Comparative study of various molecular representation for intermolecular interaction predictions

In this project we compare a number of molecular representations(vectorizers) to determine what is the most suitable way to represent a molecule as vector when intermolecular interactions are at most interest. In this study we use solvation energy as a target value and solvent and solute molecules as input. The data is obtained from MNSol Database.

To MSU AI

Please read the following file for reproducibility, models availability and other comments

Training

The training data is written to Runs folder and the results are stored in Run_results (due to large file sizes Run_results is available for manual download from Yandex Disk) including losses plot, normalization parameters, run_log and comments. The links to each result folder are presented below

Neural Networks

All training files are presented in Training_files in the format Solvent_Solute_NN. Examples of getting best val models for LinNet and ResNet are below

# LinNet

from my_nets.LinearNet import LinearNet3
from my_nets.net_func import load_ckp
import torch

in_feat = 207  # specify length 
model = LinearNet3(in_features=in_feat)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
best_model, *other = load_ckp('Examples/data/best_Class_Morgan_Lin1', model, optimizer)

# ResNet

from my_nets.ResNET import ResNet1D
from my_nets.net_func import load_ckp
import torch

Res_Dict = {'base_filters':2, 'kernel_size':3, 'stride':2, 'groups':1, 'n_block':3, 'n_classes':1, 'use_bn':True, 'use_do':True, 'verbose':False}
model = ResNet1D(in_channels=1, **Res_Dict)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
best_model, *other = load_ckp('Run_Results/ResNet/Class_Morgan_2_124_Res1/best/best_val_model.pt', model, optimizer)

Kernel Ridge Regression

KRR_training - all KRR experiments are sequentially carried out in this file. The results are available at Google Drive.

# KRR

import pickle as pkl

with open(project_path('/Run_results/KRR/Class_Morgan_KRR1/best_models.pkl'), 'rb') as f:
    KRR_Class_Morgan_kernels = pkl.load(f)  # dictionary of kernel names and the models
KRR_Class_Morgan = KRR_Class_Morgan_kernels['laplacian']

Other experiments

Experiments on another datasets (Acree and FreeSolv), G_solv distribution and feature permutation importance are presented in following Jupyter Notebook.

Repository structure

config.py

A file with some useful function used along all the project.

Training_files

A folder with .py files each of which trains the network with some parameters. All KRR training is in one file.

my_nets

A package with some .py files to create and train networks

    Create_dataset - A file that contains functions to create dataset using given vectorizers

    net_func - A file that contains functions train network and other useful functions

    LinearNet - A file that contains Linear Network used for training

    ResNET- A file that contains 1D ResNET used for training. The model is adopted from hsd1503

Vectorizers

A package vectorizers.py that contains vectorizers functions used in this project

Tables

A folder with tables used for various functions and vectorizers

Preprocess

A folder with some files used to prepare data (tables, dicts, ...)

Vectorizers

Blank

zero tensor with length one to train models without any information either on solvent or solute.

Classification

Three layer classification, described in MNSol Database.

Solute_TESA

taken from MNSol database calculated parameter of Total Exposed Surface Area. More info in MNSol Database.

Solvent_Macro_props

properties of solvent: nD, alpha, beta, gamma, epsilon, phi, psi. Sometimes called Abraham descriptors.

MorganFingerprints

calculated morgan fingerprints bit vector, described here

If troubles with installation try

pip install rdkit-pypi

BoB

Bag of Bonds.

scipy install problems solved here: https://stackoverflow.com/a/69710042/13835675

JustBonds (JB)

Bag of Bonds for bonded atoms only

BAT

Bag of Bonds with addition of Angles and Torsion angles between bonded atoms

SOAP

Smooth Overlap of Atomic Positions, thoroughly described here

Table with Solvent-Solute Experiment links

Kernel Ridge Regression

Solvent➡️ ⬇️Solute	Blank	Class	Macro	Morgan	JustBonds	BoB	BAT	SOAP
Blank		Class Blank	Macro Blank	Morgan Blank	JB Blank	BoB Blank	BAT_Blank	SOAP_Blank
Class	Blank Class	Class Class	Macro Class	Morgan Class	JB Class	BoB Class	BAT_Class	SOAP_Class
TESA	Blank TESA	Class TESA	Macro TESA	Morgan TESA	JB TESA	BoB TESA	BAT_TESA	SOAP_TESA
Morgan	Blank Morgan	Class Morgan	Macro Morgan	Morgan Morgan	JB Morgan	BoB Morgan	BAT_Morgan	SOAP_Morgan
JustBonds	Blank JB	Class JB	Macro JB	Morgan JB	JB JB	BoB JB	BAT JB	SOAP JB
BoB	Blank BoB	Class BoB	Macro BoB	Morgan BoB	JB BoB	BoB BoB	BAT_BoB	SOAP_BoB
BAT	Blank BAT	Class_BAT	Macro_BAT	Morgan_BAT	JB BAT	BoB_BAT	BAT_BAT	SOAP_BAT
SOAP	Blank SOAP	Class_SOAP	Macro_SOAP	Morgan_SOAP	JB SOAP	BoB_SOAP	BAT_SOAP	SOAP_SOAP

Linear

Solvent➡️ ⬇️Solute	Blank	Class	Macro	Morgan	JustBonds	BoB	BAT	SOAP
Blank		Class Blank	Macro Blank	Morgan Blank	JB Blank	BoB Blank	BAT_Blank	SOAP_Blank
Class	Blank Class	Class Class	Macro Class	Morgan Class	JB Class	BoB Class	BAT_Class	SOAP_Class
TESA	Blank TESA	Class TESA	Macro TESA	Morgan TESA	JB TESA	BoB TESA	BAT_TESA	SOAP_TESA
Morgan	Blank Morgan	Class Morgan	Macro Morgan	Morgan Morgan	JB Morgan	BoB Morgan	BAT_Morgan	SOAP_Morgan
JustBonds	Blank JB	Class JB	Macro JB	Morgan JB	JB JB	BoB JB	BAT JB	SOAP JB
BoB	Blank BoB	Class BoB	Macro BoB	Morgan BoB	JB BoB	BoB BoB	BAT_BoB	SOAP_BoB
BAT	Blank BAT	Class_BAT	Macro_BAT	Morgan_BAT	JB BAT	BoB_BAT	BAT_BAT	SOAP_BAT
SOAP	Blank SOAP	Class_SOAP	Macro_SOAP	Morgan_SOAP	JB SOAP	BoB_SOAP	BAT_SOAP	SOAP_SOAP

ResNET

Solvent➡️ ⬇️Solute	Blank	Class	Macro	Morgan	JustBonds	BoB	BAT	SOAP
Blank		Class Blank	Macro Blank	Morgan Blank	JB Blank	BoB Blank	BAT Blank	SOAP Blank
Class	Blank Class	Class Class	Macro Class	Morgan Class	JB Class	BoB Class	BAT Class	SOAP Class
TESA	Blank TESA	Class TESA	Macro TESA	Morgan TESA	JB TESA	BoB TESA	BAT TESA	SOAP TESA
Morgan	Blank Morgan	Class Morgan	Macro Morgan	Morgan Morgan	JB Morgan	BoB Morgan	BAT Morgan	SOAP Morgan
JustBonds	Blank JB	Class JB	Macro JB	Morgan JB	JB JB	BoB JB	BAT JB	SOAP JB
BoB	Blank BoB	Class BoB	Macro BoB	Morgan BoB	JB BoB	BoB BoB	BAT BoB	SOAP BoB
BAT	Blank BAT	Class BAT	Macro_BAT	Morgan_BAT	JB BAT	BoB BAT	BAT BAT	SOAP BAT
SOAP	Blank SOAP	Class SOAP	Macro_SOAP	Morgan_SOAP	JB SOAP	BoB SOAP	BAT SOAP	SOAP SOAP

The End