DeepDTF: Dual-Branch Transformer Fusion for Multi-Omics Anticancer Drug Response Prediction
Authors: Yuhan Zhao, Jacob Tennant, James Yang, Zhishan Guo, Young Whang, Ning Sui
Organizations: Department of Computer Science, North Carolina State University, Raleigh, NC, USA · UNC School of Medicine, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA · Department of Molecular and Structural Biochemistry, North Carolina State University, Raleigh, NC, USA
Cancer drug response varies widely across tumors due to multi-layer molecular heterogeneity, motivating computational decision support for precision oncology. Despite recent progress in deep CDR models, robust alignment between high-dimensional multi-omics and chemically structured drugs remains challenging due to cross-modal misalignment and limited inductive bias. We present DeepDTF, an end-to-end dual-branch Transformer fusion framework for joint log(IC50) regression and drug sensitivity classification. The cell-line branch uses modality-specific encoders for multi-omics profiles with Transformer blocks to capture long-range dependencies, while the drug branch represents compounds as molecular graphs and encodes them with a GNN-Transformer to integrate local topology with global context. Omics and drug representations are fused by a Transformer-based module that models cross-modal interactions and mitigates feature misalignment. On public pharmacogenomic benchmarks under 5-fold cold-start cell-line evaluation, DeepDTF consistently outperforms strong baselines across omics settings, achieving up to RMSE=1.248, R^2=0.875, and AUC=0.987 with full multi-omics inputs, while reducing classification error (1-ACC) by 9.5%. Beyond accuracy, DeepDTF provides biologically grounded explanations via SHAP-based gene attributions and pathway enrichment with pre-ranked GSEA.
Figures & tables
Fig. 1: Overall workflow of the DeepDTF, including (i) multi-omics and drug-structure inputs, (ii) transformer model (detailed in Fig. 2 ), (iii) dual task prediction (IC50 regression and sensitivity classification), and (iv) an interpretability module based on SHAP → GSEA. SHAP produces a signed gene ranking, which is used for pre-ranked GSEA: The top panel illustrates a GSEA enrichment plot (running ES with pathway “hit” positions along the ranked list), and the bottom panel shows the corresponding null/random ES distribution for assessing enrichment direction and significance.
Fig. 2: Detailed architecture of DeepDTF. The left block is a dual-branch feature extractor for multi-omics and drug structure. The right block is a feature fusion joint prediction module with a Fusion-Transformer and two task-specific heads for IC50 regression and sensitivity classification.
Setting
Method
RMSE
R2
PCC
ACC
SEN
SPEC
AUC
MUT+CNV
tCNN [ 13 ]
1.459
0.725
0.851
0.952
0.706
0.977
0.963
DeepCDR [ 14 ]
1.401
0.770
0.878
0.953
0.732
0.975
0.969
DeepTTA [ 17 ]
1.420
0.775
0.880
0.953
0.732
0.978
0.970
DeepDTF
1.403
0.780
0.883
0.954
0.738
0.978
0.975
MUT+CNV+GE
CDRscan [ 12 ]
1.483
0.715
0.846
0.950
0.667
0.979
0.960
tCNN [ 13 ]
1.459
0.725
0.851
0.952
0.706
0.977
0.963
TABLE I: Evaluation of the Proposed Method and Baselines
Fig. 3: Ablation studies on omics integration and model components, where (a) and (c) show the regression performance under different combinations of omics inputs and model variants; while (b) and (d) report the classification metrics obtained from various omics integration strategies and model variants.
Fig. 4: Cartesian-regime biological explanations via SHAP–GSEA. (a) Top genes with positive SHAP (sensitivity-associated). (b) Bottom genes with negative SHAP (resistance-associated). (c) Top enriched Hallmark pathways from signed-SHAP rankings (positive ES; sensitivity-associated programs). (d) Top enriched Hallmark pathways from negative signed-SHAP rankings (negative ES; resistance-associated programs)
Interdisciplinary Program in Bioinformatics, Seoul National University, Seoul, Republic of Korea · AIGENDRUG Co., Ltd., Seoul, Republic of Korea · BK21 FOUR Intelligence Computing, Seoul National University, Seoul, Republic of Korea +2