Widespread false negatives in DNA-encoded library data: how linker effects impair machine learning-based lead prediction

DNA-encoded chemical libraries (DECLs) have become integral to early-stage drug discovery, yielding active compounds and extensive labeled datasets for machine learning (ML)-based prediction of bioactive molecules. However, the information content of DECL selection data remains scarcely explored. This study systematically investigates for the first time the prevalence of false negatives and the influence of the linker in DECL data. Using a focused DECL targeting the poly-(ADP-ribose) polymerases PARP1/2 and TNKS1/2 as a model system, we found that our DECL selections frequently miss active compounds, with numerous false negatives for each identified hit. The presence of the DNA-conjugation linker emerged as a factor contributing to the underdetection of active molecules. This bias toward false negatives compromises the predictive power of DECL data for prioritizing hits, anticipating target selectivity, and training ML models, as determined by analyzing the effects of undersampling and oversampling techniques in learning the PARP2 data. Conversely, the linker's presence in DECLs offers advantages, such as enabling the identification of target-selective protein engagers, even when the underlying molecules themselves may not be selective. These findings highlight the challenges and opportunities of DECL data, emphasizing the need for best practices in data handling and ML model development in drug discovery.