jixing475 / molfinder

MolFinder

an evolutionary algorithm for the global optimization of molecular properties and the extensive exploration of chemical space using SMILES

MolFinder finds diverse molecules with desired properties efficiently without any training and a large molecular database. Also, it does not require a lot of computing resources. This repository contains MolFinder's code and results. MolFinder consists of a simple executable file, molfinder, and crossover and mutation process file ModSMI.py.

Getting Started

Prerequisites

• python == 3.7.*
• rdkit >= 2019.09.3.0+
• pandas
• numpy
• matplotlib

Installation instructions

We recommand to use anaconda3 virtual-env; simplest way

conda create -n molfinder_venv python=3.7
conda activate molfinder_venv
conda install -c rdkit rdkit
conda install pandas numpy

Quickstart

0. Prepare dataset

Prepare a SMILES file in CSV format. The first column must be SMILES.
(Contains headers, It doesn't have to be just only SMILES.)

SMILES
CCN(CC)CCN(C(=O)c1ccc(CCC(F)(F)F)cc1)[C@H]1CCS(=O)(=O)C1
C[C@@H](C(=O)N(C)C)N1[C@H]2CC[C@H]1CC(NC(=O)C1C(C)(C)C1(C)C)C2
CCOc1cc(N2C[C@@H]3C(NC(=O)c4ccn(C)n4)[C@H]3C2)ncn1
O=C(NC[C@@H](CO)Cc1cccnc1)c1ccnc(OC2CCC2)c1
O=C([C@@H]1C[C@H]1c1cccnc1)N1CCC(O)(CNCc2ccccn2)CC1
O=C(CCN1C(=O)[C@H]2CCCC[C@@H]2C1=O)NC1CCN(CC(F)(F)F)CC1
...

1. Run MolFinder algorithm

./molfinder -i sample.csv --max-round 5

Parameters of MolFinder

• -i, --input: (str) SMILES file (csv format) used by the model

• -r, --random-seed: (int, None) Determines the random number that selects the initial molecules

• --bank-size: (int, 100) Bank size used in the algorithm

• --seed-size: (int, 60) The number of parent molecules used to generate child molecules

• -dist, --dist-coef: (float, 0.90) Adjust the $D_{avg}$ value

• --max-round: (int, 150) The maximum number of round

• -cvg, --convergent-round: (int, 150) Determines how many rounds the Dcut will converge

• -c, --coefficient: (float, 0.9) Coefficient of objective function

• --target: (SMILES: str, None) Target molecule

• -fp, --fp-method: (str, rdkit) Fingerprint method; Morgan or RDKit (default)

• -v, --verbosity: Print RDKit Error message.

# Parameters of paper results
molfinder -r 12345678 --bank-size 1000 --seed-size 600 -dist 0.90 -c 0.994 -i sample.csv

Set objective fucntion

1. Find @@FEATURES in molfinder, Set your features.
2. Find @@REWARD in molfinder, Modifiy your objective function.

Component of MolFinder

MolFinder Algorithm
├── molfinder
│   ├── cal_avg_dist
│   ├── prepare_seed
│   ├── prepare_child
|   │   └── Crossover and Mutations from ModSMI
│   └── update_bank
└── ModSMI.py
    ├── tight_rm_branch
    ├── prepare_rigid_crossover
    ├── replace_atom
    ├── add_atom
    └── delete_atom

References

• paper: https://doi.org/10.1186/s13321-021-00501-7

---

This is my first code and was mainly written in February 2020. There are many drawbacks, but I keep learning and trying.