# AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery

**Authors:** Izhar Wallach, Michael Dzamba, Abraham Heifets  
**arXiv:** [1510.02855](https://arxiv.org/abs/1510.02855)  
**Date:** 2015-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1510.02855v1/

## Key points

- AtomNet is a deep convolutional neural network for predicting bioactivity of small molecules in drug discovery
- Deep convolutional neural networks leverage spatial and temporal structure for superior predictive performance
- AtomNet applies these concepts to modeling bioactivity and chemical interactions, resulting in highly accurate predictions
- Local convolutional filters successfully predict new active molecules for targets with no previously known modulators
- Outperforms previous docking approaches on diverse set of benchmarks by large margin, achieving AUC greater than 0.9 on 57.8% of targets in DUDE benchmark
- Invaluable tool for accelerating drug discovery efforts and reducing costs associated with traditional methods such as high-throughput screening and lead optimization studies
- Represents significant advancement in field of drug discovery with potential to revolutionize approach to research

## Summary

AtomNet is a deep convolutional neural network designed for predicting the bioactivity of small molecules in drug discovery applications. Deep convolutional neural networks are a subclass of deep neural networks that leverage the spatial and temporal structure of the domain they model, achieving superior predictive performance in areas such as speech and image recognition by hierarchically composing simple local features into complex models. AtomNet applies these concepts to the modeling of bioactivity and chemical interactions, resulting in a highly accurate prediction model. In contrast to existing DNN techniques, AtomNet's application of local convolutional filters to structural target information successfully predicts new active molecules for targets with no previously known modulators. It also outperforms previous docking approaches on a diverse set of benchmarks by a large margin, achieving an AUC greater than 0.9 on 57.8% of the targets in the DUDE benchmark. This makes it an invaluable tool for accelerating drug discovery efforts and reducing costs associated with traditional methods such as high-throughput screening and lead optimization studies. Overall, AtomNet represents a significant advancement in the field of drug discovery and has the potential to revolutionize how we approach this critical area of research.