Modeling Reactivity to Biological Macromolecules with a Deep Multitask Network.

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Drug development faces high failure rates due to late-stage toxicity detection.
• Idiosyncratic adverse drug reactions (IADRs) and drug-induced liver injury (DILI) are difficult to predict.
• Reactive metabolites forming DNA/protein adducts are a common cause of drug toxicity.

Purpose of the Study:

• To develop a computational method for early prediction of drug molecule reactivity.
• To identify potential sites of covalent binding to DNA and proteins.
• To reduce late-stage drug candidate failures caused by unexpected toxicity.

Main Methods:

• Trained a deep convolution neural network (XenoSite reactivity model) on literature data.
• Predicted sites and probability of reactivity with glutathione, cyanide, protein, and DNA.
• Developed a selectivity score to differentiate macromolecule reactivity from screening traps.

Main Results:

• XenoSite achieved high site-level prediction accuracy (AUC 89.8% for DNA, 94.4% for protein).
• The model distinguished reactive from non-reactive molecules (AUC 78.7% for DNA, 79.8% for protein).
• Identified 9.2% of molecules reactive only with DNA and 8.1% only with protein, missed by standard screening.

Conclusions:

• The XenoSite model accurately predicts drug molecule reactivity and binding sites.
• This computational approach can flag problematic drug candidates early, reducing development costs and failures.
• Predicting reactivity sites allows for molecular modification to reduce toxicity while maintaining efficacy.