Lyrebird: 3D Molecular Conformer Generation

This repository contains the Lyrebird neural network for 3D conformer generation, developed by Rowan based on the ETFlow architecture, and a Butina cluster-based train/val/test split of the GEOM dataset and CREMP.

This work was completed by Vedanth Nilabh during an internship at Rowan. For questions or issues, please open a GitHub issue or contact nilabh.v@northeastern.edu.

Overview

Lyrebird is trained on diverse molecular geometries from the GEOM datasets (DRUGS and QM9), as well as CREMP, and generates 3D molecular conformers from SMILES strings using an equivariant flow-based generative model. We also have detailed energy benchmark results (which measures energy differences pre and post optimization with groynd truth and optimizes structures using the semi-empirical method GNF2-XTB) in the associated folder, along with a detailed description of the motivation and methodology.

For the newly released MPCONF-196-GEN bechmark see: MPCONF-196-GEN-benchmark

Example Usage

import torch
from lyrebird import LyrebirdCalculator

# Initialize the calculator
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
calculator = LyrebirdCalculator("lyrebird.yaml", "lyrebird.ckpt", device=device)

# Generate conformers from SMILES
smiles = "CC(C)CC1=CC=C(C=C1)C(C)C(=O)O"  # ibuprofen
conformers = calculator.predict(smiles, num_initial_confs=5)

print(f"Generated {len(conformers)} conformers")
print(f"First conformer shape: {conformers[0].shape}")  # (N_atoms, 3)

Local Usage

Install the required packages using:

conda env create -f environment.yml
conda activate lyrebird-env

To run the example script:

python example.py

This model can run on either CPU or GPU. GPU is recommended for faster generation of multiple conformers.

License

This model is released under the MIT License. See LICENSE for details.