Drug Response Prediction

Pathway-level stratification showed gains for scpFormer across nearly all drug pathway categories. At the global cell-line and drug prediction level, scpFormer achieved Pearson correlation 0.921, compared with 0.888 for scFoundation. Among 223 tested drugs, 183 had lower mean absolute error with scpFormer than with scFoundation. Drug response prediction performance was evaluated with Pearson correlation between predicted and true IC50 and mean absolute error across drugs. The paper interprets improved IC50 prediction as evidence that scpFormer embeddings capture cellular states related to pharmacological response. Understanding why some individuals benefit from fasting while others do not is a key scientific challenge. The hypothesis is that non-responders share a distinctive microbiota profile that can be identified and used for pre-intervention outcome prediction. scpFormer embeddings were used in DeepCDR for cancer drug response prediction from paired cancer cell-line proteomic profiles and IC50 data. The trial data collection is designed to distinguish responders from non-responders based on microbiota profiles.