Cheminformatics

Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design

Haoyu Lin, Yiyan Liao, Jinmei Pan, Xinliao Ling, Luhua Lai, Jianfeng Pei

Featured July 1, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI model called Molexar uses a special molecular language and clever input tricks to design new drug molecules, letting it follow many different rules like desired properties or fitting into a protein pocket, all with one efficient system.

In depth

The paper introduces Molexar, a unified multimodal molecular foundation model for drug design. It leverages Fragment-SELFIES, a novel fragment-aware molecular language that guarantees valid molecules and explicitly encodes fragment structures. Crucially, Molexar unifies diverse conditioning modalities (e.g., molecular properties, protein pockets) by replacing value-token embeddings in a shared autoregressive decoder, eliminating the need for task-specific architectures and enabling efficient, cache-compatible generation.

Key Takeaways

1
Introduces Fragment-SELFIES, a robust molecular language that explicitly represents BRICS fragment trees, ensuring 100% validity and supporting fragment-constrained generation without corpus-specific IDs.
2
Proposes a unified multimodal conditioning mechanism via value-token embedding replacement, allowing a single autoregressive decoder to handle diverse inputs like scalar properties, pharmacophores, protein sequences, and binding pockets.
3
Achieves state-of-the-art efficiency and performance in unconditional, fragment-constrained, property-controlled, and target-conditioned molecular generation, outperforming larger models with a smaller parameter count.

Conceptual Flow

HIGH LEVEL

Methodology: Unified Molecular Design

The model takes different kinds of instructions and turns them into a special code, then uses that code to build new molecules one piece at a time.

Molecule Rules

Protein Info

Fragment Parts

Convert & Combine

Special Code

Results: Efficient & Versatile Molecule Creation

This new way of building molecules works really well, creating valid and useful drug candidates for many different design challenges, faster than older methods.

Old Methods

New Molexar

Compare Performance

More Valid Molecules

Faster Design

Better Drug-likeness