Providing a multimodal approach to protein target identification, lead compound discovery, and toxicity studies.
We encode molecular properties and structure using a waveform to enhance drug lead discovery and expedite the clinical trial process.
What is Sound of Molecules
Modern molecular AI has converged on two representations: the graph of atoms and bonds, and the sequence of text. Both work. Both omit important information. We introduce a complementary modality: audio. By encoding a molecule's structure and physical properties into an acoustic signal, we give machine learning an independent source of signal and a native bridge to self-supervised audio AI.
Atoms and bonds as nodes and edges. Powerful, but pooling and aggregation discard how structure is spatially organized.
EstablishedMolecules and proteins as strings of tokens. Scales beautifully, but carries no intrinsic spatial or temporal dimension.
EstablishedStructure encoded as a waveform. Interference between oscillators preserves the collective organization that aggregation throws away.
Our ModalityWhat we've shown
Our approach is no longer a concept. It is demonstrated at both ends of the molecular world with peer-reviewed and benchmarked results.
Our molecular sonification framework lets pre-trained speech models learn molecular properties, establishing that audio carries real, transferable signal for drug-property prediction.
A deterministic, training-free encoding of binding pockets. Acoustic superposition triples class separation of standard aggregation, and the advantage grows with more structures.
We start where our evidence is strongest, predicting the properties that make or break a drug candidate. Multimodal AI has matured to the point where a genuinely new modality can connect to the entire ecosystem of models. We're building that modality, and the science to back it.