A-S.B.20: DanSyn: Hybrid Structural-Functional Representations for Drug Synergy Prediction
Accurate prediction of synergistic drug combinations can accelerate the discovery of
cancer combination therapy, but the space of drug pairs across cell lines is far too
large to screen exhaustively, and existing deep models remain computationally expensive
for large-scale screening. We present DanSyn, a multimodal predictor built around a
hybrid structural-functional design. DanSyn represents each drug pair and cell line
through three complementary views — substructure token embeddings from an ESPF
Transformer, an unsupervised multi-omics cell line latent, and a language-model
embedding of each drug's mechanism of action — and fuses them with a cross-attention
module in which the cell context attends over drug substructures. Under a single
pre-committed preprocessing protocol applied identically to three public benchmarks
(DrugCombDB, O'Neil and NCI-ALMANAC), DanSyn attains the best in-distribution accuracy
on all three, ahead of classical tree ensembles, a molecular-graph network and strong
deep baselines retrained under the same protocol, while training several-fold faster
than transformer-based alternatives. Ablations show that predictive accuracy is
dominated by the multi-omics cell context: removing it costs far more than removing
either drug view, while the cell-conditioned fusion contributes a small but consistent
gain. On the 763 drug pairs whose measured synergy flips sign across cell lines, DanSyn
predicts the direction of the shift correctly for 92%, indicating that predictions
respond to cell context rather than to drug-pair identity alone. This study addresses
in-distribution prediction on dense screens; extrapolation to entirely unseen drugs or
cell lines is outside its scope.
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826