Related Experiment Video
Updated: Jul 10, 2026
09:19
NMR-Based Fragment Screening in a Minimum Sample but Maximum Automation Mode
Published on: June 4, 2021
3.5K
Variable-temperature token sampling in decoder-GPT molecule-generation can produce more robust and potent virtual
1Department of Chemistry, University of Reading, Reading, RG1, UK. m.cafiero@reading.ac.uk.
Physical Chemistry Chemical Physics : PCCP
|June 25, 2025
Summary
Variable temperature sampling in generative pretrained transformers (GPTs) creates larger, more effective molecular screening libraries. This method outperforms traditional techniques for generating HMG-coenzyme-A reductase inhibitors.
Area of Science:
- Computational chemistry
- Artificial intelligence in drug discovery
- Molecular modeling
Background:
- Generative Pretrained Transformers (GPTs) are increasingly used for molecular generation.
- Conventional token generation methods include greedy decoding, top-k, and top-p sampling.
- Optimizing molecular generation for specific targets like HMG-coenzyme-A reductase remains a challenge.
Purpose of the Study:
- To investigate the efficacy of variable temperature sampling for generating molecules with GPTs.
- To compare variable temperature sampling against conventional methods for generating HMG-coenzyme-A reductase inhibitors.
- To identify optimal variable temperature schemes for enhanced molecular library generation.
Main Methods:
- Utilized a GPT model trained for HMG-coenzyme-A reductase inhibitor generation.
- Implemented and compared various temperature-based sampling techniques, including single-temperature and variable temperature ramps (e.g., sigmoidal).
- Evaluated generated molecular libraries based on size, predicted IC50 values, docking scores, and synthetic accessibility.
Main Results:
- Variable temperature sampling, particularly with a sigmoidal ramp early in generation, produced larger screening libraries compared to greedy decoding or single-temperature sampling.
- The libraries generated using variable temperature sampling exhibited lower predicted IC50 values and docking scores.
- Synthetic accessibility scores were also improved with the variable temperature approach, especially for shorter prompt lengths.
Conclusions:
- Variable temperature sampling is a superior method for generating diverse and high-quality molecular libraries with GPTs.
- A sigmoidal temperature ramp early in the generation process is recommended for optimizing inhibitor discovery.
- This approach enhances the efficiency of identifying potential drug candidates for targets like HMG-coenzyme-A reductase.
