Fig 1.
The operating principle of APDCA.
APDCA iteratively optimizes the low-rank matrices (Gi) of multiple relational data matrices through matrix tri-factorization, and updates (Gi) by selecting the top k relationships based on sparsity. Then construct an extrapolated point Yi. Then iterate Sij, Yi by LS-1 if the condition is satisfied, otherwise iterate Sij and Gi by LS-2. Finally reconstruct the relation matrix and get association scores and rank.
Table 1.
Initialization of variables in the APDCA algorithm.
Fig 2.
Schematic illustration of the multi-type biological network used in APDCA.
Six types of biological entities are represented: RNA-binding proteins (RBPs), miRNAs, genes, alternative splicing (AS) events, diseases, and drugs. Edges between entities denote observed associations Rij with corresponding weights Wij. Self-loop parameters and
are used for gene and drug regularization, respectively.
Fig 3.
Comparison of predictive performance under 5-fold cross-validation.
(a) ROC curves with AUC values. (b) Precision-Recall curves with AUPR values. APDCA achieves the highest scores across both metrics.
Fig 4.
Convergence comparison of APDCA, SIMCLDA, MLFDA, and SDLDA over 100 iterations using relative error as the evaluation metric.
Fig 5.
Network showing the RBP–AS event associations during EMT.
Table 2.
Top 10 QKI–AS Event Associations Predicted by APDCA and Their Evidence in OncoSplicing.
Fig 6.
AUC surface with respect to regularization parameters λ and .
Performance improves as both parameters decrease and stabilizes at low values.
Table 3.
AUC values under different combinations of regularization parameters λ and .
Fig 7.
AUC sensitivity of APDCA to the sparse control parameter k under 5-fold cross-validation.
Performance peaks at k = 75 and remains stable across a wide range.