Foundations of Molecular Generation with GP-MoLFormer on AMD Instinct MI300X Accelerators

2026-02-03 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Chemistry · Depth: Intermediate, long

Summary

The article introduces GP-MoLFormer, an open-source generative foundation model from IBM designed for molecular generation, trained on over a billion canonical SMILES strings. It is based on the MoLFormer architecture, utilizing a decoder-only transformer with causal language modeling, linear attention, and Rotary Positional Embeddings (RoPE) to learn molecular syntax and patterns. Unlike MatterGen, which focuses on 3D crystalline structures, GP-MoLFormer operates on symbolic molecular representations, treating generation as a sequence completion problem. The model supports unconditional generation, scaffold-constrained decoration, and property-oriented design via pair-tuning for properties like QED, penalized logP, and DRD2 binding affinity. Evaluation uses the MOSES benchmark, assessing fragment and scaffold similarity, diversity (IntDivp), global distributional similarity (FCD), validity, uniqueness, and novelty. The article also provides a practical guide for setting up and running GP-MoLFormer on AMD Instinct MI300X accelerators using Docker and the AMD Container Toolkit.

Key takeaway

For AI Engineers and Research Scientists working on drug discovery or materials science, GP-MoLFormer offers a robust, open-source solution for molecular generation. You should consider integrating this sequence-based model for exploring chemical space, especially when designing molecules with specific properties via pair-tuning, leveraging its compatibility with AMD GPUs for efficient execution.

Key insights

GP-MoLFormer enables efficient, large-scale molecular generation using sequence-based language models on AMD GPUs.

Principles

• Molecular design uses discrete symbolic descriptions.
• Generative models can explore chemical landscapes with intent.
• Open-source ecosystems foster reproducibility and collaboration.

Method

GP-MoLFormer uses a decoder-only transformer with causal language modeling on SMILES strings. It supports unconditional, scaffold-constrained, and pair-tuned property-oriented generation by learning prompt vectors to bias chemical space exploration.

In practice

• Run GP-MoLFormer on AMD Instinct MI300X accelerators.
• Use Docker with AMD Container Toolkit for setup.
• Apply pair-tuning for property-guided molecular design.

Topics

• GP-MoLFormer
• Molecular Generation
• AMD Instinct MI300X
• Generative Models
• SMILES

Code references

• ROCm/rocm-blogs
• IBM/gp-molformer
• IBM/gp-molformer
• molecularsets/moses

Best for: AI Engineer, Machine Learning Engineer, Research Scientist

Related on AIssential

• De novo molecular generation with optical property preconditioning at the token level — Token-conditioned autoregressive models can generate OLED molecules with targeted optical properties, but control varies by chemical motif.
• GROMACS Performance on AMD Instinct MI355X — AMD's MI355X GPU significantly boosts GROMACS molecular dynamics performance, especially with compute partitioning.
• DanceOPD: On-Policy Generative Field Distillation — DanceOPD unifies conflicting image generation capabilities by distilling expert velocity fields into a single flow-matching model.
• Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules — Generative models should expand their "generable set" to discover new-to-nature designs beyond training data.
• Augmenting Molecular Language Models with Local $n$-gram Memory — Integrating local $n$-gram memory via MolGram efficiently resolves the locality gap in molecular language models, improving performance with fewer parameters.
• From Data to Theory: Autonomous Large Language Model Agents for Materials Science — Autonomous LLM agents can perform end-to-end scientific theory development, from equation selection to code execution and validation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.