Dataset Information

Using attribution to decode binding mechanism in neural network models for chemistry.

ABSTRACT: Deep neural networks have achieved state-of-the-art accuracy at classifying molecules with respect to whether they bind to specific protein targets. A key breakthrough would occur if these models could reveal the fragment pharmacophores that are causally involved in binding. Extracting chemical details of binding from the networks could enable scientific discoveries about the mechanisms of drug actions. However, doing so requires shining light into the black box that is the trained neural network model, a task that has proved difficult across many domains. Here we show how the binding mechanism learned by deep neural network models can be interrogated, using a recently described attribution method. We first work with carefully constructed synthetic datasets, in which the molecular features responsible for "binding" are fully known. We find that networks that achieve perfect accuracy on held-out test datasets still learn spurious correlations, and we are able to exploit this nonrobustness to construct adversarial examples that fool the model. This makes these models unreliable for accurately revealing information about the mechanisms of protein-ligand binding. In light of our findings, we prescribe a test that checks whether a hypothesized mechanism can be learned. If the test fails, it indicates that the model must be simplified or regularized and/or that the training dataset requires augmentation.

SUBMITTER: McCloskey K

PROVIDER: S-EPMC6575176 | biostudies-literature | 2019 Jun

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Using attribution to decode binding mechanism in neural network models for chemistry.

McCloskey Kevin K Taly Ankur A Monti Federico F Brenner Michael P MP Colwell Lucy J LJ

Proceedings of the National Academy of Sciences of the United States of America 20190524 24

Deep neural networks have achieved state-of-the-art accuracy at classifying molecules with respect to whether they bind to specific protein targets. A key breakthrough would occur if these models could reveal the fragment pharmacophores that are causally involved in binding. Extracting chemical details of binding from the networks could enable scientific discoveries about the mechanisms of drug actions. However, doing so requires shining light into the black box that is the trained neural networ ...[more]

PMID: 31127041

Similar Datasets

Project description:BackgroundLiver cirrhosis is a chronic liver disease with hepatocyte necrosis and lesion. As one of the TCM formulas Wuling Powder (WLP) is widely used in the treatment of liver cirrhosis. However, it's key functional components and action mechanism still remain unclear. We attempted to explore the Key Group of Effective Components (KGEC) of WLP in the treatment of Liver cirrhosis through integrative pharmacology combined with experiments.MethodsThe components and potential target genes of WLP were extracted from published databases. A novel node importance calculation model considering both node control force and node bridging force is designed to construct the Function Response Space (FRS) and obtain key effector proteins. The genetic knapsack algorithm was employed to select KGEC. The effectiveness and reliability of KGEC were evaluated at the functional level by using gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis. Finally, the effectiveness and potential mechanism of KGEC were confirmed by CCK-8, qPCR and Western blot.Results940 effective proteins were obtained in FRS. KEGG pathways and GO terms enrichments analysis suggested that effective proteins well reflect liver cirrhosis characteristics at the functional level. 29 components of WLP were defined as KGEC, which covered 100% of the targets of the effective proteins. Additionally, the pathways enriched for the KGEC targets accounted for 83.33% of the shared genes between the targets and the pathogenic genes enrichment pathways. Three components scopoletin, caryophyllene oxide, and hydroxyzinamic acid from KGEC were selected for in vivo verification. The qPCR results demonstrated that all three components significantly reduced the mRNA levels of COL1A1 in TGF-β1-induced liver cirrhosis model. Furthermore, the Western blot assay indicated that these components acted synergistically to target the NF-κB, AMPK/p38, cAMP, and PI3K/AKT pathways, thus inhibiting the progression of liver cirrhosis.ConclusionIn summary, we have developed a new model that reveals the key components and potential mechanisms of WLP for the treatment of liver cirrhosis. This model provides a reference for the secondary development of WLP and offers a methodological strategy for studying TCM formulas.

Dataset Information

Using attribution to decode binding mechanism in neural network models for chemistry.

Publications

Using attribution to decode binding mechanism in neural network models for chemistry.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets