LEARNING GENERAL TRANSFORMATIONS OF DATA FOR OUT-OF-SAMPLE EXTENSIONS.

Area of Science:

• Machine Learning
• Computational Biology
• Genomics

Background:

• Generative models like GANs map data distributions but struggle to generalize transformations to unseen data.
• Existing models memorize training domains, limiting their ability to apply learned transformations out-of-sample.

Purpose of the Study:

• To develop a novel neural network, the Neuron Transformation Network (NTNet), capable of isolating and generalizing data transformations.
• To enable the application of learned transformations to new datasets with different distributions.

Main Methods:

• Proposed a Neuron Transformation Network (NTNet) architecture.
• NTNet isolates transformation signals from internal data variations.
• The learned transformation signal can be removed from new datasets.

Main Results:

• Demonstrated NTNet effectiveness on over a dozen synthetic and biomedical single-cell RNA sequencing datasets.
• NTNet successfully learned transformations from genetic and drug perturbations.
• Applied learned transformations to predict treatment outcomes on new cell samples.

Conclusions:

• NTNet overcomes the generalization limitations of traditional generative models.
• This approach allows for the transfer of learned data transformations across different distributions.
• NTNet shows promise for analyzing and predicting outcomes in complex biological systems.