schrodinger.application.matsci.flywheel.ldmol package

LDMol: Text-Conditioned Molecule Diffusion Model.

Copyright (c) Jinho Chang and Jong Chul Ye (Original LDMol) Copyright (c) Meta Platforms, Inc. and affiliates (DiT) Copyright (c) Salesforce.com, Inc. (BERT cross-attention) Copyright (c) OpenAI and contributors (Diffusion) Modified by Schrodinger, LLC.

This module implements LDMol, a latent diffusion model for text-to-molecule generation. LDMol uses a chemically informative latent space produced by a molecule autoencoder and combines it with natural language conditioning via a diffusion transformer (DiT).

Key Components:

• Text Encoder: MolT5-based ONNX encoder

• VAE Autoencoder: BERT-based encoder/decoder for SMILES

• DiT: Diffusion Transformer with cross-attention

• Diffusion: Gaussian diffusion with timestep respacing

References:

• LDMol: https://arxiv.org/abs/2405.17829

• DiT: https://arxiv.org/abs/2212.09748

• BERT/ALBEF: https://arxiv.org/abs/2107.07651

• IDDPM: https://arxiv.org/abs/2102.09672

Subpackages

• schrodinger.application.matsci.flywheel.ldmol.diffusion package
  • Submodules
    • schrodinger.application.matsci.flywheel.ldmol.diffusion.gaussian_diffusion module
    • schrodinger.application.matsci.flywheel.ldmol.diffusion.respace module
    • schrodinger.application.matsci.flywheel.ldmol.diffusion.utils module

Submodules

• schrodinger.application.matsci.flywheel.ldmol.autoencoder module
• schrodinger.application.matsci.flywheel.ldmol.bert_decoder module
• schrodinger.application.matsci.flywheel.ldmol.constants module
• schrodinger.application.matsci.flywheel.ldmol.dit_model module
• schrodinger.application.matsci.flywheel.ldmol.generator module
• schrodinger.application.matsci.flywheel.ldmol.smiles_utils module
• schrodinger.application.matsci.flywheel.ldmol.text_encoder module
• schrodinger.application.matsci.flywheel.ldmol.tokenizer module
• schrodinger.application.matsci.flywheel.ldmol.utils module