GxVAEs: Two Joint VAEs Generate Hit Molecules from Gene Expression Profiles

Authors: Chen Li, Yoshihiro Yamanishi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments and case studies on the generation of therapeutic molecules show that Gx VAEs outperforms current state-of-the-art baselines and yield hit-like molecules with potential bioactivity and drug-like properties.
Researcher Affiliation Academia Chen Li and Yoshihiro Yamanishi Graduate School of Informatics, Nagoya University, Chikusa, Nagoya, 464-8601, Japan li.chen.z2@a.mail.nagoya-u.ac.jp, yamanishi@i.nagoya-u.ac.jp
Pseudocode Yes Algorithm 1 in Appendix C shows the procedures of Gx VAEs.
Open Source Code Yes Code is available at: https://github.com/naruto7283/GxVEAs
Open Datasets Yes Chemically induced profiles were collected from the LINCS L1000 database (Duan et al. 2014)... Target perturbation profiles were obtained from the LINCS database... Disease-specific profiles were collected from the crowd extracted expression of differential signatures (CREEDS) database (Wang et al. 2016).
Dataset Splits Yes Figure F.2 in Appendix F shows the violin plots for the QED scores from the validation set (red) and the molecules generated by Mol VAE (green).
Hardware Specification No All the experiments were run on GPUs with CUDA. This statement is too general and does not provide specific hardware models (e.g., NVIDIA A100, RTX 2080 Ti) or detailed specifications.
Software Dependencies No The paper mentions software components such as 'RDKit tool' but does not specify version numbers for any software dependencies.
Experiment Setup Yes In Profile VAE, the encoder and decoder contained three feedforward layers of size of 512, 256, and 128. The learning rate and dropout probability were set to 1e 4 and 0.2, respectively. Mol VAE had an embedding size of 128 and three hidden layers of size 256. The learning rate, dropout probability, and temperature β were set to 5e 4, 0.1, and 1.0. The maximum length of generated SMILES strings was 100. The dimensionality and batch size of the latent vectors for the two VAEs were 64. Profile VAE and Mol VAE were trained for 2000 and 200 epochs, respectively.