Unveiling the bioactive landscape of drug inactive ingredients (DIGs) using deep transfer learning

In a drug product, the major components by mass are the drug inactive ingredients (DIGs), which raises great concerns about their unwanted effects and clinical toxicities. It is demanded to unveil their proteome-wide bioactive landscape using computational methods. However, existing methods are impeded by either incapability to scan human proteome or inaccuracy in DIGs' bioactivity prediction. Here, a cross-attention transformer model, titled TransDIG, leveraging cross-module deep transfer learning was therefore developed to map the bioactive landscape of DIGs using minimal experimental data. First, the generalizability and interpretability of this model was verified by the prediction of zero-shot proteins and identification of key atoms/residues, respectively. Then, the bioactive landscape of hundreds of DIGs was unveiled using TransDIG, and thousands of potential bioactivities were found for the DIGs currently employed in pharmaceutical industry. Finally, the bioactivities of four popular DIGs were identified based on the landscape and experimentally validated by activity assay. As a result, the colorant β-carotene was validated to inhibit a critical drug transporter, and our study presented the first in vitro evidence of the bioactivity of the antioxidant dodecyl gallate that has not previously been reported to regulate any human protein. This study might offer insights for the design of drug formulation and its clinical utilization.