Although the exposure to PM2.5 has serious health implications, indoor PM2.5 monitoring is not a widely applied practice. Regulations on the indoor PM2.5 level and measurement schemes are not well established. Compared to other indoor settings, PM2.5 prediction models for large office buildings are particularly lacking. In response to these challenges, statistical models were developed in this paper to predict the PM2.5 concentration in well-mixed indoor air in a commercial office building. The performances of different modeling methods, including multiple linear regression (MLR), partial least squares regression (PLS), distributed lag model (DLM), least absolute shrinkage selector operator (LASSO), simple artificial neural networks (ANN), and long-short term memory (LSTM), were compared. Various combinations of environmental and meteorological parameters were used as predictors. The root-mean-square error (RMSE) of the predicted hourly PM2.5 was 1.73 μg/m3 for the LSTM model and in the range of 2.20-4.71 μg/m3 for the other models when regulatory ambient PM2.5 data were used as predictors. The LSTM models outperformed other modeling approaches across the performance metrics used by learning the predictors' temporal patterns. Even without any ambient PM2.5 information, the developed models still demonstrated relatively high skill in predicting the PM2.5 levels in well-mixed indoor air.