Transfer Learning and Permutation-Invariance Improving Predicting Genome-Wide, Cell-Specific and Directional...

Area of Science:

• Computational Biology
• Pharmacology
• Systems Biology

Background:

• Precision medicine requires predicting complex system interventions beyond single drugs.
• Extensive single-compound data and deep learning models exist.
• Transfer learning offers a promising approach to leverage existing data for complex systems.

Purpose of the Study:

• To develop a predictive model for complex system intervention effects using transfer learning.
• To fine-tune a deep model pre-trained on single-compound data for complex systems like natural products.
• To evaluate the model's accuracy and generalizability on unseen complex systems.

Main Methods:

• Utilized a permutation-invariant deep learning model as the core.
• Pre-trained the model on large-scale single-compound intervention data in cell lines.
• Fine-tuned the model on limited complex system intervention data in cell lines, creating SETComp (Set Embedding and Transfer learning model for Complex systems).

Main Results:

• SETComp achieved high accuracy (93.86% and 92.70%) on complex system-cell-gene association tests, outperforming baselines by 5.82%–27.59%.
• Predictive accuracy for novel complex systems improved by up to 24.83% compared to baselines.
• In vitro validation confirmed up to 88.65% of predictions, showing a significant positive correlation with real-world fold change.

Conclusions:

• SETComp effectively predicts complex system intervention effects, demonstrating the power of transfer learning.
• The model shows significant potential for applications in mechanism uncovering and drug repositioning.
• SETComp advances the integration of deep learning with complex biological systems for biomedical research.