Human seminal protein concentration is among the least measured analytes in human physiology — produced billions of times a day and quantified almost never. We describe a consumer assay that quantifies total protein content from a colorimetric reaction imaged with an unmodified smartphone camera. A fixed sample volume is combined with a Coomassie-based reagent whose absorbance shifts in proportion to bound protein; the reaction is photographed in transmission against a backlit second screen — the “lightboard” — that renders fiducial markers, a tube socket, and on-screen gray and black reference patches. We recover a calibrated absorbance proxy from a glare-trimmed, gray-normalized channel ratio, gate out captures with insufficient light, correct for per-batch reagent variation, and report the result directly as the Cum Capacity Index (CCI) — the recovered protein concentration, in mg/mL, rounded to the nearest whole number. Against a 9-level bovine serum albumin dilution series spanning 0–200 mg/mL, a four-parameter logistic model fits the chromatic response with R² = 0.99 and a median intra-assay coefficient of variation near 1.7%, approaching entry-level laboratory spectrophotometry while requiring no instrument, clinic, or appointment.
1Introduction
The quantification of protein in biological fluids is a solved problem in the laboratory and an unsolved problem everywhere else. The dye-binding method introduced by Bradford in 1976 remains a workhorse of analytical biochemistry precisely because it is fast, sensitive, and requires only a color change to read [1]. Yet for one of the most frequently produced biofluids in human biology, that color change has never been made available to the person producing it.
Cumomics closes this gap. Our thesis is simple: a reaction that can be read by a benchtop spectrophotometer can also be read by the camera every adult already carries, provided the imaging conditions are constrained and calibrated. The remainder of this paper describes the kit architecture, the assay chemistry, the image-capture and signal-recovery pipeline, the calibration strategy, and the model that converts a recovered concentration into a single comparable number.
2Materials & Kit Architecture
Each kit contains a single-use reaction tube pre-loaded with lyophilized reagent, a fixed-volume transfer dropper, and a QR code encoding the kit and reagent-lot identifiers. The dropper delivers a standardized sample volume, removing the single largest source of operator variance in any home assay; the QR binds each capture to the exact lot that produced it.
The capture target itself is not printed. Earlier kits used a reflective scan card, but a card illuminated by ambient room light is at the mercy of whatever color and intensity that light happens to be. We therefore moved the target onto a second screen — the “lightboard” — which the user opens on any laptop, tablet, or monitor. The lightboard renders four ArUco fiducial markers that define the imaging plane and scale [2], a socket that seats the tube, and two on-screen reference patches — a mid-gray normalization reference and a black patch — all on a controlled white backlight. The tube is then photographed in transmission against this emissive surface. Because the screen is its own light source, it provides a uniform, known-spectrum illuminant whose emitted patches travel the same optical path as the sample, anchoring the color normalization to a known target rather than to the room. This is the single largest improvement in data quality over a reflective card.
3Colorimetric Assay Principle
The reagent is a Coomassie Brilliant Blue G-250 formulation. In its free, cationic state the dye absorbs maximally near 465 nm and appears red-brown. On binding protein — primarily via arginine and aromatic residues — the anionic form is stabilized and the absorbance maximum shifts to approximately 595 nm, producing the characteristic blue. The magnitude of the shift is, over the working range, proportional to total protein concentration.
Because the readout is a shift in absorbance, it maps cleanly onto the channels of a consumer image sensor. We define a chromatic response R = B / G — the ratio of the blue- and green-channel means over the extracted reaction region. Blue tracks the bound-dye band near 595 nm, while green lies outside both dye maxima and serves as an internal exposure reference; their ratio therefore increases monotonically with protein content and is robust to overall illumination.
4Sample Acquisition & Smartphone Capture
The user produces a sample, draws a fixed volume with the dropper, and dispenses it into the reaction tube; color development completes within the published reaction window. On a second device — a laptop, tablet, or monitor — the user opens cumomics.com/light to display the lightboard, then opens cumomics.com/scan on the phone — no application is installed — and scans the kit QR to bind the session to a kit and lot. The phone camera is then aimed at the tube held in the lightboard's socket, lit from behind by the screen.
Capture is triggered automatically once all four fiducial markers are detected, guaranteeing a known pose and scale. The four marker positions define a homography onto a canonical millimetre frame, so the tube and reference regions always rectify to the same pixels. From the rectified frame we (i) normalize each channel against the on-screen gray reference, (ii) localize the reaction region of interest within the tube, and (iii) extract per-channel statistics as a glare-trimmed mean. The reference is mid-gray rather than white deliberately: a white patch saturates the sensor near full code value and therefore cannot track changes in backlight level, whereas a gray patch retains the headroom to do so. Because the screen is both the illuminant and the reference, normalization is anchored to a known emissive target rather than to the handset's auto-white-balance — which is what makes cross-device comparison tractable.
Three measures harden the readout against real-world capture conditions. First, the B/G signal is a channel ratio, so it is intrinsically invariant to overall brightness and, empirically, to screen tilt across ±30° — both scale the channels together and cancel in the ratio. Second, because the curved tube can throw a specular highlight from the bright screen, per-channel values are taken as a trimmed mean that discards the brightest quartile of pixels in the region, rejecting that glare; the sampling window is also enlarged to average more of the reaction column, reducing the residual variance of a hand-held capture. Third, a light-adequacy gauge — the ratio of the gray reference to the black patch — is evaluated on every frame: when it falls below a fixed floor, the signal is approaching the sensor noise floor (the screen is too dim or held too obliquely) and the B/G ratio degenerates, so the capture is flagged for retake rather than silently scored. This single gate subsumes both the low-brightness and the extreme-tilt failure modes, since each ultimately reduces to too little light reaching the sensor.
5Signal Recovery & Per-Batch Calibration
Reagent lots vary. Each manufactured batch is characterized against a dilution series of bovine serum albumin (BSA) to produce a standard curve relating chromatic response to known concentration. The resulting coefficients are stored under the lot identifier encoded in the kit's QR code, so every measurement is interpreted through the calibration of the exact batch that produced it. This batch-specific correction is the mechanism by which two users' numbers are made comparable.
We model the standard curve as a four-parameter logistic (4PL) rather than a polynomial: R(c) = d + (a − d) / (1 + (c/e)^b), where a is the lower asymptote (blank ratio), d the upper asymptote, e the half-maximal concentration (EC50), and b the Hill slope. The curve is monotonic, and therefore uniquely invertible — a measured R maps to exactly one concentration — which a polynomial fit cannot guarantee. Coefficients are fit by Nelder–Mead minimizing relative residuals, so the floor stays anchored at the measured blank instead of being dragged up by the high-concentration points.
The current production lot (BR-2026-06-14) was characterized over a 9-level, two-fold serial dilution of BSA from 0.78 to 200 mg/mL plus a blank, in five replicates (n = 48). The fit gives a = 0.360, d = 3.870, e = 297 mg/mL, and b = 0.510, with R² = 0.996 across the full 0–200 mg/mL range (Fig. 4). At scoring time the relation is inverted in closed form, c = e·((a − d)/(R − d) − 1)^(1/b), and clamped to the calibrated range.
Assay performance for the characterized lot is summarized in Table 1. Limits of detection and quantitation follow the usual blank-noise convention — the signal at 3.3 and 10 times the blank standard deviation above the mean blank, inverted through the fitted curve — while inter-device reproducibility awaits the multi-handset validation cohort.
| Metric | Value | Unit |
|---|---|---|
| Calibrated range | 0 – 200 | mg/mL |
| Half-maximal concentration (e) | 297 | mg/mL |
| 4PL fit R² | 0.996 | — |
| Limit of detection (LoD) | 0.22 | mg/mL |
| Limit of quantitation (LoQ) | 1.93 | mg/mL |
| Intra-assay CV (median) | 1.7 | % |
| Inter-device CV | — | % |
Table 1. Assay performance for production lot BR-2026-06-14 (BSA standard, n = 48). LoD/LoQ are the concentrations at 3.3 and 10 × blank SD above the mean blank, inverted through the fitted curve; intra-assay CV is the median replicate CV of B/G across the nine non-blank levels (range 1.3–6.1%). Inter-device CV pending the multi-handset cohort.
6The Cum Capacity Index
The per-batch calibration of §5 places every measurement on a common, lot-independent footing, so the CCI needs no further rescaling. What it reports is the apparent protein concentration on the BSA-calibrated scale — the true concentration modulated by the seminal-plasma matrix and Bradford's protein-specific response. The general form below collapses to exactly that; operationally it is read out directly, in mg/mL, rounded and floored at zero.
The physical terms are ρ, the true protein concentration; M, the seminal-plasma matrix factor; and βBSA→SP, the BSA-to-seminal-plasma response offset (Coomassie binding is protein-specific). The calibrated covariance Σθ and partition function Z(θ) are corrections that evaluate to unity under per-batch calibration, so the index collapses to CCI = ⌊ρ · M · β⌉ — the apparent, BSA-equivalent concentration, in mg/mL, rounded and floored. M and β are constant across the cohort, so they set the absolute scale without affecting comparability.
7Cohort Normalization & the Leaderboard
Scores are posted, anonymized, to a live leaderboard stratified by group, state, and country. Because every CCI is a lot-calibrated apparent concentration on a common BSA-anchored scale — with the matrix and response factors constant across the cohort — scores remain directly comparable as the population grows; cohort statistics contextualize where a given number falls without altering the number itself. The leaderboard is therefore both the product surface and the cohort lens.
8Reproducibility & Validation
To assess repeatability, replicate measurements of the same sample across devices and lighting conditions are compared. We report agreement via a Bland–Altman analysis of paired captures; the acceptance criterion is that the limits of agreement fall within the assay's clinically irrelevant range.
9Limitations
Total-protein colorimetry does not resolve individual analytes, and, as a dye-binding method, its response is protein-specific: the curve is anchored to a bovine serum albumin standard, so reported values are relative to that basis rather than an absolute, composition-independent mass. Capture quality still depends on the second screen — a dim, color-shifted, or low-gamut display weakens the normalization anchor and the transmission signal — but rather than scoring such a frame silently, the light-adequacy gate flags it for retake. The index is comparative, not diagnostic: Cumomics measures a protein-proportional signal and a person's standing relative to the cohort, and nothing in this paper should be construed as a medical claim.
10Conclusion
We have shown that a constrained, calibrated smartphone capture can recover a protein-proportional signal from a Coomassie reaction and normalize it into a single comparable index. The approach moves a laboratory-grade measurement out of the laboratory without sacrificing comparability. The cohort is open and growing; this is a preprint, and we are only at the beginning.
References
- [1] Bradford, M. M. (1976). A rapid and sensitive method for the quantitation of microgram quantities of protein utilizing the principle of protein-dye binding. Analytical Biochemistry, 72, 248–254.
- [2] Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F. J., & Marín-Jiménez, M. J. (2014). Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6), 2280–2292.
- [3] Cumomics Research (2026). Per-batch calibration of consumer colorimetric assays. Internal Technical Note, CMX-TN-004.