Developing Collaborative QSAR Models Without Sharing Structures.

Area of Science:

• cheminformatics
• computational chemistry
• drug discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) models improve with more data, but predictive performance degrades when chemical structures diverge from the training set.
• Expanding the applicability domain of QSAR models requires increasing the diversity of training data, often by combining diverse data sources.
• Integrating proprietary data into QSAR model development is challenging due to intellectual property concerns and the need to protect confidential structural information.

Purpose of the Study:

• To present a novel method for collaborative development of linear regression QSAR models.
• To address the challenge of incorporating diverse datasets, including proprietary data, into QSAR model development.
• To enable the creation of QSAR models with expanded applicability domains without compromising data confidentiality.

Main Methods:

• Development of a collaborative linear regression modeling approach.
• Utilizing aggregated data sharing to prevent disclosure of individual data points and confidential structural information.
• Comparing models developed through collaborative aggregated data sharing with those built using combined datasets.

Main Results:

• The proposed method allows for the collaborative development of QSAR models using aggregated data.
• Confidential structural information is protected by not sharing individual data points.
• The final QSAR models developed through this collaborative approach are equivalent in performance to models built with directly combined datasets.

Conclusions:

• Collaborative QSAR model development can be effectively achieved by sharing data in an aggregated form.
• This approach successfully expands the applicability domain of QSAR models while safeguarding proprietary information.
• The method offers a viable solution for integrating diverse data sources in a privacy-preserving manner for improved QSAR modeling.