Purpose

On paper, Raman spectroscopy is close to ideal for biology. Raman provides information about a sample's chemical composition while offering properties that make its application to biology easy: label-free, information-rich, easy sample prep, minimally damaging, and can be used with water-containing samples. All of this suggests that Raman should be widely used in high-throughput biological data acquisition.

In practice, there’s a gap between its potential and its use. We’ve run into a set of common challenges: fluorescence can bury the weak Raman signal, biological spectra are broad and overlapping, excitation light can damage what we're measuring, and keeping cells alive and comparable across runs is difficult.

We’re now assessing how to overcome these issues by improving our hardware, experimental design, and data processing. There are known strategies to mitigate each of these problems, but it’s unclear which ones are most impactful and practicable. We're seeking feedback from people with hands-on experience overcoming the limitations of Raman spectroscopy to help us prioritize our next steps.

Motivation

We've spent the past two years developing spontaneous and coherent Raman spectroscopy for biological research at Arcadia. Our interest is in label-free, sensitive techniques with low sample preparation complexity that we can use for applications like high-throughput phenotyping. We built and applied a low-cost, open-source spontaneous Raman spectrometer 1–2, automated it for throughput 3, and benchmarked commercial instruments and excitation wavelengths for phenotyping algae 4–5. We've detected chemical perturbation of protein structures 6, imaged single live yeast cells and their mutants with stimulated Raman (SRS) and coherent anti-Stokes Raman (CARS) 7–8, shown that run-to-run batch effects can dominate strain-level signal 9, built a model to predict and simulate spectra 10, and outlined where Raman might speed up therapeutic discovery 11. Throughout this work, we've encountered the familiar challenges of Raman spectroscopy in biology.

As we move from exploration to application of this technology, we're now deciding how to refine our hardware, experimental design, and processing methods. This pub lays out the challenges we find most limiting, summarizes the approaches we’ve seen proposed in the literature to address each one, with our current read on their pros and cons, and asks for input from people who've actually built or applied these techniques. We also know that some of the real answers to our questions may be “it depends” on the sample and the question of interest. We’d still like to know how others at this intersection are thinking about the same issues, what factors into their decisions, and what works and doesn’t work for specific problems.

Challenges applying Raman to biological questions

Five problems account for most of the difficulty we've encountered applying Raman to biology, and the literature documents them as well 12. Some are intrinsic to Raman; others come specifically from working with living or biological material.

Fluorescence

Endogenous fluorophores emit broadly and efficiently, often orders of magnitude more strongly than the Raman signal, so their background can swamp the sharp Raman bands. NAD(P)H, flavins, and porphyrins contribute in many cells, and in photosynthetic cells, chlorophyll dominates the background under visible excitation 12. We've compared different excitation wavelengths on Chlamydomonas reinhardtii at 473, 532, and 785 nm, and observed that the visible lines drove a sample fluorescence background, while at 785 nm the glass substrate contributed its own 5.

Weak scattering

Only about one in 10⁷ incident photons scatters inelastically, so spontaneous Raman can yield weak signals 13. Signal-to-noise improves with integration time and laser power, but only as the square root of the photons collected (shot-noise limit), and a living sample tolerates only so much of either, so the achievable signal-to-noise is bounded. We’ve tested surface-enhanced Raman scattering (SERS) 11 and drop coating deposition Raman (DCDR) to an extent to mitigate these issues.

Chemical complexity and band overlap

Cells and tissue are chemically heterogeneous and largely amorphous rather than ordered or crystalline, so their bands are often broad and overlapping; unmixing and assignment can be difficult 12. Sharp, well-separated reference spectra tend to come from high-purity, well-ordered standards, which are well represented in the literature and spectral databases but don't always resemble what we measure in living samples. Also, Raman also doesn't report chemical composition evenly: band intensities follow scattering cross-sections, and when excitation overlaps an electronic transition, resonance enhancement can raise a few modes by orders of magnitude. Label-free doesn't mean unbiased, so we care as much about what a spectrum doesn't show as about the peaks it does.

Sample alteration

The excitation light that produces the signal can also heat, bleach, or photochemically alter the sample. Shorter, higher-energy wavelengths tend to do the most damage, though longer near-infrared wavelengths might also alter the sample, especially by heating.

Constrained and variable measurement conditions

Keeping cells alive and physiologically intact over a long acquisition, or reaching tissue below the surface, tends to conflict with the geometry and dose that yield the best signal. Run-to-run batch effects, especially due to environment fluctuations or specific setups, can also dominate the biological differences we're trying to measure 9. We’ve developed methods to assess the prevalence of batch effects and are carefully characterizing our instruments and experimental setups day to day using calibration standards.

Mitigation strategies

There are multiple proposed solutions to some of the issues we outlined above. We group them below by the problem they address, established and emerging methods together (Table 1). While we’ve already used some of these, our specific implementations could be improved or modified. Others on this list we've read about more than we've run, and that's where outside experience would help us most; we’ve found no clear consensus in literature.

Other directions we're watching address the challenges above: compressive and single-pixel Raman, spatial heterodyne Raman, and waveguide-enhanced (on-chip) Raman for speed and compactness; transmission Raman for bulk samples; and deep-learning denoising and fluorescence removal. We're also already applying Raman in parallel with other imaging or spectroscopy methods (e.g., fluorescence-lifetime imaging microscopy, FLIM, and brightfield) to provide complementary or confirmatory information, and with other sample processing steps (e.g., liquid chromatography–mass spectrometry, LC-MS).

What we’re optimizing for

We're building toward discovery at scale, and this shapes which solutions are practical for us. We realize Raman spectroscopy is somewhat of a generalist chemistry tool in a field that often prioritizes specialist instruments; we believe its strengths and the high-dimensional chemical information it yields make it uniquely useful for high-throughput biology.

We include this section on our company's unique considerations so readers understand why certain solutions might work in some cases, but not for us. We hope the following goals can add context and help guide any feedback you leave on this pub.

Reproducibility at scale

We care about results that hold up across samples, days, and operators, not a single demonstration. For hardware, an unstable, homebuilt rig that only its builder can run is hard to scale, and we weigh robustness and reproducibility heavily, sometimes more than peak performance. In terms of processing, many methods are used in literature and analysis software (e.g., for baseline fitting, denoising, smoothing, normalization, peak finding, signal-to-noise quantification, etc.), but few people consider how robust their results are. In both cases, people often report what worked (in their hands) for a specific experiment; we’re interested in what works consistently, robustly, and across implementations.

Untargeted discovery

Much of our work is label-free phenotyping meant to surface novel phenotypes. Methods that rely on strong prior knowledge or targeted labels reintroduce the problem of deciding in advance what you'll see, the same constraint tagging imposes. We lean heavily on machine learning, which readily latches onto batch structure instead of biology, so reproducibility across runs is critical.

Throughput

Scaling to many samples means experimental design and strategy are paramount. Ideally, we want to routinely acquire data on hundreds or thousands of samples, so we're looking for ways to streamline sample prep and acquisition while maintaining signal.

Substrates

The substrates and containers typically used in biological research are often poor for Raman. Polystyrene plates and glass-bottom multiwells add their own bands and background signal, so we often use mirror-finish stainless steel plates instead. All-quartz or CaF₂ substrates are ideal optically but expensive at scale, and we'd rather not replate everything for every measurement, so the practical question is if and where the workable middle ground exists.

Cost and openness

Our open-science mandate pushes us toward approaches that others can afford and reproduce. The less a method depends on costly hardware, the more useful it is to the people who might build on our work.

Weigh in!

We’re aiming to scale this technique and are deciding how to do so. If you've worked with any of these strategies, especially on a biological or biochemical investigation, we'd value your perspective. Please comment on this pub with your experience, including positive, negative, and inconclusive results. Point us to work we've missed, and tell us where our understanding is off. You can comment on specific questions in-line above, or overall at the bottom of this page.

Pub preparation

We used Grammarly Enterprise to help match Arcadia's style. We also used Claude (Opus 4.8) to help copy-edit draft text to match Arcadia's style, to help clarify and streamline text that we wrote, and to get feedback on how we could better tailor our draft to serve our audience and goals.