Collaborative Profile-QSAR: A Natural Platform for Building Collaborative Models among Competing Companies

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Multitask learning models significantly outperform single-task models in bioactivity prediction.
• Sharing assay data across organizations could yield superior predictive models, but data privacy is a major concern.
• Profile-QSAR (pQSAR) is a stacked, multitask model that uses predictions from single-task models as compound descriptors.

Purpose of the Study:

• To evaluate the effectiveness of collaborative model sharing using a profile-QSAR approach.
• To simulate internal model sharing within a large pharmaceutical company (Novartis) to address data privacy concerns.
• To determine the impact of collaborative modeling on prediction accuracy for both internal and external compounds.

Main Methods:

• Implemented a two-level, multitask, stacked profile-QSAR model.
• Simulated collaborative model sharing by training models on internal and external compound/assay data subsets.
• Assessed prediction performance using various combinations of internal and external model profiles.

Main Results:

• Multitask pQSAR models consistently outperformed single-task models.
• Collaborative modeling did not improve predictions for internal compounds.
• Predictions for external compounds improved with collaboration, but less than purely internal multitask models.
• Increased overlap in compound collections enhanced collaborative model performance for external compounds.
• A consensus model combining internal and external profiles performed best for external compounds and offered a compromise for internal ones.

Conclusions:

• Collaborative model sharing via profile-QSAR is a viable strategy for improving predictions on external compounds.
• Internal data remains superior for predicting internal compound activity, but external data aids in broader applicability.
• The findings suggest that model sharing, without direct data sharing, can yield benefits similar to direct data aggregation.