Cheminformatics

Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design

Haoyu Lin, Jianfeng Pei

Featured July 21, 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AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content.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 fragment-based chemical language and a clever trick to inject different design rules directly into its brain, letting it create new drug-like molecules for many purposes with one simple system.

In depth

The paper introduces Molexar, a unified multimodal molecular foundation model for drug design. Its core innovation lies in Fragment-SELFIES, a novel fragment-aware molecular language that guarantees valid molecules and explicitly encodes fragment structures. This model unifies diverse design constraints—from scalar properties to protein pockets—by injecting them as value-token embeddings into a single autoregressive decoder, eliminating the need for task-specific architectures.

Key Takeaways

1
The authors developed Fragment-SELFIES, a robust molecular language that explicitly represents BRICS fragments, ensuring 100% validity in generated molecules and facilitating fragment-constrained design.
2
Molexar unifies multimodal conditioning (e.g., molecular properties, pharmacophores, protein sequences, binding pockets) into a single autoregressive decoder via value-token embedding replacement, simplifying the architecture and enabling efficient, cache-compatible generation.
3
The model achieves high performance across unconditional, fragment-constrained, property-controlled, and target-conditioned generation tasks, demonstrating superior efficiency and drug-likeness compared to existing specialized models.

Conceptual Flow

HIGH LEVEL

Methodology: Unified Molecular Design

The model learns a special chemical language and then uses a clever trick to add different design rules, like desired properties or protein shapes, into the same system to create new molecules.

Chemical Language

Property Rules

Protein Shape

Combine & Learn

New Molecules

Results: Efficient and Effective Drug Candidates

The new system creates valid, diverse, and high-quality drug-like molecules much faster and with less computer memory than older methods, even when given many specific design instructions.

Old Methods

New Method

Compare Performance

Slower, More Memory

Faster, Less Memory

Better Molecules