The Goldilocks paradigm: comparing classical machine learning, large language models, and few-shot learning for drug...

Area of Science:

• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) models like transformers (large language models, LLMs) and few-shot learning (FSLC) show promise, but the 'no-free lunch' theorem implies no single algorithm is universally best.
• Classical ML, FSLC, and transformer models have distinct strengths and weaknesses depending on the task.

Purpose of the Study:

• To evaluate the performance of classical (SVR), FSLC, and transformer (MolBART) models across various dataset sizes and diversity levels.
• To identify the optimal ML model strategy based on dataset characteristics.

Main Methods:

• Comparative analysis of Support Vector Regression (SVR), few-shot learning (FSLC), and transformer (MolBART) models.
• Testing model performance on datasets varying in size (small, medium, large) and diversity (feature distribution).

Main Results:

• FSLC models outperform others on small datasets (<50 molecules).
• Transformers demonstrate superior performance on diverse, small-to-medium datasets (50-240 molecules).
• Classical ML models perform best on large, sufficiently sized datasets.

Conclusions:

• The optimal ML model choice is contingent upon dataset size and diversity, defining a 'goldilocks zone' for each model type.
• This research provides guidance for selecting appropriate ML algorithms for new datasets in scientific applications.