Abstract
<jats:p>Molecular property prediction underpins virtual screening, reaction design, and toxicity assessment, yet most deep learning models reason from a single molecular view—typically a two-dimensional graph or a SMILES string. Such single-modality representations discard complementary signals that chemists routinely exploit: the three-dimensional conformation governs quantum behavior, the textual description encodes sub-structural intent, and spectroscopic measurements provide direct experimental evidence. A further difficulty is that these modalities contribute unequally across tasks and training stages, so naive fusion lets a dominant modality suppress the others and fails when a view is missing. We propose ChemMoB (Chemistry-aware Multi-modal Balance), a framework that ports contribution-aware dynamic multi-modal balancing to chemical representation learning. ChemMoB encodes four complementary views of a molecule—a 2D graph, a 3D conformation, a SMILES/Transformer text stream, and a 1D spectroscopic signal—and fuses them through a Contribution-aware Dynamic Balance (CDB) module that re-estimates per-modality weights every epoch. To handle missing or noisy inputs, ChemMoB adds a generative spectral enhancement head, a static-to-dynamic lift that reconstructs an approximate 3D embedding from the 2D graph, and a provenance watermark that flags unreliable spectra. We evaluate ChemMoB on QM9 (quantum properties), ESOL (aqueous solubility), and ADMET binary tasks (blood-brain barrier penetration and hERG liability). ChemMoB consistently outperforms single-modality and fixed-fusion baselines, reducing QM9 average MAE by 12.0% over fixed fusion and improving BBBP and hERG AUROC to 0.873 and 0.831. Ablations confirm that the dynamic balance module and the static-to-dynamic enhancement are the principal sources of gain, and that ChemMoB remains robust when up to 30% of spectral inputs are dropped.</jats:p>