Can large language models predict antimicrobial peptide activity and toxicity?

Area of Science:

• Computational biology
• Medicinal chemistry
• Drug discovery

Background:

• Antimicrobial peptides (AMPs) offer a promising solution to the antimicrobial resistance crisis.
• Clinical development of AMPs is hindered by challenges in controlling their toxicity to human cells.

Purpose of the Study:

• To investigate the potential of large language models (LLMs) in predicting AMP activity and toxicity.
• To compare the performance of LLMs against traditional machine learning models for AMP data analysis.

Main Methods:

• Fine-tuning LLMs, including GPT-3 and GPT-3.5, using data from the Database of Antimicrobial Activity and Structure of Peptides (DBAASP).
• Training recurrent neural networks (RNNs) on sequence-activity data and support vector machines (SVMs) on MAP4C molecular fingerprint-activity data.
• Evaluating model performance for AMP activity prediction and hemolysis (as a toxicity proxy).

Main Results:

• GPT-3 demonstrated promising but irreproducible results for activity and hemolysis prediction.
• GPT-3.5 performed less effectively than GPT-3.
• RNNs and SVMs outperformed LLMs in predicting AMP activity and toxicity.

Conclusions:

• Simpler machine learning models (RNNs, SVMs) are currently recommended for predicting AMP activity and toxicity over LLMs.
• The rapid advancement of LLMs necessitates ongoing re-evaluation of their predictive capabilities in this field.