Autonomous Assignment of Melanin Spectra amidst Dynamical Heterogeneity

The assignment of the exact features of the melanin spectra has been challenging---due to its extreme heterogeneity. Melanin in biologically active form does not have a single structure but a heterogeneous ensemble of compositionally diverse structures. As a result, traditional computational spectroscopy-based assignment is ineffective because they involve starting with a known molecular structure that is varied to observe the effect in the spectrum. To overcome this issue, we propose an inverse solution by training machine learning models to assign from an ultraviolet/visible absorption spectrum of a melanin mixture, the underlying structures in varying oxidation states and conformational preferences. Our analyses reveal the dynamic heterogeneity to have a dominant impact on the spectral features than the effect due to oligomer sizes. We demonstrate that the structures recreated using the data-driven approach accurately depict the known melanin structures. We finally apply our methodology to assign an experimental spectrum of melanin and rationalize the model's predictions.