Overview

Introduction

GenMol is a masked diffusion model trained on Sequential Attachment-based Fragment Embedding (SAFE)¹ representations for fragment-based molecule generation, which can serve as a generalist model for various drug discovery tasks. As the most important feature of GenMol compared to other molecular generation models, the use of SAFE format enables users to design a template for a generation schema, such as:

1. Specifying fixed fragment(s), which will remain unchanged in generation.
2. Specifying specific positions that generated fragments will attach to.
3. Generating a partial or full fragment or generating multiple fragments.
4. Generating fragments at any range of lengths specified.

The inference by this model is a masking-unmasking process inspired by the idea of masked discrete diffusion². It accepts a SAFE-formatted text sequence (representing a molecule) with some fragments masked (represented by asterisk symbols) and translates it into tokens (with masking tokens) before supplying it to the Transformer/BERT-based neural network model. In each forward pass, the model will try to predict the tokens at all masked positions, and only one or a few tokens with the highest probit values will be picked in a step, which will be repeated until all masking positions are recovered.

Applications

GenMol can be applied for a variety of scenarios for molecular generation:

• De Novo generation​ - random sampling of valid molecular sequences with certain lengths.
• Conditioned generation - sequence completion from a given molecular structure, such as:
  • Motif extension and scaffold decoration - Adding a new fragment sequence to a specified attachment point in a molecule.
  • Superstructure generation - Adding a new fragment sequence to any of the possible attachment points (randomly) in a molecule.
  • Linker design - Generation of a sequence linking two separated molecular fragments on specified attachment points.
  • One-step linker design - Direct generation of a linker connecting two specified fragments in a single inference pass. v2 substantially improves the success rate over v1.
• Molecule optimization with Oracle methods, such as:
  • Hit-generation - Screening drug candidates generated from a library of molecular fragments for effective binding affinity to the target.
  • Lead-optimization - Sampling molecules from a pre-identified lead compound with improved drug-like properties, such as ADMET.

NIM Features

The GenMol model is wrapped and delivered as NVIDIA Inference Microservice (NIM) with the following features:

• Scalable deployment - GenMol NIM is a Docker image that you deploy to supported Linux systems with NVIDIA GPUs. You can scale it to more GPUs by running one container instance per GPU.

• Standard API - GenMol inferences are OpenAPI standard requests to the HTTP(S) endpoints of a NIM hosted in the NVIDIA API Catalog or self-deployed on premises. You can integrate it into pipelines and workflows for molecular design.

• Enterprise-level optimization and support - NIM products are optimized for the performance, reliability, and security of AI inference, and NVIDIA monitors and patches CVEs in them.

• NIM telemetry support - Optional telemetry controls help improve deployment reliability and compatibility while preserving user privacy.

References

¹ Gotta be SAFE: A New Framework for Molecular Design, Noutahi et al, 2023. Link

² Simple and Effective Masked Diffusion Language Models, Sahoo et al, 2024. Link