iORPred - iOR Function Prediction

We propose a transfer learning-based iOR function prediction strategy aimed at addressing the bottlenecks of sparse experimental data and cross-species generalization difficulties.

Two-Step Prediction Strategy

The strategy consists of two steps: First, the deep learning model is pre-trained using large-scale molecular docking data (approximately 6 million interactions recorded in the DockingSet module), enabling it to grasp the overall distribution of binding features between ORs and VOCs.

Subsequently, fine-tuning is performed on limited high-quality experimental data (from species such as Drosophila and Anopheles, already collected in the iORExp module), allowing the model to adapt to real biological response patterns. Through this approach, the model effectively combines the breadth of virtual screening with the precision of experimental data.

Model Performance Validation

Validation on multiple independent test sets demonstrates that the transfer learning model significantly outperforms the randomly initialized baseline model in both in-distribution (random split) and out-of-distribution (unseen VOC) prediction tasks. The Pearson correlation coefficient between predicted and true results exceeds 0.5 in most datasets, reaching up to 0.8.

More importantly, we conducted functional experimental validation using an independent pest species, Bactrocera dorsalis. Results showed that the positive hit rate from model predictions in the top predictions can reach 2 times the random level (enrichment factor = 2), confirming its cross-species prediction capability.