This code implements seq3seq fingerprint method, the paper and results were published by ACM BCB 2018.
We use docker/nvidia-docker for both development and production running, mostly with docker image utasmile/tensorflow:1.4.1-gpu.
For detailed instructions, you might want to refer to https://github.com/NVIDIA/nvidia-docker#quickstart。
ZINC is used for our experiments, which is a free database of commercially-available compounds for virtual screening. You can download ZINC datasets from http://zinc.docking.org/
To train a seq2seq model from a debug data with only 5 SMILE representations in tests/data/debug.smi.
Assume you clone the code under your $HOME dir and your current work directory is the code folder root.
Example usage:
nvidia-docker run --rm -it -v $HOME:$HOME -w $PWD
utasmile/tensorflow:1.4.1-gpu smilectl examples/train_debug.sml runlocalExample output:
I0615 21:19:51.832336 140140613834496 seq3seq_model.py:313] Initializing a fresh training...
I0615 21:19:51.835680 140140613834496 seq3seq_model.py:88] Initializing model with hparams: {"loss_ty
pe": "SparseXentLoss", "dropout_rate": 0.5, "learning_rate_decay_factor": 0.99, "pred_net_type": "MLP
", "buckets": [[30, 30], [60, 60], [90, 90]], "label_states": false, "target_vocab_size": 40, "batch_
size": 256, "source_vocab_size": 40, "use_recovery": true, "mlp_output_nums": [2], "num_layers": 3, "
metric_type": "BCMetric", "alpha": 1.0, "max_gradient_norm": 5.0, "learning_rate": 0.5, "size": 128}
I0615 21:21:09.036127 140140613834496 seq3seq_model.py:344] Initialize fresh parameters...
I0615 21:21:14.931850 140140613834496 train.py:178] Reading train data from tests/data/debug.tokens..
.
I0615 21:21:14.932178 140140613834496 train.py:180] Reading test data from tests/data/debug.tokens...
I0615 21:22:24.285237 140140613834496 train.py:240] global step 200 learning rate 0.500000 step-time
0.143644 perplexity6.374212
I0615 21:22:24.285351 140140613834496 train.py:242] loss_unsupervised: 1.7795321 loss_supervised:
0.0
I0615 21:22:36.758291 140140613834496 train.py:279] eval: bucket 0 perplexity 4.509512, em_acc 0.00
0000, acc_sup None
I0615 21:22:39.320552 140140613834496 train.py:279] eval: bucket 1 perplexity 5.055138, em_acc 0.00
0000, acc_sup None
I0615 21:22:39.320717 140140613834496 train.py:255] eval: empty bucket 2
I0615 21:22:58.310951 140140613834496 train.py:240] global step 400 learning rate 0.500000 step-time
0.094621 perplexity3.357505
I0615 21:22:58.311058 140140613834496 train.py:242] loss_unsupervised: 0.83335733 loss_supervised
: 0.0
I0615 21:23:08.053030 140140613834496 train.py:279] eval: bucket 0 perplexity 3.724646, em_acc 0.00
0000, acc_sup None
I0615 21:23:08.102361 140140613834496 train.py:279] eval: bucket 1 perplexity 3.564833, em_acc 0.00
0000, acc_sup None
I0615 21:23:08.102560 140140613834496 train.py:255] eval: empty bucket 2
We includes few training examples using smilectl script in examples/ folder.
- train_debug_sup.sml: Example for training with a supervised task.
If our work is helpful for your research, please consider citing:
@inproceedings{zhang2018seq3seq,
title={Seq3seq fingerprint: towards end-to-end semi-supervised deep drug discovery},
author={Zhang, Xiaoyu and Wang, Sheng and Zhu, Feiyun and Xu, Zheng and Wang, Yuhong and Huang, Junzhou},
booktitle={Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics},
pages={404--413},
year={2018},
organization={ACM}
}
@article{xu2017seq2seqfingerprint,
title={Seq2seq Fingerprint: An Unsupervised Deep Molecular Embedding for Drug Discovery},
author={Zheng Xu, Sheng Wang, Feiyun Zhu, and Junzhou Huang},
journal={BCB’17, Aug 2017, Boston, Massachusetts USA},
year={2017}
}