Generalized molecular property imputation using a flexible transformer architecture

Chemical data is fundamentally sparse, with molecular structures serving as database keys for countless properties. Current machine learning methods map structures to properties with remarkable accuracy, yet they do not leverage available property information when predicting unknowns, creating unutilized partial information. Here, a flexible transformer architecture is developed that addresses this gap by predicting molecular properties from any combination of available properties and structural information. The model employs custom embedding layers differentiating present from missing values, self-attention mechanisms learning property-property relationships, and property-specific decoders handling variable output sets. Testing on 16M organic molecules with 23 properties through 710 independently trained models subject to fixed missingness reveals R\textsuperscript{2} values of 0.75-0.85 for property-to-property predictions without structural data, confirming the imputation utility of extant property data. Models combining structure and property information consistently outperform either alone. In harder random missingness scenarios, self-supervised protocols accurately complete sparse datasets with average R\textsuperscript{2} values of 0.79 and 0.85 in two case studies. This work establishes a new paradigm enabling models to utilize all available molecular information flexibly.