MolAudioNet: Towards Molecular Foundation Models Through Multimodal Learning

MolAudioNet introduces audio as a third modality for molecular machine learning, complementing traditional text (SMILES) and graph representations. We present a systematic framework for encoding molecular structures as audio waveforms and release MolAudioNet-50K, the first large-scale molecular audio dataset containing 50,284 compounds with synchronized text, graph, and audio representations. Preliminary validation demonstrates promising results: 15-23% improvements over text-only baselines on drug classification and molecular property prediction tasks. Audio representations uniquely capture stereochemical and conformational information missed by other modalities. This preprint represents initial validation results. Comprehensive benchmarking across diverse molecular property prediction tasks is ongoing and will be reported in future versions. Dataset, code, and audio samples available at: https://molaudionet.com/downloads/ Part of the MolecularWorld ecosystem for democratizing molecular knowledge and enabling multimodal molecular AI.