Learning to Optimize in Swarms, guided by posterior estimation (with application to protein docking)

Overview

Learning to optimize has emerged as a powerful framework for various optimization and machine learning tasks. Current such "meta-optimizers" often learn in the space of continuous optimization algorithms that are point-based and uncertainty-unaware. To overcome the limitations, we propose a meta-optimizer that learns in the algorithmic space of both point-based and population-based optimization algorithms. The meta-optimizer targets at a meta-loss function consisting of both cumulative regret and entropy. Specifically, we learn and interpret the update formula through a population of LSTMs embedded with sample- and feature-level attentions. Meanwhile, we estimate the posterior directly over the global optimum and use an uncertainty measure to help guide the learning process. Empirical results over non-convex test functions and a protein docking application demonstrate that this new meta-optimizer outperforms existing competitors.

http://papers.nips.cc/paper/9641-learning-to-optimize-in-swarms

Documentation

Please look at here.

Citation

@incollection{NIPS2019_9641,
title = {Learning to Optimize in Swarms},
author = {Cao, Yue and Chen, Tianlong and Wang, Zhangyang and Shen, Yang},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {15018--15028},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/9641-learning-to-optimize-in-swarms.pdf}
}

Dependencies

• TensorFlow >=1.13
• cNMA (For protein docking application: sampling)
• CHARMM (For protein docking application: initial structure minimization)

Important Source Files

• meta.py: The architecture of the model.
• train.py: The training program.
• problems.py: Store the optimization problems.
• util.py: Store the utility subroutines.
• evaluate.py: The evaluation programs.

Training

Example: 2D Rastrigin functions (see more problem options in the code)

• python train.py --problem=square_cos_2 --save_path=./square_cos_2

Evaluation

Example: 2D Rastrigin functions (see more problem options in the code)

• python evaluate.py --problem=square_cos_2 --optimizer=L2L --path=./square_cos_2

Trained Models:

In "trained_models/":

• 2D Rastrigin: "square_cos_2/"
• 10D Rastrigin: "square_cos_10/"
• 20D Rastrigin: "square_cos_20/"

Application to Protein Docking

File Explanation

• get_12basis.py: Get the 12 basis vectors through cNMA
• force_field.py: The python implementation of CHARMM19 force field (including atom charge and rdii).
• prepocess_prot.py: Generate input data.
• dataloader.py: Load input data into the model.
• charmm_setup.prl: Set up prerequsites before runing CHARMM.

Datasets:

• You can download data from https://drive.google.com/file/d/1x-Jye87YTWk_8WiooPJ23QJ_0zxiDT9V/view?usp=sharing.
• You can also generate your own data using above programs.
• ZDOCK data are from (https://zlab.umassmed.edu/zdock/) without CHARMM minimization.

Data Generation Steps:

• 1.Prepare your own pdbs in a folder.
• 2.Modify the path in prepocess_prot.py and get_12basis.py to be the folder of your own pdbs.
• 3.Download CHARMM27 and minimize the pdb complex through CHARMM.
• 3.Download cNMA, and change the path of cNMA in get_12basis.py
• 4.run get_12basis.py, force_filed.py and then prepocess_prot.py.