Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prob-cGAN: A Probabilistic Conditional Generative Adversarial Network for LSD1 Inhibitor Activity Prediction

Hanyang Wang

TL;DR

The Probabilistic Conditional Generative Adversarial Network (Prob-cGAN), designed to predict the activity of LSD1 inhibitors, was evaluated against state-of-the-art models using the ChEMBL database, demonstrating superior performance.

Abstract

The inhibition of Lysine-Specific Histone Demethylase 1 (LSD1) is a promising strategy for cancer treatment and targeting epigenetic mechanisms. This paper introduces a Probabilistic Conditional Generative Adversarial Network (Prob-cGAN), designed to predict the activity of LSD1 inhibitors. The Prob-cGAN was evaluated against state-of-the-art models using the ChEMBL database, demonstrating superior performance. Specifically, it achieved a top-1 $R^2$ of 0.739, significantly outperforming the Smiles-Transformer model at 0.591 and the baseline cGAN at 0.488. Furthermore, it recorded a lower $RMSE$ of 0.562, compared to 0.708 and 0.791 for the Smiles-Transformer and cGAN models respectively. These results highlight the potential of Prob-cGAN to enhance drug design and advance our understanding of complex biological systems through machine learning and bioinformatics.

Prob-cGAN: A Probabilistic Conditional Generative Adversarial Network for LSD1 Inhibitor Activity Prediction

TL;DR

The Probabilistic Conditional Generative Adversarial Network (Prob-cGAN), designed to predict the activity of LSD1 inhibitors, was evaluated against state-of-the-art models using the ChEMBL database, demonstrating superior performance.

Abstract

The inhibition of Lysine-Specific Histone Demethylase 1 (LSD1) is a promising strategy for cancer treatment and targeting epigenetic mechanisms. This paper introduces a Probabilistic Conditional Generative Adversarial Network (Prob-cGAN), designed to predict the activity of LSD1 inhibitors. The Prob-cGAN was evaluated against state-of-the-art models using the ChEMBL database, demonstrating superior performance. Specifically, it achieved a top-1 of 0.739, significantly outperforming the Smiles-Transformer model at 0.591 and the baseline cGAN at 0.488. Furthermore, it recorded a lower of 0.562, compared to 0.708 and 0.791 for the Smiles-Transformer and cGAN models respectively. These results highlight the potential of Prob-cGAN to enhance drug design and advance our understanding of complex biological systems through machine learning and bioinformatics.

Paper Structure

This paper contains 10 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Feature descriptor
  3. Prob-cGAN
  4. Dataset description
  5. Evaluation Metrics
  6. Result
  7. Result: Comparsion against Machine Learning Algorithms
  8. Result: Comparison against Random Forest and Smile Transformers.
  9. Results: Prob-cGAN vs. cGAN.
  10. Conclusion

Figures (3)

  • Figure 1: Overview of the proposed Probabilistic Conditional Generative Adversarial Networks (CGAN) with molecule activity prediction. (1) SMILES strings of each molecule are converted into Morgan Fingerprint sandfort2020structure and word embedding vector levy2014neural descriptors, which are combined to form 812-dimensional vectors. (2) An autoencoder is applied to remove noisy features and extract a 203-dimensional latent space $x$. (3) The noise-injection generator incorporates the noise vector $z$ by concatenating it with the hidden representation at each layer, introducing randomness and variability into the generated molecules. (4) The discriminator receives input from the generated fake molecule activities, latent space vectors, and true activities, and produces a final one-dimensional output without any activation function, corresponding to "T" in the f-GAN framework.
  • Figure 2: Model results of activity prediction.
  • Figure 3: Model results of activity prediction distribution