PCAM: proteochemometric assay-aware modeling for robust bioactivity estimation and uncertainty quantification
Year: 2026
Type: article
Abstract: Public bioactivity databases aggregate measurements from heterogeneous experimental assays, introducing systematic variability that can degrade model performance and distort uncertainty estimates. We present proteochemometric assay-aware modeling (PCAM), an extension of proteochemometrics that integrates assay descriptors alongside compound and protein features to capture assay context. Using a curated ChEMBL-derived benchmark spanning 167 G protein-coupled receptor targets, 77,081 compounds, and 5,161 assays, we compare models that omit assay information to models using either categorical assay identifiers or text-derived embeddings from assay descriptions. We evaluate generalization under three regimes of increasing distribution shift (random, scaffold-based, and assay-held-out splits) and assess both point accuracy and probabilistic quality. Overall difficulty is dominated by the split strategy, with assay-held-out evaluation inducing the strongest distribution shift. When assay identifiers are shared between training and evaluation, categorical assay descriptors provide the most consistent gains in accuracy and probabilistic scoring relative to models without assay context. Under assay-held-out evaluation, differences between assay representations contract, while text-derived embeddings show a consistent but typically small advantage. This regime corresponds to practical assay-transfer scenarios, such as protocol or vendor changes, new readouts, or sparse early measurements for a newly introduced assay, where leveraging assay text/metadata can help transfer signal from related assays. In addition, assay representations can reduce reliance on brittle cross-assay normalization in heterogeneous public datasets by allowing the model to learn assay-dependent shifts directly; however, we observed optimization instability for one configuration (a probabilistic neural network using BioBERT embeddings under the random split). We introduce an assay-aware extension of proteochemometrics that explicitly models assay context and evaluate it under increasing distribution shift. We clarify when simple assay identifiers suffice and when text-derived assay representations better support transfer to unseen assay protocols while maintaining uncertainty-aware evaluation.
Source: Journal of Cheminformatics
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,390
OpenAlex ID: W7224435772