Workflow
Figure 1: The workflow of the training process of our molecular generator underthe reinforcement learning framework. During this loop, each token in the SMILESsequence might be also generated by the pre-trained model with a smallprobability, which was set manually as exploring rate.
De novo drug molecule generation with generative
adversarial networks
X.Liu1, K.Ye2, H.W.T.van Vlijmen1,3, A.P.IJzerman1, G.J.P.van Westen11 Drug Discovery and Safety, Leiden Academic Center for Drug Research, Einsteinweg 55, Leiden, The Netherlands, 2 Omics and Omics informatics, Xi’an Jiaotong University, 28 Xianning W Rd, Xi’an, China,
3 Janssen Pharmaceutica NV, Beerse, Belgium
Introduction• Quantitative structure-activity relationship (QSAR) modeling is
important for drug discovery and well established[1] .• Recently, deep learning has been introduced as a new
method for the construction of QSAR models[2].• Now deep learning can also be applied to generate
molecules with a desired activity, or inverse QSAR[3,4].
AimAlthough several groups investigated the performance ofdeep learning for de novo drug design[3,4], the diversity ofthe generated molecules was unsatisfactory[5]. Here, wepropose a method that integrates an exploring strategy intogenerative adversarial networks to generate ligands for theA2A adenosine receptor as various as possible.
Conclusions1. Our exploring strategy decreased the number of duplicate
molecules.2. A higher exploring rate could increase the diversity of the
generated molecules, but it also output more undesiredmolecules.
3. The baseline helps the model generate more desired moleculesand increase their scores. But this effect at the lower exploringrate was not as significant as at the higher exploring rate.
Contact InformationXuhan Liu, MSc
Email: [email protected]: +31 71 527 4332
References1. Cherkasov, A., et al., QSAR modeling: where haveyou been? Where are you going to? J Med Chem,2014. 57(12): p. 4977-5010.2. Lenselink, E.B., et al., Beyond the hype: deepneural networks outperform established methodsusing a ChEMBL bioactivity benchmark set. JCheminform, 2017. 9(1): p. 45.3. Olivecrona, M., et al., Molecular de-novo designthrough deep reinforcement learning. JCheminform, 2017. 9(1): p. 48.4. Benjamin, S.-L., et al., Optimizing distributionsover molecular space. An Objective-ReinforcedGenerative Adversarial Network for Inverse-designChemistry (ORGANIC). 2017.5. Benhenda, M. ChemGAN challenge for drugdiscovery: can AI reproduce natural chemicaldiversity? ArXiv e-prints, 2017. 1708.
Performance
Figure 2: (a) The average score at each epoch during the training processwith different parameters. (b) Receiver operator characteristic (ROC) curvefor the predictor with different machine learning methods. (c, d) Scatter plotshowing the logP ~ weight distribution (c) and the principle componentsanalysis (d) for predicted active molecules generated by our proposed modeland REINVENT[3] through 10,000 times sampling, compared with real ligandsof the A2A adenosine receptor.
Exploring rate=0.01 Exploring rate=0.1 REINVENTPre-
trained
Baseline 0.0 0.1 0.0 0.1 -- --
Valid SMILES
99.32% 99.55% 95.85% 98.75% 99.38% 93.88%
DesiredSMILES
98.66% 99.23% 74.60% 80.09% 98.70% 0.70%
UniqueSMILES
98.26% 96.86% 73.02% 80.03% 97.51% 0.70%
Diversity 0.7946 0.7866 0.8452 0.8510 0.8090 0.8724
Loss Function
𝐿(𝜃) = (𝑅 𝑌1:𝑇 − 𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒)
𝑡=1
𝑇
log 𝑃( 𝑦𝑡 | 𝑌1:𝑡−1)
Molecule Diversity[5]
𝐷(𝑀) =1
|𝑀|2
𝑎,𝑏 ∈(𝑀,𝑀)
1 −|𝑓𝑝𝑎 ∩ 𝑓𝑝𝑏|
|𝑓𝑝𝑎 ∪ 𝑓𝑝𝑏|
(b)(a)
Figure 3: The generated molecules by our proposed method (a) and REINVENT(c) and known ligands of A2A adenosine receptor (b) in 5 clusters based onprinciple component analysis.
(d)(c)
(1)
(3)
(2)
(4)
(5)
(b)(a) (c)