Early Prediction of Donor-Dependent Expansion Capacity in NK Cells Using FTIR Spectroscopy and Machine Learning

Introduction:Natural killer(NK) cells are candidates for cancer immunotherapy due to their antitumor-activity and lower risk of graft-versus-host disease than T-cell therapies. However, NK-cell manufacturing is constrained by donor-to-donor variability in expansion capacity, evident after culture days. This limitation reduces opportunities for early-process adjustment, making identification of high-performing donors crucial for improving process efficiency and product consistency.

Objectives:This study combines Fourier-transform infrared(FTIR) spectroscopy and machine-learning to identify molecular signatures in culture media associated with NK-cell expansion potential. The goal is to predict donor-dependent expansion performance at early culture stages, before significant cell-proliferation occurs.

Methods:NK-cells were isolated from umbilical cord blood samples under legal procedures. Culture media samples were analysed using an FTIR-spectrometer across 400–4000cm⁻¹. First-derivative-normalization was applied as a preprocessing method, and supervised classification models were developed using 578 best features from preprocessed spectra, selected using the Information-Gain-algorithm.

Results:Culture samples from ten donors were grouped based on principal component analysis(PCA), generating two donor groups and a third group containing the remaining donors. Using samples collected between days 9 and 20, a k-nearest neighbours model achieved perfect classification (F1-score=1.000). Using early-stage samples (days 0-9), an AdaBoost model achieved an F1-score of 0.896, demonstrating early prediction of donor-related behaviour. Donors were grouped according to a fold-change expansion threshold of 15, generating classes of four and six donors, respectively. For samples collected between days 9 and 20, a Naïve Bayes model achieved an F1-score of 0.958, while an SVM model trained on samples from the first nine culture days achieved an F1-score of 0.867.

Conclusions:FTIR-spectroscopy combined with machine-learning can classify NK-cell donors according to expansion potential, including at early-culture-stages. These findings highlight the potential of this cost-effective approach to support donor selection and process optimization in NK-cell immunotherapy manufacturing.