RamanMAE

RamanMAE applies masked autoencoding to Raman spectra of biological samples. Raman spectroscopy reads out the chemical composition of cells and tissue label-free, capturing information that stained histopathology discards — but only about one in a million photons scatters inelastically, so spectra are slow to acquire and noisy when they are not. That combination has kept deep learning largely out of the field: the datasets needed to train a network from scratch are expensive to collect, and models built for one study rarely carry over to another.

The model was developed by Santosh Kumar Paidi and Parul Maheshwari, with Paidi affiliated with Genentech. It was posted as a preprint in May 2025 and published in Analytical Chemistry in November 2025. The design choice that defines it is deliberate restraint: rather than engineer a bespoke spectral architecture, the authors take the standard ViT-MAE and change only what is necessary — input shape, layer sizes, and embedding dimensions — so that the result is a turnkey recipe a spectroscopy group can adopt without masked-autoencoder expertise.

RamanMAE sits alongside masked-autoencoder models in adjacent measurement domains, notably OpenPhenom-S/16 for Cell Painting microscopy, and complements text-based spectral interpreters such as SpectraLLM. It learns directly in the spectral domain, outputting both a denoised reconstruction and a low-dimensional embedding.

#Key Features

• 1D spectra as 2D patches: A 400-point fingerprint-region spectrum is wrapped into a 20×20 image so that contiguous 1D spectral patches map onto raster-scanned 2D patches, making the unmodified ViT-MAE patch and positional encoding directly applicable to one-dimensional data.
• Denoising by reconstruction: The decoder acts as a learned smoother. Reconstructions from low-SNR spectra reach a cosine similarity of 0.98 with reconstructions from the matched high-SNR spectra, matching or slightly beating a tuned Savitzky-Golay filter without its sensitivity to window and polynomial choices.
• Biologically structured embeddings: k-means clustering in the encoder embedding space separates nucleus, cytoplasm, and cell membrane in single-cell hyperspectral maps, and does so on low-SNR spectra where clustering the raw spectra confuses membrane with nucleus.
• Cross-application transfer: A model trained on one dataset produces useful embeddings for a different biological question at a different spatial scale, provided the spectra were acquired over the same wavenumber range.
• Small and reproducible: Encoder projection dimension 64, decoder projection dimension 32, three encoder layers and one decoder layer — trained on a single T4 GPU in Google Colab.

#Technical Details

Pretraining used published Raman datasets rather than newly collected data. The primary corpus is the DeepR dataset of over 159,000 cell spectra paired at high and low signal-to-noise ratio with their spatial map coordinates; supporting datasets include over 20,000 single-cell spectra spanning parental MDA-MB-231 breast cancer cells, matched circulating tumour cells, and lung metastatic cells; 16,500 spectra from mouse tumours of varying metastatic potential; 5,000 spectra from mice treated with anti-PD-L1 immunotherapy; and an articular cartilage tissue dataset. Training used a patch size of 16 points, a masking ratio of 0.5, batch size 256, mean absolute error loss, and the AdamW optimiser in TensorFlow with a Keras backend. Reconstruction loss converged to 0.040 training and 0.047 validation MAE, with a test PSNR of 27.88 and little sensitivity to masking ratios between 0.25 and 0.75.

Downstream evaluation used random forests on frozen embeddings. Per-spectrum classification of the three metastatic states reached 87.3%, 91.9%, and 87.6% accuracy for parental, circulating tumour, and lung metastatic cells; misclassified lung metastatic cells were consistently assigned to the adjacent circulating tumour class, matching the continuum visible in UMAP projections. In the transfer experiment, a model pretrained on the metastasis dataset produced embeddings that predicted anti-PD-L1 response in a separate mouse study at accuracy comparable to the full spectra.

#Applications

The clearest payoff is acquisition time. Because embeddings from low-SNR spectra cluster as well as those from high-SNR spectra, experiments can trade exposure time for throughput — more cells measured, better coverage of biological heterogeneity, and access to temporal dynamics that long dwell times foreclose. The encoder also serves as a general dimensionality-reduction front end for Raman workflows that currently rely on PCA or MCR-ALS, which matters most where labelled data is scarce: patient biopsies, surgical margins, and preclinical therapy-response studies.

#Impact

RamanMAE is an argument for shared infrastructure in biomedical Raman spectroscopy. By showing that representations transfer across scale and application whenever acquisition parameters align, it gives labs a concrete reason to standardise wavenumber ranges and protocols and pool data toward larger spectral models. The authors are explicit that the model is not optimised — it is a minimal adaptation of ViT-MAE chosen for adoptability over performance — and transfer is shown only for spectra captured in the same wavenumber region. A code repository accompanying the paper is now public, though it carries no license file.