Natural-product chemists find new drugs by splitting a crude plant or microbial extract into fractions, testing each fraction in a cell assay, and following only the fractions that respond. The inactive ones get thrown away. A fraction can look inactive for two quite different reasons. It may contain nothing potent. Or it may contain a potent molecule whose effect is hidden because there is too little of it, or because other compounds in the same fraction blunt the cells' response. Current practice cannot distinguish these cases. Physicists face the same problem when a dark-matter search sees nothing, and they have a mature answer: turn the null result into an exclusion region, the set of hypotheses (particle mass, interaction strength) that would have produced a detectable signal and are therefore ruled out at 95% confidence. Their uncertain astrophysical inputs are handled as nuisance parameters, quantities that are estimated and then marginalized or profiled out rather than assumed.
We propose to carry this machinery into bioassay-guided fractionation. The missing piece is supplied by analytical chemistry's matrix-matched calibration: measuring an instrument's response to a known standard added to the actual sample background instead of to clean solvent. Here the "instrument" is the cell assay. We spike a known active compound into each inactive fraction at graded doses. This measures how strongly that fraction's background masks activity, and the masking factor becomes the nuisance parameter. The output for every discarded fraction is a statement a chemist can act on, for example: "no compound making up more than 0.3% of this fraction's mass with potency better than 100 nM is present, at 95% confidence." Because fractionation is sequential, these limits pass down to child fractions through mass balance. The whole purification tree then carries quantified risk rather than intuition.
Core Claim
A null bioassay result can be converted into a calibrated upper limit on the (abundance, potency) plane of any hidden active compound. The conversion needs two things: an assay-masking factor measured in the fraction's own matrix, which is then profiled out, and mass balance across purification stages. Under planted-truth tests, the resulting limits will meet their nominal false-exclusion rate. The thresholds used in practice today will not, because they ignore masking. If the claim holds, a fraction can be discarded with a stated miss probability, and a fraction can be flagged as "masked, not empty" and re-cleaned rather than abandoned.
The Mechanism
Consider a fraction tested at mass concentration $D$ in the assay. Suppose it hides one active compound that makes up mass fraction $w$ of the fraction and has half-maximal potency $E$ (in the same mass units). A simple saturating model predicts an observed effect
$$R(w,E) = m\,R_{\max}\,\frac{wD}{wD+E},$$
where $R_{\max}$ is the assay's maximal response and $m \in [0,1]$ is the fraction-specific masking factor. A value of $m=1$ means the background does not interfere. We estimate $m$ by standard addition: a reference active is spiked at several doses into the fraction and into clean medium, and the ratio of the two dose-response curves gives $m$ with an uncertainty. A point $(w,E)$ is excluded when the profile likelihood ratio test rejects it at 95% given the fraction's observed response. The masking factor $m$, the assay noise, and the extraction recovery are all profiled over.
This mirrors the structure of the MAGIC combined dwarf-galaxy branon search, which derives 95% upper limits on annihilation cross-section versus mass while treating the uncertain dark-matter density ($J$-factor) as a nuisance parameter. That paper supplies the statistical method. Its own open problem motivates our emphasis on measuring the nuisance directly instead of borrowing it: in that analysis, switching between two published $J$-factor sets moves the limits by up to an order of magnitude. Our $m$ plays that role, but it can be measured per fraction by spiking, which astronomers cannot do with a galaxy. Coogan, Profumo and Shepherd supply the logic of the exclusion itself: a hypothesis whose predicted signal exceeds the observed limit is ruled out.
Cuoco et al. used a second observation channel to sidestep an unconstrained nuisance. That suggests our second readout. Liquid chromatography–tandem mass spectrometry (LC-MS/MS) of the same fraction, itself calibrated by matrix matching, bounds the largest single-component abundance and so caps $w$ independently of the assay.
Sequential purification enters through mass balance. If a parent fraction of mass $M_p$ is split and a compound ends up in a child of mass $M_c$ with stage recovery $\rho$, then $w_c = \rho\,w_p M_p/M_c$. An exclusion computed on the parent therefore maps into every child, and nulls at successive stages combine the way multi-galaxy searches combine targets. The recovery $\rho$ is a second nuisance parameter. We calibrate it with matrix-matched LC-MS/MS of the spiked standard across each column, following the practice in the 79-pesticide QuEChERS study, which reports recoveries of 78.2–119.8% with matrix-matched curves. The same practice underlies the Fe3O4@COF mycotoxin work, which found matrix matching still necessary even after selective sorbent cleanup. That finding is the main reason we do not assume the masking factor or the recovery can be engineered away.
Research Program
The testbed is the Erythrina variegata stem-bark workflow of Li et al.. It partitions an ethanol extract into ethyl acetate, butanol and water fractions, then chains silica gel, Sephadex LH-20 (a size-based gel), reverse-phase ODS and medium-pressure columns, with a UMR106 osteoblast-like proliferation assay as the readout. It has two useful features. It reports concrete activity levels: the ethyl acetate fraction gave a 47% proliferation increase, the new isoflavanone (compound 4) 30.67% at $5\times10^{-7}$ mol/L, and laburnetin (compound 8) 27.00% at $5\times10^{-6}$ mol/L, still active at $5\times10^{-8}$ mol/L. It also yields purified actives that can serve as spike standards.
We will first rebuild the fractionation tree. Then we will run standard-addition curves with laburnetin in each inactive fraction to measure $m$, and matrix-matched LC-MS/MS to measure per-stage recovery. The statistical engine will be a small, open implementation of profile-likelihood exclusion on the $(w,E)$ plane with propagation through the tree.
A second, simulation arm uses synthetic fractionation trees whose masking and recovery distributions are fitted to the measured values. This lets us test coverage over thousands of trees.
Baselines
The naive reference point is current practice, as in Li et al. and in the Aconitum alkaloid and Gentiana glucoside isolations. A fraction is discarded when its effect falls below a fixed threshold relative to control. We will run it on the same fractions, at the same test concentration, with the threshold chosen so its nominal false-discard rate matches our 5%. It is the strongest competitor on practicality because it costs nothing extra.
The stronger practical baseline is a dose-response retest. Each inactive fraction is re-assayed at a higher concentration and in a dilution series, and discarded only if flat throughout. It gets the same number of assay wells as our method uses for spiking, so the extra experimental budget is identical. This is the fairest test of whether spiking is worth its wells.
The ablation that isolates our key ingredient keeps the full exclusion statistics but fixes $m=1$ (no masking). A second variant plugs in the point estimate of $m$ without profiling its uncertainty. If either ablation reaches nominal coverage, the matrix-matched nuisance is unnecessary.
For the strongest existing method, no reported result was found on Lacuna for false-exclusion rates or coverage of any inactive-fraction decision rule in bioassay-guided fractionation. The only reported numbers are the activity levels above, and they will set planting doses. Beating the baselines means two things together: empirical false-exclusion at or below 5% where the baselines exceed it, and an excluded region at least as large as the fixed-$m$ ablation on unmasked fractions.
Why not just…?
Why not just clean every fraction until masking disappears? The file's evidence says exhaustive cleanup leaves residual matrix effects even with tailored sorbents (Fe3O4@COF mycotoxins). Each extra column also costs yield in multi-stage isolations like Li et al.. Our recovery measurements will quantify this loss directly. If measured $m$ turns out to be near 1 for most Erythrina fractions after the second stage, cleanup is a sufficient answer from that stage on, and we will report that.
Why not just retest inactive fractions at several concentrations? This is the real threat. A dilution series detects low abundance, but it cannot detect masking, because the masking compounds scale with dose too. The planted-truth experiment below settles it. If dose-response retesting matches our coverage at equal well budget, the spiking step is not earning its cost, and the contribution shrinks to the mass-balance propagation.
Why not just run a spike-recovery check and discard if recovery passes? A pass/fail spike check is our method with its information thrown away. It gives a binary gate instead of a limit on $(w,E)$, and it cannot combine evidence across stages. The point-estimate ablation is close to this baseline. If that ablation matches full profiling in coverage, we will recommend the simpler check.
Why not just dereplicate inactive fractions by LC-MS and database retrieval? Retrieval from spectra is still weak. MassSpecGym reports that the best baseline, MIST, reaches Hit@1 of 14.6%. Conformal retrieval shows prediction sets growing to about 80% of candidates under distribution shift, which matrix-distorted spectra from crude fractions would induce. We therefore use LC-MS only to cap abundance, where matrix-matched quantification is reliable, not to name compounds.
Decisive Experiments
The central experiment is blinded planting. Purified compounds 4, 5 and 8 from the Erythrina isolation are reconstituted into inactive fractions at known mass fractions spanning the region where they would be sub-threshold. They are planted both into fractions with measured strong masking and into fractions with weak masking, and the analyst is blind to which fractions are planted. The primary metric is the empirical false-exclusion rate: the share of planted $(w,E)$ points that the method declared excluded, which should be at most 5%. We report it separately for high- and low-masking fractions, since that split is where the fixed-$m$ ablation and the thresholds should fail. Secondary metrics are the excluded area on the log-$(w,E)$ plane, as a measure of power, and assay wells per fraction.
The second experiment tests propagation. Parent-level limits are compared against direct measurements on children to check that mass-balance transfer, with profiled recovery, holds its coverage over two and three stages.
The third, run in simulation, varies the spread of masking factors to find where profiling stops mattering. This tells practitioners when the simpler rules are adequate.
The proposal is falsified if the fixed-threshold rule or the dose-response retest achieves nominal coverage on high-masking planted fractions at equal budget.
References
- Combined search in dwarf spheroidal galaxies for branon dark matter annihilation signatures with the MAGIC telescopes
- Monochromatic gamma rays from dark matter annihilation to leptons
- Constraining heavy dark matter with cosmic-ray antiprotons
- Determination of 79 pesticide residues in fruits and vegetables by QuEChERS-ultra performance liquid chromatography-tandem mass spectrometry
- Fe3O4@COF(TAPT–DHTA) Nanocomposites as Magnetic Solid-Phase Extraction Adsorbents for Simultaneous Determination of 9 Mycotoxins in Fruits by UHPLC–MS/MS
- Four New Isoflavonoids from the Stem Bark of Erythrina variegata
- Identification of Diterpenoid Alkaloids from the Roots of Aconitum kusnezoffii Reihcb.
- Terpenoid Glucosides from Gentiana macrophylla That Attenuate TNF-α Induced Pulmonary Inflammation in A549 Cells
- MassSpecGym: A benchmark for the discovery and identification of molecules
- Reliable Molecular Retrieval from Mass Spectra using Conformal Prediction