Measurement noise is random or systematic variation in observed values that does not represent the underlying property. In drug discovery, assay protocols, instruments, reagent lots, and experimental conditions can change measured ADMET properties substantially, even for the same molecule. section 13.4.3
Noise becomes a modeling problem when heterogeneous measurements are combined. Property-prediction models may then produce uncertain predictions because the training labels vary experimentally, rather than because the algorithm is intrinsically inadequate. section 13.4.3
In high-throughput screening, experimental noise interacts with other practical difficulties:
- Large assay collections contain greater experimental noise and more diverse chemical structures, complicating structure–activity relationship analysis. section 7.1.1
- High-dimensional measurements can obscure meaningful patterns for traditional QSAR modeling; dimensionality reduction can organize the chemical space while preserving structural relationships. section 7.1.1
The same molecule, different laboratories
A molecule's measured ADMET property can differ by an order of magnitude between laboratories. Differences in assay protocol, instrumentation, reagent lot, or experimental conditions can create variation that the model learns as if it were molecular signal. section 13.4.3















