Molecule Design by Latent Prompt Transformer

Authors: Deqian Kong, Yuhao Huang, Jianwen Xie, Edouardo Honig, Ming Xu, Shuanghong Xue, Pei Lin, Sanping Zhou, Sheng Zhong, Nanning Zheng, Ying Nian Wu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that LPT not only effectively discovers useful molecules across single-objective, multi-objective, and structure-constrained optimization tasks, but also exhibits strong sample efficiency.
Researcher Affiliation Collaboration 1Department of Statistics and Data Science, UCLA 2Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University 3Bio Map Research 4Akool Research 5Institute of Engineering in Medicine, UCSD 6Shu Chien-Gene Lay Department of Bioengineering, UCSD
Pseudocode Yes Algorithm 1 MLE learning of Latent Prompt Transformer (LPT); Algorithm 2 Online learning of LPT
Open Source Code Yes Project page: https://sites.google.com/view/latent-prompt-transformer. (Footnote 1). In the NeurIPS Paper Checklist, it states: "Code is released in our project page."
Open Datasets Yes For molecule design tasks, we use SELFIES representations of the ZINC (Irwin et al., 2012) dataset, which comprises 250K drug-like molecules as our offline dataset D. [...] The TF Bind 8 task focuses on identifying DNA sequences that are 8 bases long... This task contains a training set of 32,898 samples... The GFP task involves generating protein sequences of 237 amino acids... we use a subset of 5,000 samples as the training set by following the methodology outlined in Trabucco et al. (2022).
Dataset Splits No The paper mentions training sets and evaluating on certain numbers of samples, but does not explicitly provide percentages or counts for validation splits, nor does it specify a cross-validation setup for all experiments.
Hardware Specification Yes Training was conducted on an NVIDIA A6000 GPU, requiring 20 hours for pre-training, 10 hours for fine-tuning, and 12 hours for online learning. Additional details can be found in App. A.3.
Software Dependencies No The paper mentions RDKit and Auto Dock GPU as tools used for property computation and the Adam W optimizer for training, but does not specify their version numbers or other crucial software dependencies with specific versions (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes The prior model, pα(z), of LPT is a one-dimensional UNet... The sequence generation model, pβ(x|z), is implemented as a 3-layer causal Transformer, while a 3-layer MLP serves as the predictor model, pγ(y|z). As described in Sec. 3.2, we pre-train LPT on molecules for 30 epochs and then fine-tune it with target properties for an additional 10 epochs... We use the Adam W optimizer... with a weight decay of 0.1.