Predicting targeted- and immunotherapeutic response outcomes in melanoma with single-cell Raman spectroscopy and AI
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (Arcadia Science)
Abstract
PURPOSE
Identifying reliable predictors of immunotherapeutic response in melanoma remains an outstanding challenge. Existing transcriptomic and proteomic profiling methods for the tumor-immune microenvironment (TIME) are costly and may not faithfully capture modifications actively impacting tumor behavior. Here, we present a non-destructive, single-cell approach combining Raman spectroscopy and machine learning (ML) that enables rapid cell profiling and therapeutic response prediction.
METHODS
We analyzed single-cell Raman spectra of mouse and human melanoma cell lines alongside nine melanoma patient-derived samples with known resistance profiles to targeted and immunotherapeutic inhibitors bemcentinib, cabozantinib, dabrafenib, nivolumab, and a combination of nivolumab and relatlimab. We assessed cell phenotyping classification and treatment resistance using random forests and feature importance analysis. For patient samples, we constructed a two-stage evaluation workflow to determine clinical drug resistance through aggregated single-cell predictions and identified corresponding highly variant spectral signatures using computational methods adapted from single-cell RNA sequencing methods.
RESULTS
In cell lines, our approach achieved >96% differentiation accuracy across tumor microenvironment cell types and induced functional phenotypes. Persistent (drug-resistant) cells formed subclusters based on genetic mutations rather than sample origin, with Raman signatures reflecting biochemical changes relevant to therapeutic pathways. For patient samples, our workflow correctly inferred resistance likelihoods for 30 of 33 clinically-relevant patient-drug combinations (91% accuracy).
CONCLUSION
Single-cell Raman spectroscopy combined with machine learning offers a scalable, prognostic platform to predict therapeutic resistance likelihood, with further potential to advance clinical, multi-omic biomarker efforts for melanoma. Our approach may improve first-and second-line therapy selection assessments for precision medicine by providing rapid, non-destructive prediction of therapeutic response based on cellular spectral profiles.
Context Summary
Key objective
Can label-free, single-cell Raman spectroscopy and machine learning approach accurately profile melanoma cell states and therapeutic resistance likelihood to targeted and immunotherapeutic agents?
Knowledge generated
Raman spectroscopy with machine learning differentiated tumor microenvironment cell types and functional phenotypes with >96% accuracy in cell lines. When applied to patient-derived metastatic melanoma samples, the approach correctly inferred patient response to a panel of targeted and immunotherapeutic inhibitors with 91% accuracy (30 of 33 cases). High-likelihood persistent and sensitive cells across diverse patients exhibited recurrent spectral features.
Relevance
Single-cell Raman-based profiling supports functional-diagnostic assessment or resistance likelihood and may contribute to improved therapeutic selection and precision oncology strategies for melanoma patients.
Article activity feed
-
The data that support the findings in the work are available in the article and the supplementaryinformation file
I was not able to find the data supplement in the most recent preprint
-
RF achieved 96% classification accuracy discriminatingcell lines, with >91% accuracy for parent and derivative lines (Figure 2b)
This is impressive; however, it would also be beneficial to rule out batch effects. The supplementary methods say that " Duplicate wells were used for both treated and untreated cells". I assume the data from the duplicates were combined but it would be meaningful to run a classifier to predict each duplicate. If the classifier is truly learning cell type-specific features, it should fail at predicting the duplicate.
-
Supplementary Figure 3
in this figure, why are the "counts" very different day to day? e.g. in C, the counts range from 100 to >2,000. Is this a result of a different number of points being collected on each day or different numbers of spectra being excluded as outliers?
-
B) Biological Raman window, also known as the optical fingerprint regime, of a cell. Averagedspectra of RAW264.7 macrophages, with major band assignments highlighting cellular makeup. Differential Raman“expression” profiles or spectral barcodes generated from cell spectra are used to delineate samples and drugresponse. C) UMAP embeddings of melanoma spectra are analyzed by nonlinear dimensionality reduction toexamine cell states and phenotypes.
It's not clear from a first pass what aspects of this figure are real collected data vs a representative diagram. e.g. the Raman spectrum looks real, the UMAP looks like a diagram, and the spectral barcode looks like a diagram, but it's not completely clear.
-
