Evolutionary Strategy for Optimizing Neural Models for Compound Similarity Estimation
Molecular similarity evaluation is a key aspect of bioinfor- matics and poses a significant challenge when dealing with compounds with unknown structures. In this context, graph neural networks have proven effective in generating representations based on the topology of chemical reactions. However,...
Guardado en:
| Autores principales: | , , |
|---|---|
| Formato: | Articulo |
| Lenguaje: | Inglés |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | http://sedici.unlp.edu.ar/handle/10915/177971 |
| Aporte de: |
| Sumario: | Molecular similarity evaluation is a key aspect of bioinfor- matics and poses a significant challenge when dealing with compounds with unknown structures. In this context, graph neural networks have proven effective in generating representations based on the topology of chemical reactions. However, designing these models and selecting their hyperparameters requires exploring a vast range of options. Evolutio- nary algorithms naturally arise as a solution for searching these exten- sive spaces, including the hyperparameter space of neural architectures.
This study presents a comparison between a traditional hyperparameter search approach, based on expert knowledge, and a method leveraging evolutionary computation for the same task, specifically in compound similarity estimation. Using a predefined architecture, experiments are conducted to compare both approaches across different datasets. The re- sults indicate that the evolutionary computation-based method success- fully identifies suitable hyperparameters for the evaluated architecture, achieving a performance comparable to the expert-driven approach while eliminating the need for human intervention in the selection process. |
|---|