetKcat,

cd RetKcat

put any sample you want to predict in the RetKcat/input.json and allow the format below

    {   
        "id": "['WT-4-HPAA']",
        "Smiles": "Oc1ccc(cc1)C/C=N/CCc2cc(O)c(O)cc2",
        "Sequence": "MMKMEVVFVFLMLLGTINCQKLILTGRPFLHHQGIINQVSTVTKVIHHELEVAASADDIWTVYSWPGLAKHLPDLLPGAFEKLEIIGDGGVGTILDMTFVPGEFPHEYKEKFILVDNEHRLKKVQMIEGGYLDLGVTYYMDTIHVVPTGKDSCVIKSSTEYHVKPEFVKIVEPLITTGPLAAMADAISKLVLEHKSKSNSDEIEAAIITV",
        "Value": 21.35
    }

and run predict.py, out file will appear in the current folder, a temporary file input. pkl will be created and then removed after predicting.

Introduction

The neural network can be divided into two parts: The first part utilizes a retentive network (RetNet) to extract protein features. This is achieved through a combination of causal masking and exponential decay along relative distances, which are combined into a single matrix. The second part employs graph convolutional networks (GCN) to capture substrate characteristics.

(A) RetKcat learning performance analysis. The trained model is tested on the training set, and R-square is used to measure whether the model has correctly learned the training set. (B) NCS samples prediction test. On the test set derived from the experiment, by comparing RetKcat with the currently better DLKcat. (C) RetKcat schematic diagram. RetKcat is composed of two parts, GCN is used to read molecular information, and RetNet is used to read protein information.

Construction of RetKcat

In this work, we developed an end-to-end learning approach for in vitro Kcat value prediction by combining a GCN for substrates and a RetNet for proteins. Molecular structures which are atoms linked with chemical bonds can be naturally converted into a graph and protein sequences can also be seen as a special format of the list.

First, substrate SMILES information was loaded with RDKit v.2022.9.5 (https://www.rdkit.org) and then each node will update itself via its neighbor around, which can be seen as dividing atoms with its chemical environment. Moreover, the adjacency of the molecule was extracted, and the molecule was finally represented as adjacency and an ordered node list. Then the edge information and node information have been convoluted. The final output of the GCN is a real-valued matrix M.

The protein sequence is manually split into ‘words’ which contain N amino acids. Every word is corresponding with a real number. Windows was set to limit the length of the word list, every N amino acid is transferred into numbers and held by Windows respectively. Then the matrix may be embedded to the appointed dimension. The protein representation and molecule representation will have the same dimension and will be concentrated as the input of RetNet.

The outcome of RetNet will forward an output layer, which consists of several Linear, and then the vector will be turned to predict value via a single layer Linear.

Data and Code Availability

Databases including BRENDA (https://www.brenda-enzymes.org), and SABIO-RK (http://sabiork.h-its.org/) were used in the evaluation of the DLKcat performance. Protein Data Bank in Europe database (https://www.ebi.ac.uk/pdbe/).

Source data are provided in this paper. To facilitate further usage, we provide all codes and detailed instructions in the GitHub repository (https://github.com/CPU-CHINA/RetKcat). Files and results related to simulation calculations can also be found in the GitHub repository (https://github.com/CPU-CHINA/collation).