Discrimination of molecular subtypes of breast cancer with ATR-FTIR spectroscopy in blood plasma coupled with partial least square-artificial neural network discriminant analysis (PLS-ANNDA)

Abstract

Attenuated total reflection-Fourier transform infrared (ATR-FTIR) spectroscopy is a simple, rapid, reagent-free, and non-invasive technique that was used as bioscreening tool for breast cancer (with blood plasma) in this study. Three partial least square-artificial neural network discriminant analysis (PLS-ANNDA) models were created (n = 74): i) to differentiate cancer (n = 56) from non-cancer subjects (control, n = 18); ii) grouping the molecular subtypes considering the therapeutic options: differentiating control from Luminal A (LA, n = 32) + Luminal B (LB, n = 10) and HER2 (n = 12) + Triple-negative (TN, n = 3). iii) differentiating control and molecular subtypes individually (control vs. LA vs. LB vs. HER2 vs. TN). The sensitivity (%)/specificity (%) for the three models are as follows: i) control (100/100) and breast cancer (100/100); ii) control (100/100), LA + LB (100/70), and HER2 + TN (40/100); iii) control (100/100), LA (66.7/76.9), LB (50/94.4), HER2 (75/94.4), and TN (0/100). In spectral analysis, four intervals in the biofingerprint were identified by the Kruskal-Wallis test with significant difference (p < 0.05) for the molecular subtypes and control (cm−1): 1800–1700 [ν(C=O)], 1437–1326 [νs(COO−) + δs(CH3)], 1236–934 [ν(C–O) + ν(C–C) + ν(CH2OH) + δ(C–O) + νas(CO–O–C) + νs(PO2−)], and 919–900 (ribose ring). Furthermore, the 1739 cm−1 peak was verified to be the most sensitive for screening (breast cancer vs. control) directly by absorbance normalized between 0 and 1 [AUC = 0.82 (0.70–0.93)]. Therefore, ATR-FTIR coupled with PLS-ANNDA is a promising screening option for breast cancer. It is feasible to discriminate between control and breast cancer as well as individually discriminate molecular subtypes, when a dynamic database is created with a routine for inserting new samples to improve the model.