ComplexMatrixComb: Predicting Drug Combination IC50 Doses via Complex Numbers and Matrix Factorization
Abdollahi, M.; Mirzaei Koli, A.; Ghiam, S.; Eslahchi, C.
Show abstract
Determining precise drug concentrations to inhibit cancer cell growth is a critical but resource-intensive challenge, especially for combinations requiring many dose pairs. Existing computational methods often predict synergy or classify interactions but rarely estimate exact concentration pairs for a defined inhibitory effect. We introduce ComplexMatrixComb, a matrix factorization framework predicting drug combination dosages achieving IC50 for specific cancer cell lines. It encodes drug concentrations as real and imaginary components of complex numbers, modeling joint dose-response dynamics and enabling accurate IC50 estimation. Across ONeil, NCI-ALMANAC, and AZ-Dream datasets, ComplexMatrixComb outperformed traditional machine learning in regression and classification, showing robustness to drug order and adaptability to varied designs. Integration with models such as ComboFM and ComboLTR showed predicted concentrations closely matched ground-truth doses. Experimental validation of five drug pair-cell line predictions using MTT assays confirmed 50% growth inhibition. By reducing reliance on exhaustive screens, ComplexMatrixComb offers a scalable, data-driven tool to accelerate preclinical research and support personalized oncology strategies.
Matching journals
The top 6 journals account for 50% of the predicted probability mass.
Similar papers in this journal
Similar papers in this journal
- A regularized functional regression model enabling transcriptome-wide dosage-dependent association study of cancer drug response 97%
- Controlling astrocyte-mediated synaptic pruning signals for schizophrenia drug repurposing with Deep Graph Networks 96%
- Tissue-guided LASSO for prediction of clinical drug response using preclinical samples 96%
Similar papers in this journal
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.