Cell-line drug sensitivity must reach patients: MODAE+ aligns the two domains while keeping tumour lineage, and its predicted sensitivity tracks survival
Synopsis
Transferring drug sensitivity from cancer cell lines to patient tumours is hard because aligning the two domains can erase tumour-lineage structure. MODAE+ jointly optimizes domain alignment, deconfounding, drug-response prediction, survival prediction and lineage classification in one shared latent space, with adaptive multitask weighting to balance them. On matched treated TCGA subsets, its predicted sensitivity reached a 41.5% Cox-significant rate for overall survival, versus 4.9% for CODE-AE-ADV.
Interpretation
Cross-domain drug-sensitivity transfer can now keep tumour-lineage structure: cell lines and patient tumours group by tissue of origin in the shared latent space instead of collapsing into an unstructured mixture. Earlier alignment methods, including ComBat, Celligner and CODE-AE-ADV, mainly reduced global domain differences without constraining alignment to biologically corresponding cancer lineages; CODE-AE-ADV removed the domain signal by collapsing samples into a continuous, regular structure. Classifying latent samples by tissue of origin with a k-nearest-neighbour classifier gave MODAE+ a balanced accuracy of 0.92 versus 0.15 for CODE-AE-ADV; ablations removing lineage supervision or deconfounding sharply reduced lineage recovery, and removing adaptive loss balancing destabilized performance across tissues.
The transferred sensitivity scores retain survival-related information in real treated populations. On the same treated TCGA subsets evaluable for both methods, the earlier approach's sensitivity scores showed little association with survival in a paired comparison. For overall survival MODAE+ reached a mean C-index of 0.684 versus 0.506 for CODE-AE-ADV, with Kaplan–Meier significance in 46.3% versus 2.4% of pairs and Cox significance in 41.5% versus 4.9% (41 pairs); for progression-free interval the C-indices were 0.617 versus 0.549, with significance rates of 32.4% versus 5.4% and 10.8% (37 pairs).
The model encodes cell-line and patient expression profiles into one latent space, where a shared part carries transfer and private parts absorb dataset-specific variation. The shared code is aligned across domains with Wasserstein adversarial training, and a deconfounding norm penalty enforces orthogonality between shared and private codes; the shared representation then drives three supervised heads for drug sensitivity, survival and lineage. Training used about 9,696 TCGA patient tumours, about 1,293 CCLE cell lines, and about 321,567 drug–cell-line pairs from 544 compounds screened across 738 cell lines in CTRP; FAMO dynamically weights the eight optimization objectives so no single one dominates training.
Perspective
The demonstrated scope is pan-cancer cell-line–tumour integration plus survival associations in treated TCGA subsets, drug-target expression differences, and concordance with external breast cancer signatures for paclitaxel and doxorubicin; the authors state that direct clinical predictive accuracy still requires evaluation against observed treatment outcomes.
The external signature validation is post hoc, limited to the selected breast cancer treatments, and its cohorts overlap and include combination regimens, limiting independence and drug specificity; predictive reliability may also vary with cancer type and with how well a drug is represented in CCLE and CTRP.