Transitive prediction of small-molecule function through alignment of high-content screening resources

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• High-content image-based phenotypic screens (HCSs) generate valuable compound profile data.
• Existing HCS datasets are isolated due to experimental and computational variability, hindering integration.
• A unified approach is needed to leverage the growing volume of HCS data for drug discovery.

Purpose of the Study:

• To develop a computational framework for integrating heterogeneous HCS profile datasets.
• To enable accurate prediction of compound functions across different HCS studies.
• To accelerate early drug discovery by unifying HCS resources.

Main Methods:

• A contrastive, deep-learning framework was developed.
• Sparse overlapping profiles were used as fiducials for dataset alignment.
• Datasets were aligned into a shared latent space.

Main Results:

• The framework successfully aligned heterogeneous HCS datasets.
• Accurate 'transitive' predictions of compound function were achieved.
• The method allows predicting functions of uncharacterized compounds by leveraging data from other datasets.

Conclusions:

• In silico alignment of HCS resources is feasible and effective.
• This approach unifies isolated HCS datasets, creating a more comprehensive resource.
• The framework accelerates early drug discovery by enabling cross-dataset compound profiling.