MolEmb: Multimodal Large Language Models Can Be Strong Molecular Embedding Models

Traditional molecular embedding models typically map a molecule to a single, fixed vector, which limits their utility for tasks requiring different semantic lenses (e.g., toxicity vs. solubility). This paper investigates whether Multimodal Large Language Models (MLLMs)—which natively process images, text, and symbolic data—can serve as general molecular embedding models that produce flexible, context-aware representations.

The authors introduce MolEmb, a lightweight framework that adapts MLLMs into molecular embedding models. The framework represents molecules through multi-view profiles (2D depictions and SMILES strings) and uses a bidirectional contrastive objective to align these profiles with textual descriptions in a shared embedding space. To evaluate the model's ability to handle semantic context, the authors introduce MolCAR (Molecular Context-Aware Retrieval), a diagnostic benchmark that tests whether a model can route a molecular query to the correct task-specific document when provided with a natural-language instruction.

MolEmb demonstrates that MLLMs are a viable route for molecular representation learning. The model achieves competitive performance on standard molecular property prediction benchmarks and supports effective cross-modal retrieval between molecules and text. Crucially, the authors find that while generic molecule-text alignment creates a shared space, it does not inherently provide context-aware routing. By using task-diverse, outcome-grounded supervision (continued alignment on MolCAR-Train), the model learns to use natural-language instructions to successfully distinguish between different valid descriptions of the same molecule, effectively expanding the embedding space to be task-sensitive.

This work shifts the paradigm of molecular representation learning from fixed-view specialist encoders to flexible, instruction-following foundation models. It demonstrates that the limitation of current molecular embeddings is often a property of the training data rather than the model architecture. By curating task-aware, outcome-grounded datasets, researchers can build more versatile molecular tools that adapt to the specific scientific questions being asked.