Functionally Regionalized Knowledge Transfer for Low-resource Drug Discovery

Authors: Huaxiu Yao, Ying Wei, Long-Kai Huang, Ding Xue, Junzhou Huang, Zhenhui (Jessie) Li

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically evaluate the effectiveness of FRML on two diverse drug discovery tasks: drug activity prediction and ADMET property prediction. Empirical results on both virtual screening and ADMET prediction validate the superiority of FRML over state-of-the-art baselines powered with interpretability in assay relationship.
Researcher Affiliation Collaboration Huaxiu Yao1 , Ying Wei2, Long-Kai Huang3, Ding Xue3 Junzhou Huang4, Zhenhui Li5 1Stanford University, 2City University of Hong Kong, 3 Tencent AI Lab 4University of Texas at Arlington, 5Pennsylvania State University
Pseudocode Yes Algorithm 1 Meta-training Process of FRML
Open Source Code No The paper does not contain any statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes For drug activity prediction, we use the dose-response activity assays from ChEMBL[1], where 4,276 assays are selected in this problem. The AMDET Prediction problem is constructed by combining 4 benchmark datasets from the Molecule Net [36] with biophysiology and physiology targets. The 4 datasets are MUV [29], SIDER [11], Tox21 and Tox Cast [27].
Dataset Splits Yes Here, we randomly sample 100 assays as the meta-testing set, 76 assays as the meta-validation set, and the rest of assays for meta-training.
Hardware Specification No The paper does not specify any particular hardware components such as GPU models, CPU types, or memory used for running the experiments. It only generally refers to the computational environment.
Software Dependencies No The paper mentions RDKit [12] as an implementation tool ('implemented in RDKit [12]') but does not provide a specific version number for it or any other software dependencies in the main text.
Experiment Setup Yes We construct a neural network consisting of two fully connected layers as the predictive model, which also serves as the base learner f. We denote the weights for the base learner f to be θ. With the base learner f, we introduce gradient-based meta-learning as the backbone meta-learning framework, which regards the initialization θ0 for the base learner as the transferable knowledge. Here α denotes the learning rate for assay adaptation. where β is the learning rate for meta-updating. where τ is the temperature and qml ml+1 i is sampled from the Gumbel distribution... where the hyperparameter λ balances between two losses... Detailed hyperparameters for both applications are listed in Appendix D.