Chemical recognition using machine learning based on detection by gas sensors relies on the accuracy and sensitivity of the sensors at capturing the key features of target classes. In some cases, however, the electronic signal transduced from the detection of analytes does not completely represent the key attributes, resulting in inaccurate classification results when trained from signal data alone. To overcome this shortcoming, we propose a novel “chemistry-informed” machine learning framework composed of two modules. From available sensor response data, Module 1 identifies and predicts the chemical properties of the analytes that give rise to the sensitivity and selectivity of the sensors, and Module 2 performs final classifications using the dataset concatenating predicted chemical properties and raw sensor responses. To evaluate the performance and generalizability of our methodology, we conducted experiments with three gas sensor array datasets for gas detection. In all the cases, the performance of gas species classification was improved when the raw features were combined with the predicted chemical property features. The main contribution of our framework is that it bridges the gap between the gas sensor signals and the target analytes, thereby improving classification performance beyond that of models trained exclusively on sensor response data.