Drug Discovery
Tox 21 Dataset and the TrimNet Architecture
Tox 21 is a Drug Discovery dataset
12,060 training samples and 647 test samples that represent chemical compounds.
Chemical descriptors, such as molecular weight, solubility or surface area
12 binary labels that represent the outcome (active/inactive) of 12 different toxicological experiments
TrimNet: learning molecular representation from triplet messages for biomedicine
3rd place deep learning algorithm on this dataset
First place open source architecture
Improvement on AUC score using Perforated Backpropagation + TrimNet
We downloaded the open source TrimNet code and added our technology
Graph shows Test AUC scores which is the metric used on the Tox 21 competition
Best result Increased AUC score by .078
40.9% closer to a perfect score
Average gap between top 5 papers on leaderboard the for this dataset is 0.0062
Over 7 experiments, first PB cycle on average reduced error 8.2%
Average of 20.2% at point of highest overall validation score
