Prediction of lncRNA–cervical cancer associations using a multi-view collaborative variational autoencoder with knowledge distillation transfer learning

Abstract Long non-coding RNAs (lncRNAs) play crucial regulatory roles in the pathogenesis of cervical cancer. Accurate identification of lncRNA-cervical cancer associations is of great significance for understanding disease molecular mechanisms and discovering potential therapeutic targets. However, the extreme scarcity of cervical cancer-specific lncRNA association data has become the core bottleneck constraining the performance of existing computational prediction methods. To address this challenge, we propose a Multi-View Collaborative Variational Autoencoder based on Knowledge Distillation Transfer Learning (MVCVAE), which innovatively integrates the two technical advantages of multi-source heterogeneous data fusion and cross-disease knowledge transfer. The multi-view collaborative variational autoencoder constructs dual-view feature representations of lncRNA similarity and disease similarity, utilizing symmetric encoder-decoder architecture to learn complex lncRNA-disease association patterns in latent space. The knowledge distillation transfer learning framework adopts a teacher-student network design, where the teacher model is pre-trained on a large-scale dataset containing 12,865 lncRNA-disease associations to learn universal association patterns, and effectively transfers rich prior knowledge to the student model targeting 92 cervical cancer associations through feature distillation and prediction distillation strategies. In five-fold cross-validation, MVCVAE achieved an AUC of 0.9228±0.0058. Through the synergistic effect of multi-view heterogeneous information fusion and cross-domain knowledge transfer, MVCVAE achieves excellent performance in lncRNA-cervical cancer association prediction, effectively alleviating the bottleneck constraint of cervical cancer-specific lncRNA association data scarcity.