Molecular embedding models serve as foundational infrastructure for computational chemistry and drug discovery, offering reusable vectors for property prediction, virtual screening, and retrieval. Most existing encoders focus on a single molecular view and output unconditional vectors, lacking a natural‑language interface to modulate representations. We investigate whether multimodal large language models (MLLMs), which natively handle images, text, and symbolic inputs, can act as general molecular embedding models that produce embeddings conditioned on both a molecular profile and a natural‑language semantic context.
To this end we introduce MolEmb, a lightweight framework that aligns molecular profiles with textual descriptions in a shared embedding space via a bidirectional contrastive objective. The resulting model is competitive on molecular property prediction and enables cross‑modal molecule‑text retrieval within the same space. We also present MolCAR, a diagnostic benchmark for context‑aware retrieval, and find that the effectiveness of context‑aware molecular embeddings is primarily a property of the supervision data.
These findings suggest that MLLMs are not merely chemistry assistants or generators, but a viable and extensible route to general molecular embedding models.\
Blogger's Review: MolEmb brings the powerful representation capabilities of multimodal language models into molecular science, offering flexible, context‑conditioned embeddings that open new possibilities for retrieval and prediction in drug discovery. With larger and more diverse chemical datasets, the model’s universality and practical impact could be further amplified.