Fragment-based Pretraining and Finetuning on Molecular Graphs

Authors: Kha-Dinh Luong, Ambuj K Singh

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our graph fragment-based pretraining (Graph FP) advances the performances on 5 out of 8 common molecular benchmarks and improves the performances on long-range biological benchmarks by at least 11.5%. Code is available at: https://github.com/lvkd84/Graph FP. (Abstract) and '4 Experiments'
Researcher Affiliation Academia Kha-Dinh Luong, Ambuj Singh Department of Computer Science University of California, Santa Barbara Santa Barbara, CA 93106 {vluong,ambuj}@cs.ucsb.edu
Pseudocode No The paper describes the methodology using text and mathematical equations but does not include structured pseudocode or algorithm blocks with explicit labels such as 'Algorithm' or 'Pseudocode'.
Open Source Code Yes Code is available at: https://github.com/lvkd84/Graph FP.
Open Datasets Yes We use a processed subset containing 456K molecules from the Ch EMBL database [24] for pretraining.
Dataset Splits Yes For downstream evaluation, we consider 8 binary graph classification tasks from Molecule Net [36] with scaffold split [15]. Moreover, to assess the ability of the models in recognizing global arrangement, we consider two graph prediction tasks on large peptide molecules from the Long-range Graph Benchmark [7]. Long-range graph benchmarks are split using stratified random split.
Hardware Specification Yes All experiments are run on individual Tesla V 100 GPUs.
Software Dependencies No The paper mentions software components like GIN and Adam W optimizer but does not specify their version numbers or the versions of other key software dependencies (e.g., Python, PyTorch/TensorFlow libraries).
Experiment Setup Yes All pretrainings are done in 100 epochs, with Adam W optimizer, batch size 256, and initial learning rate 1 × 10−3. We reduce the learning rate by a factor of 0.1 every 5 epochs without improvement. On graph classification benchmarks, to ensure comparability, our finetuning setting is mostly similar to that of previous works [15, 23]: 100 epochs, Adam optimizer, batch size 256, initial learning rate 1 × 10−3, and dropout rate chosen from {0.0, 0.5}.