Fourier transform infrared (FT-IR) spectroscopy has become a powerful tool for biodiagnostics and cell line classification. Typical experimental perturbations included in spectra are baseline shift and scale variation between spectra. They have to be removed by data preprocessings to allow further data analysis and classification. In this work, we addressed baseline shift corrections and normalizations in attenuated total reflection (ATR) FT-IR spectra. We compared the efficiency of several preprocessing methods with series of spectra containing typical perturbations (baseline shift, scaling factor, and noise) and a priori known definite spectral difference. Several baseline-correction and normalization possibilities were evaluated. Our results were generally sensitive, selective, and robust with respect to baseline and scaling. Full-range scaling generated more false-positive results. Use of first- and second-derivative spectra was tested. Results obtained on model spectra were confirmed with series of spectra from sensitive and multidrug-resistant leukemia K562 cells. We showed that the use of derived spectra did not provide more efficiency and required additional preprocessing such as smoothing to obtain results similar to those obtained from non-derived ones. On the other hand, results obtained with derivatives were less sensitive to scaling, a useful feature when scaling is problematic.