CoChat
Sign Up
Cited by 317Computational Drug Discovery MethodsMachine Learning in Materials ScienceNeural Networks and Applications

The DOI matches its Crossref record; 71 of the 71 references it cites were found in a registry or an index.

Deep-learning: investigating deep neural networks hyper-parameters and comparison of performance to shallow methods for modeling bioactivity data

Open access
Koutsoukas, Alexios, Monaghan, Keith J., Li, Xiaoli, Huan, Jun·2017·Journal of Cheminformatics
DOI 10.1186/s13321-017-0226-yPMC PMC5489441
Open this paper in CoChat — free accountOpen PDF

Abstract

BACKGROUND: In recent years, research in artificial neural networks has resurged, now under the deep-learning umbrella, and grown extremely popular. Recently reported success of DL techniques in crowd-sourced QSAR and predictive toxicology competitions has showcased these methods as powerful tools in drug-discovery and toxicology research. The aim of this work was dual, first large number of hyper-parameter configurations were explored to investigate how they affect the performance of DNNs and could act as starting points when tuning DNNs and second their performance was compared to popular methods widely employed in the field of cheminformatics namely Naïve Bayes, k-nearest neighbor, random forest and support vector machines. Moreover, robustness of machine learning methods to different levels of artificially introduced noise was assessed. The open-source Caffe deep-learning framework and modern NVidia GPU units were utilized to carry out this study, allowing large number of DNN configurations to be explored. RESULTS: We show that feed-forward deep neural networks are capable of achieving strong classification performance and outperform shallow methods across diverse activity classes when optimized. Hyper-parameters that were found to play critical role are the activation function, dropout regularization, number hidden layers and number of neurons. When compared to the rest methods, tuned DNNs were found to statistically outperform, with p value <0.01 based on Wilcoxon statistical test. DNN achieved on average MCC units of 0.149 higher than NB, 0.092 than kNN, 0.052 than SVM with linear kernel, 0.021 than RF and finally 0.009 higher than SVM with radial basis function kernel. When exploring robustness to noise, non-linear methods were found to perform well when dealing with low levels of noise, lower than or equal to 20%, however when dealing with higher levels of noise, higher than 30%, the Naïve Bayes method was found to perform well and even outperform at the highest level of noise 50% more sophisticated methods across several datasets.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1186/s13321-017-0226-y

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 a A feed-forward deep neural network with two hidden layers, each layer consists of multiple neurons, which are fully connected with neurons of the previous and following layers. b Each artificial neuron receives one or more input signals x 1, x 2,…, x m and outputs a value y to neurons of the next layer. The output y is a nonlinear weighted sum of input signals. Nonlinearity is achieved by passing the linear sum through non-linear functions known as activation functions. c Popular neurons activation functions: the rectified linear unit (ReLU) (red), Sigmoid (Sigm) (green) and Tanh (blue)
Fig. 1 a A feed-forward deep neural network with two hidden layers, each layer consists of multiple neurons, which are fully connected with neurons of the previous and following layers. b Each artificial neuron receives one or more input signals x 1, x 2,…, x m and outputs a value y to neurons of the next layer. The output y is a nonlinear weighted sum of input signals. Nonlinearity is achieved by passing the linear sum through non-linear functions known as activation functions. c Popular neurons activation functions: the rectified linear unit (ReLU) (red), Sigmoid (Sigm) (green) and Tanh (blue)CC BY · PubMed Central
Fig. 2 Comparison of activation functions rectified linear units (ReLU), Tanh and Sigmoid (Sigm) on the performance of DNNs. DNNs with a single hidden layer and variable number of neurons were trained and tested using the activation functions ReLU (red), Sigm (green) and Tanh (blue) over fivefold cross validation. The performance was measured using MCC as evaluation metric
Fig. 2 Comparison of activation functions rectified linear units (ReLU), Tanh and Sigmoid (Sigm) on the performance of DNNs. DNNs with a single hidden layer and variable number of neurons were trained and tested using the activation functions ReLU (red), Sigm (green) and Tanh (blue) over fivefold cross validation. The performance was measured using MCC as evaluation metricCC BY · PubMed Central
Fig. 3 Effect of the hyper-parameters (i) number of hidden layers, (ii) number of neurons and (iii) dropout regularization on the performance of DNNs measured by MCC as evaluation metric. DNN configuration A shows results obtained by DNN with a single hidden layer and 10 neurons, ReLU activation function and no regularization averaged over the seven activity datasets, B a two hidden layered DNN with 500 neurons in each layer, ReLU activation function and no regularization, C two hidden layers with 3000 neurons per hidden layer and dropout regularization (0% for the input and 50% for hidden layers), D two hidden layers with 3000 neurons per hidden layer and dropout regularization (20% for the input and 50% for hidden layers), E two hidden layers with 3000 neurons per hidden layer and dropout regularization (50% for both the input and hidden layers), F three hidden layers with 3000 neurons per layer and dropout regularization (50% for both the input and hidden layers) and G four hidden layers with 3500 neurons per layer and dropout regularization (50% for the input and hidden layers)
Fig. 3 Effect of the hyper-parameters (i) number of hidden layers, (ii) number of neurons and (iii) dropout regularization on the performance of DNNs measured by MCC as evaluation metric. DNN configuration A shows results obtained by DNN with a single hidden layer and 10 neurons, ReLU activation function and no regularization averaged over the seven activity datasets, B a two hidden layered DNN with 500 neurons in each layer, ReLU activation function and no regularization, C two hidden layers with 3000 neurons per hidden layer and dropout regularization (0% for the input and 50% for hidden layers), D two hidden layers with 3000 neurons per hidden layer and dropout regularization (20% for the input and 50% for hidden layers), E two hidden layers with 3000 neurons per hidden layer and dropout regularization (50% for both the input and hidden layers), F three hidden layers with 3000 neurons per layer and dropout regularization (50% for both the input and hidden layers) and G four hidden layers with 3500 neurons per layer and dropout regularization (50% for the input and hidden layers)CC BY · PubMed Central
Fig. 4 Boxplot of differences between performances achieved by tuned DNN and the rest algorithms measured using MCC as evaluation metric on the validation sets over the seven activity classes. Results are ranked by decreased mean differences. The differences ranged on average from 0.149 MCC units between DNN and NB, 0.092 DNN and kNN, 0.052 DNN and SVM with linear kernel, 0.021 DNN and RF and 0.009 DNN and SVM with “rbf” kernel
Fig. 4 Boxplot of differences between performances achieved by tuned DNN and the rest algorithms measured using MCC as evaluation metric on the validation sets over the seven activity classes. Results are ranked by decreased mean differences. The differences ranged on average from 0.149 MCC units between DNN and NB, 0.092 DNN and kNN, 0.052 DNN and SVM with linear kernel, 0.021 DNN and RF and 0.009 DNN and SVM with “rbf” kernelCC BY · PubMed Central
Fig. 5 Robustness of machine learning methods to different levels of noise for 4 out of 7 activity classes. At low levels of noise, lower that 20%, non-linear methods performed well achieving performance higher than 0.7 MCC units for most of the tested datasets. Instead, at higher level of noise, equal to or higher than 30%, performance for most algorithms dropped below 0.7 MCC and in several occasions even lower than 0.6 at 50% of noise. Naïve Bayes method was found to be the least affected method achieving in several tested datasets performance higher than 0.6 MCC even at the highest level of noise tested 50% and outperforming more complex methods
Fig. 5 Robustness of machine learning methods to different levels of noise for 4 out of 7 activity classes. At low levels of noise, lower that 20%, non-linear methods performed well achieving performance higher than 0.7 MCC units for most of the tested datasets. Instead, at higher level of noise, equal to or higher than 30%, performance for most algorithms dropped below 0.7 MCC and in several occasions even lower than 0.6 at 50% of noise. Naïve Bayes method was found to be the least affected method achieving in several tested datasets performance higher than 0.6 MCC even at the highest level of noise tested 50% and outperforming more complex methodsCC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 69 references

Research works printed but not linked: 2

  1. Aspuru-Guzik A, Duvenaud D, Maclaurin D, Aguilera-Iparraguire J, Gomez-Bombarelli R, Hirzel TD, Adams RP (2015) Convolutional networks on graphs for learning molecular fingerprints. In: Proceedings of neural information processing systems, Quebec, pp 2224–2232
  2. Barrett SJ, Langdon WB (2006) Advances in the application of machine learning techniques in drug discovery. Des Dev 36:99–110

Cited by

Sorted by citations

Related

See all

Metadata

Authors
Koutsoukas, A., Monaghan, K. J., Li, X., & Huan, J.
Affiliations
University of Kansas (US)
Corresponding author
Huan, Jun
ORCID
Venue
Journal of Cheminformatics
Publisher
Springer Science and Business Media LLC
Published
2017-06-28
Volume
9
Issue
1
Pages
42-42
ISSN
1758-2946 (ISSN-L)
DOI
10.1186/s13321-017-0226-y
PMC
PMC5489441
OpenAlex
W2731161895
Cited by
315
References
69
Citations by year
2026: 12 · 2025: 25 · 2024: 40 · 2023: 42 · 2022: 53 · 2021: 55 · 2020: 32 · 2019: 33 · 2018: 22 · 2017: 1
FWCI
18.06
Type
journal article
Language
en
Open access
Licence
Locations
  • Journal of Cheminformatics (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • National University of Singapore (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
National Science Foundation (CNS 1337899)
Topics
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
Keywords
deep neural network, hyperparameter optimization, bioactivity prediction, QSAR, predictive toxicology, feedforward neural networks, noise robustness, activation functions, dropout regularization, Caffe, drug discovery, naive Bayes, k-nearest neighbors, random forest, support vector machine

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review