Is Multitask Deep Learning Practical for Pharma?

Abstract

Multitask deep learning has emerged as a powerful tool for computational drug discovery. However, despite a number of preliminary studies, multitask deep networks have yet to be widely deployed in the pharmaceutical and biotech industries. This lack of acceptance stems from both software difficulties and lack of understanding of the robustness of multitask deep networks. Our work aims to resolve both of these barriers to adoption. We introduce a high-quality open-source implementation of multitask deep networks as part of the DeepChem open-source platform. Our implementation enables simple python scripts to construct, fit, and evaluate sophisticated deep models. We use our implementation to analyze the performance of multitask deep networks and related deep models on four collections of pharmaceutical data (three of which have not previously been analyzed in the literature). We split these data sets into train/valid/test using time and neighbor splits to test multitask deep learning performance under challenging conditions. Our results demonstrate that multitask deep networks are surprisingly robust and can offer strong improvement over random forests. Our analysis and open-source implementation in DeepChem provide an argument that multitask deep networks are ready for widespread use in commercial drug discovery.

Learning Drug Function from Chemical Structure with Convolutional Neural Networks and Random Forests

“…Although there are several chemistry problems where DNNs outperform other shallow machine learning methods 49,59,60 , here the MFP+RF performed best with the small dataset of 676 molecules in the 5-and 12-class predictions. However, in the 3class task with the small dataset, and all the tasks with the large dataset, the two models produced accuracies that were nearly indistinguishable.…”

“…Random forests 39 are ensembles of decision trees, where each tree is learned on a subsample of data points and features (in this case, bits in a molecular fingerprint). Benchmarking studies often include MFP+RF models because they are easy to train and have strong performance on a variety of computational chemistry tasks 12,40,[46][47][48][49][50] . The random forest model was implemented with scikit-learn 51 .…”

Prediction of hERG K+ channel blockage using deep neural networks

“…Deep neural networks model is similar to neural networks in the brain. It stimulates the process of signal conduction between neurons (Junshui, Sheridan, Andy, Dahl, & Vladimir, 2015;Korotcov, Tkachenko, Russo, & Ekins, 2017;Lecun, Bengio, & Hinton, 2015;Ramsundar et al, 2017). Three types of layers constitute the whole network, as shown in Figure S1.…”

Practical Model Selection for Prospective Virtual Screening

“…Random Forest (RF): Random forests are ensembles of decision trees that are often used as a baseline in virtual screening benchmarks. 35,36 We used scikit-learn 37 to train a random forest classifier for each binary label with Morgan fingerprints as features. The RF models used 4,000 to 16,000 estimators, 1 to 1000 minimum samples at a leaf node, a bound on the maximum number of features, and other hyperparameters described in Table S6.…”

“…Other metrics like AUC[ROC] or AUC[PR], which is more appropriate for problems where the inactive compounds far outnumber the actives, 58 may still be reasonable for benchmarking virtual screening methods on large existing datasets where the entire ranked list of compounds is evaluated 36. Some recent studies3,35,59 reported that deep learning models substantially outperform traditional supervised learning approaches, including random forests. Our finding that a random forest model was the most accurate in both cross-validation and our prospective screen does not refute those results.…”