State variables throughout non-induced and induced cultivations of Pichia pastoris for the heterologous Rhizopus oryzae lipase (ROL) production were monitored with a multi-wavelength on-line fluorescence sensor. Based on this work, the use of in situ multi-wavelength fluorometry combined with chemometrics models (PLS-1 models) provided a quantitative prediction of biomass and substrates (glycerol and methanol) during non-induced and induced ROL production. The mean prediction errors for both variables were about 7% and 10%, respectively. ROL is also quite satisfactory estimated in the exponential growth phase with prediction errors similar to biomass and substrate variables. However, in the stationary phase, where proteolytic degradation of ROL is observed, the prediction error could get a value about 20%. This fact is due to the lower reproducibility of protein production from batch to batch.