Development of a standardized methodology for transfer learning with QSAR models: a purely data-driven approach for...

Area of Science:

• Computational chemistry
• Machine learning applications in drug discovery

Background:

• Transfer learning is effective for chemical endpoints but limited by the need for domain expertise in source task selection.
• Existing methods require significant prior knowledge, hindering broader application of transfer learning.

Purpose of the Study:

• To develop a purely data-driven approach for selecting optimal source tasks in transfer learning, eliminating the need for domain knowledge.
• To create a predictive model for transferability to guide the selection of effective source tasks.

Main Methods:

• A supervised learning framework was established to predict transfer outcome (positive/negative).
• Six transferability metrics were calculated using source and target dataset information as predictive features.
• A transferability prediction model (TP-Model) was trained and evaluated on 100,000 random transfers from the ChEMBL database.

Main Results:

• The TP-Model demonstrated a 135-fold increase in precision with 92% sensitivity, significantly outperforming random search.
• Transfer learning application resulted in an average Matthews Correlation Coefficient (MCC) increase of 0.19 (single source) and 0.44 (multiple sources).

Conclusions:

• The developed data-driven TP-Model effectively identifies suitable source tasks for transfer learning in cheminformatics.
• This approach enhances the efficiency and applicability of transfer learning, leading to substantial performance improvements in chemical endpoint prediction.