Presentation Information
[P01-044]Disentangling Raman spectra, noise, and autofluorescence via a Physics-Informed AI
○Thomas David Muir1, William Mills2, Mohammadrahim Kazemzadeh2, Huabing Yin2, Alexander Krull1 (1. University of Birmingham (UK), 2. University of Glasgow (UK))
Keywords:
Raman Spectroscopy,AI,Denoising,Decomposition,Unsupervised
Live-cell Raman spectroscopy provides a powerful, non-destructive measurement method for analysing molecular compositions that can be applied across multiple disciplines. When a laser is focused onto a sample, inelastic Raman scattering occurs, where the received photons correspond directly to the vibrational modes of molecular bonds. This map of vibrational modes is a spectral fingerprint that allows for the distinction of complex molecules, as well as an understanding of their underlying structures.
However, the Raman effect is an inherently weak phenomenon, which introduces two steep hurdles: autofluorescence and noise. Not only does autofluorescence raise the shape of the true Raman signal onto a broad baseline hill, but the massive signal-dependent Poisson shot noise from this emission, alongside standard Gaussian readout noise, heavily obfuscates the far less frequent Raman photons.
Consequently, downstream analysis of Raman spectra proves difficult without a preprocessing method to remove the noise and baseline. While traditional denoising (e.g., Savitzky-Golay filtering) and baseline correction algorithms (such as airPLS) exist, their effectiveness depends heavily on parameter settings tuned by the experience of the operator.
To address this, we present a fully unsupervised, physics-informed temporal Variational Autoencoder (VAE) designed to perform mathematically rigorous, probabilistic blind source separation on time-series photobleaching data. In order to remove the fluorescence baseline physically, biologists take advantage of photobleaching - a process exploiting the exponential decay of fluorophores under continuous laser exposure. While fluorophores decay exponentially, the Raman signal is static. We train our network unsupervised on only a fraction of this decay.
The model is trained and evaluated on synthetically mixed, mathematically pure baseline-free biological macromolecules subjected to simulated camera noise. Preliminary results demonstrate a highly robust capacity to disentangle overlapping chemical signatures and accurately extract the static Raman spectra, using only 5% of the typical physical decay process. Consequently, this rapid extraction could enable significantly higher experimental throughput and reduce phototoxic damage to the live cells. Furthermore, because fluorophore decay rates are highly dependent on local cellular environments, extracting the fluorophore information itself could prove biologically useful. Additionally, the VAE's inherently probabilistic outputs provide vital uncertainty quantification, enabling far more informed downstream decision-making.
However, the Raman effect is an inherently weak phenomenon, which introduces two steep hurdles: autofluorescence and noise. Not only does autofluorescence raise the shape of the true Raman signal onto a broad baseline hill, but the massive signal-dependent Poisson shot noise from this emission, alongside standard Gaussian readout noise, heavily obfuscates the far less frequent Raman photons.
Consequently, downstream analysis of Raman spectra proves difficult without a preprocessing method to remove the noise and baseline. While traditional denoising (e.g., Savitzky-Golay filtering) and baseline correction algorithms (such as airPLS) exist, their effectiveness depends heavily on parameter settings tuned by the experience of the operator.
To address this, we present a fully unsupervised, physics-informed temporal Variational Autoencoder (VAE) designed to perform mathematically rigorous, probabilistic blind source separation on time-series photobleaching data. In order to remove the fluorescence baseline physically, biologists take advantage of photobleaching - a process exploiting the exponential decay of fluorophores under continuous laser exposure. While fluorophores decay exponentially, the Raman signal is static. We train our network unsupervised on only a fraction of this decay.
The model is trained and evaluated on synthetically mixed, mathematically pure baseline-free biological macromolecules subjected to simulated camera noise. Preliminary results demonstrate a highly robust capacity to disentangle overlapping chemical signatures and accurately extract the static Raman spectra, using only 5% of the typical physical decay process. Consequently, this rapid extraction could enable significantly higher experimental throughput and reduce phototoxic damage to the live cells. Furthermore, because fluorophore decay rates are highly dependent on local cellular environments, extracting the fluorophore information itself could prove biologically useful. Additionally, the VAE's inherently probabilistic outputs provide vital uncertainty quantification, enabling far more informed downstream decision-making.
Comment
To browse or post comments, you must log in.Log in