In the quiet laboratories of computational chemistry, scientists have long relied on a specific kind of digital fingerprint to understand the invisible world of molecules. These fingerprints, known as embeddings, are mathematical summaries that capture the essence of a chemical structure, allowing computers to predict how a substance will behave, whether it will dissolve in water, or if it might be toxic. For years, creating these fingerprints required specialized tools designed for a single purpose: taking a picture of a molecule or reading its chemical code and spitting out one fixed number. This approach worked well for specific tasks, but it lacked flexibility. It could not easily shift its focus based on what a scientist was asking. If a researcher wanted to know about a molecule's ability to cross the blood-brain barrier, the tool gave the same answer as if they were asking about its ability to dissolve in oil, because the tool did not understand the question, only the shape of the molecule.

This limitation has become a bottleneck as the field moves toward more complex drug discovery and material design. Scientists needed a way to talk to these digital fingerprints, to ask them to highlight different features depending on the context of the inquiry. The question became whether the powerful, flexible artificial intelligence systems that can already read text and see images could be repurposed to create these adaptable chemical fingerprints. Instead of building a new, narrow tool for every chemical problem, researchers wondered if a single, general-purpose system could learn to represent molecules in a way that changes based on the language used to describe them.

A team of researchers from the Chinese University of Hong Kong, Shenzhen, and the Shanghai Artificial Intelligence Laboratory set out to answer this question with a new framework they call MolEmb. They started with a type of artificial intelligence known as a multimodal large language model. These are the same kinds of systems that can look at a photograph, read a sentence, and understand the relationship between the two. The researchers realized that these models were naturally suited to handle the different ways a molecule can be described: as a two-dimensional drawing, as a string of text characters representing its chemical structure, or as a set of instructions about what property to investigate.

The team built a lightweight system to teach these large models to act as molecular embedding experts. They did not try to teach the model to generate new molecules or write chemical papers from scratch. Instead, they focused on training the model to align the visual and textual description of a molecule with a specific text instruction. Imagine a molecule as a complex object that can be viewed through different lenses. The researchers showed the model a picture of a molecule and its chemical code, then paired that with a text instruction like "analyze for toxicity" or "analyze for solubility." The goal was to make the model produce a unique digital fingerprint for that molecule that shifted slightly depending on the instruction. When the instruction was about toxicity, the fingerprint would emphasize the parts of the molecule relevant to safety. When the instruction changed to solubility, the fingerprint would shift to highlight the parts relevant to dissolving in water.

To test if this approach worked, the researchers ran a series of rigorous experiments. First, they checked if the new system could predict chemical properties as accurately as the specialized tools currently in use. They found that the system, even without any special training on specific chemical tasks, could predict properties like solubility and toxicity with a level of accuracy that rivaled the best existing methods. This was a significant finding because it suggested that the general-purpose model had already learned enough about chemical structures to be useful, simply by understanding the relationship between the molecule's image, its text code, and the language used to describe it.

Next, the team tested whether these fingerprints could be used to search for molecules using text. In traditional systems, you cannot search a database of molecules by typing a sentence like "find me a molecule that is toxic to liver cells" and expect to get a relevant result, because the chemical data and the text data live in separate worlds. The researchers showed that their new system created a shared space where molecules and text descriptions could be compared directly. When they asked the system to find a molecule matching a specific description, it succeeded with high accuracy, proving that the model had truly learned to connect the visual and textual aspects of chemistry in a single, unified representation.

The most critical test, however, was to see if the system could truly change its mind based on the context. The researchers created a diagnostic benchmark called MolCAR to see if the model could distinguish between different valid descriptions of the exact same molecule. They presented the model with a single molecule and asked it to find the correct description for a specific task, such as toxicity, while ignoring other valid descriptions for different tasks, such as hydration. Before they added a final layer of specific training, the model struggled to make this distinction. It treated the molecule as having one fixed identity, regardless of the question asked. However, after the researchers trained the model on a dataset specifically designed to pair molecules with different task-focused descriptions, the system learned to shift its focus. It began to produce different fingerprints for the same molecule depending on whether the question was about toxicity or solubility.

The results suggest that multimodal large language models are not just tools for generating chemical ideas or answering questions; they can serve as the foundation for a new kind of molecular representation. The study indicates that the ability to create context-aware chemical fingerprints is not an inherent flaw in the model's architecture, but rather a matter of the data it is trained on. When the training data includes diverse examples of the same molecule described under different scientific lenses, the model learns to adapt. This finding points toward a future where a single, general-purpose system can replace a collection of narrow, specialized tools, allowing scientists to query chemical data with the same flexibility they use to search for information in a library. The work demonstrates that by aligning molecular profiles with natural language instructions, these powerful models can become a versatile infrastructure for the next generation of drug discovery and chemical research.