Machine-learning based classification of 2D-IR liquid biopsies enables stratification of melanoma relapse risk

Abstract

Non-linear laser spectroscopy methods such as two-dimensional infrared (2D-IR) produce large, information-rich datasets, while developments in laser technology have brought substantial increases in data collection rates. This combination of data depth and quantity creates the opportunity to unite advanced data science approaches, such as Machine Learning (ML), with 2D-IR to reveal insights that surpass those from established data interpretation methods. To demonstrate this, we show that ML and 2D-IR spectroscopy can classify blood serum samples collected from patients with melanoma according to diagnostically-relevant groupings. Using just 20 μL samples, 2D-IR measures 'protein amide I fingerprints', which reflect the protein profile of blood serum. A hyphenated Partial Least Squares-Support Vector Machine (PLS-SVM) model was able to classify 2D-protein fingerprints taken from 40 patients with melanoma according to the presence, absence or later development of metastatic disease. Area under the receiver operating characteristic curve (AUROC) values of 0.75 and 0.86 were obtained when identifying samples from patients who were radiologically cancer free and with metastatic disease respectively. The model was also able to classify (AUROC = 0.80) samples from a third group of patients who were radiologically cancer-free at the point of testing but would go on to develop metastatic disease within five years. This ability to identify post-treatment patients at higher risk of relapse from a spectroscopic measurement of biofluid protein content shows the potential for hybrid 2D-IR-ML analyses and raises the prospect of a new route to an optical blood-based test capable of risk stratification for melanoma patients.

Fig. 1. Average 2D-IR spectra obtained with T_w = 250 fs for each of the different patient classification groups (a) control (Cont, 48 spectra), (b) developed metastasis (Dev, 33 spectra) and (c) metastatic (Met, 63 spectra). Difference spectra determined for subtraction of the averaged spectra of the three classes from each other are shown in panels (d) metastatic – control, (e) developed metastasis – control and (f) metastatic – developed metastasis. (g) to (i) show the difference spectra from (d)–(f) expanded by the multiplication factor shown.

Fig. 2. Average performance metric outputs across outer folds for test samples, (a) average confusion matrix, (b) average ROC curves for each class: control (green), developed metastasis (gold) and metastatic (pink) (c) prediction probability box plots across all outer folds for each class, with jitter points added showing each individual probability value and range. Clear separation of the target category from the others show the model's confidence in producing a classification.

Fig. 3. (a) Average variable importance of projection scores across training set of outer folds. VIP scores greater than 1 indicate important latent variables used in making model predictions. Panels (b) to (d) shows the spectral loadings of the three variables considered to be the most important, (b) = LV 15, (c) = LV 13, (d) = LV 14. Panels (e–g) reproduce the difference spectra from Fig. 1(g–i) for comparison. Coloured arrows highlight points of interest as discussed in the text.