ncats / ld50-multitask

Official repository for multitask deep learning models.

Large-scale Modeling of Multi-Species Acute Toxicity Endpoints using Consensus of Multi-Task Deep Learning Methods

This repository contains multitask deep learning models developed using acute toxicity data, primarily focusing on the endpoints: lethal dose fifty (LD₅₀); lethal dose low (LDLO); and toxic dose low (TDLO). Please note that the data was obtained from ChemIDPlus.

Results

Our best models are based on a consensus of best developed individual models. We compared our best models against the multi-task deep learning models by Sosnin et al. While they report models for a total of 29 toxicity endpoints, our models are based on a total of 59 endpoints. A total of 18 LD₅₀ endpoints were in common. The results for these 18 endpoints are listed below. The performance measure reported is root mean squared error (lower is better).

species	route	cpds (ours)	cpds (Sosnin et al)	score (ours)	scorea (Sosnin et al)
mouse	intraperitoneal	36295	37202	0.41	0.41
mouse	oral	23373	24355	0.39	0.42
mouse	intravenous	16978	17742	0.43	0.43
rat	oral	10190	10743	0.52	0.53
mouse	subcutaneous	6769	7221	0.51	0.51
rat	intraperitoneal	5021	5041	0.52	0.55
rat	intravenous	2472	2538	0.52	0.54
rat	subcutaneous	1896	2014	0.63	0.64
mouse	unreported	1739	1804	0.47	0.51
rabbit	skin	1495	1734	0.53	0.56
mammalb	unreported	1129	1121	0.42	0.40
rabbit	oral	894	910	0.58	0.58
rat	skin	835	930	0.61	0.63
rat	unreported	806	838	0.58	0.60
rabbit	intravenous	792	764	0.59	0.68
guinea pig	oral	793	799	0.66	0.70
rat	oral	322	966	0.63	0.61
rat	intraperitoneal	318	1029	0.52	0.43

^a the scores are from the supplementary information of the original article; ^b the mammalian species and route are unspecified

Other Models

We also report single-task models using baseline methods: random forest and deep neural networks. The scripts used for modeling can be found under scripts/. An example notebooks/create_fold_data.ipynb to create the training and test sets by joining the descriptors and task details for different folds of cross-validation is provided.

ld50-multitask's Issues

ld50 multitask

What packages do I need to install in my environment to run the ld50 multitask script?

Where are descriptors?

Hi!

It seems like data files have to contain descriptors (rdkit, fp, pairs, etc.). Where are they? Or where is the code that calculates them? I cannot run you code because of that.

Thanks.