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

Area of Science:

• Drug Discovery
• Medicinal Chemistry
• Computational Chemistry

Background:

• DNA-encoded chemical libraries (DECLs) are crucial for identifying bioactive molecules in early-stage drug discovery.
• DECLs generate large datasets for machine learning (ML) model development.
• The information content and potential biases within DECL selection data are not fully understood.

Purpose of the Study:

• To systematically investigate the prevalence of false negatives in DECL selections.
• To determine the influence of the DNA-conjugation linker on the detection of active compounds.
• To assess the impact of DECL data biases on ML model performance for drug discovery.

Main Methods:

• Utilized a focused DECL targeting PARP1/2 and TNKS1/2 enzymes as a model system.
• Analyzed DECL selection data to quantify false negatives and identify underdetected active compounds.
• Evaluated the effect of the DNA-conjugation linker on compound detection.
• Applied undersampling and oversampling techniques to assess ML model performance using PARP2 data.

Main Results:

• DECL selections frequently yielded false negatives, missing numerous active compounds.
• The DNA-conjugation linker was identified as a factor contributing to the underdetection of active molecules.
• False negatives compromised the predictive power of DECL data for hit prioritization and ML model training.
• The linker also enabled the identification of target-selective protein engagers.

Conclusions:

• DECL data contains significant biases, particularly false negatives, impacting its utility in drug discovery.
• The DNA-conjugation linker presents both challenges (underdetection) and opportunities (selectivity identification) in DECLs.
• Best practices for data handling and ML model development are essential for maximizing the value of DECL data.