QSARs

Quantitative structure–activity relationships (QSARs) are mathematical models that relate chemical structure to a biological or physicochemical property. They are used to predict the activity or properties of new compounds based on data from known compounds, enabling exploration of chemical space without direct testing. The approach originated in the 1960s with foundational work by Corwin Hansch and colleagues, who established the idea of correlating substituent constants and other descriptors with activity.

A typical QSAR workflow begins with data collection and curation of a dataset of compounds with known activities. Molecules are encoded into descriptors that capture structural and physicochemical information, such as logP, molecular weight, hydrogen-bond donors and acceptors, topological indices, quantum-chemical properties, or binary fingerprints. The chosen descriptors are paired with the target property to build a statistical or machine learning model. Common modeling approaches include linear methods like multiple linear regression and partial least squares, as well as nonlinear techniques such as random forests, support vector machines, and neural networks. Model performance is assessed using internal validation (e.g., cross-validation) and external validation, with metrics such as R-squared, Q-squared, RMSE, and predictive accuracy.

An important concept is the applicability domain, the chemical space over which the model’s predictions are considered reliable. QSARs are widely used in drug discovery, toxicology, and environmental risk assessment to prioritize candidates, screen large libraries, and support decision making. Regulatory acceptance follows guidelines such as the OECD principles for QSAR validation, which require defined endpoints, a transparent algorithm, a defined domain of applicability, appropriate measures of goodness-of-fit and predictivity, and, when possible, mechanistic interpretation. Limitations include data quality, overfitting, descriptor selection, and extrapolation beyond the model’s domain.