The Match That Depends on the Light

I.The First Known Thing — the colour match that holds only under its own lamp

colour science · the metameric match and what it is permitted to conceal

What reaches the eye from a surface is a spectral power distribution: a curve of energy against wavelength, which for industrial purposes is recorded at some thirty or forty points across the visible range. What the eye takes from it is three numbers. There are three classes of cone, each with its own sensitivity curve, and the response of each is the integral of the arriving spectrum against that curve. Colour matching is therefore a projection from a function onto a three-dimensional space, and the map is many-to-one by a very large margin. Two surfaces whose spectra differ everywhere may return the same three numbers, and when they do, an observer sees no difference at all: not a small difference, none. Such a pair are metamers, and the match between them is a metameric match. A pair whose spectra coincide point by point are an isomeric match, and that is a different thing entirely.

the difference the match is allowed to carry

Because the projection has a kernel, the difference between two metamers is not arbitrary: it is an element of that kernel — a spectral function whose three integrals are all zero. Wyszecki named such a function a metameric black, since it can be added to any spectrum without changing the colour seen, and the decomposition was later put on a formal footing by Cohen and Kappauf, who split any spectrum into a fundamental part, which the observer sees, and a metameric black, which it cannot. The consequence that matters here is arithmetical rather than perceptual. Under the conditions in which it was declared, a metameric match is exact; there is no residual, no margin, nothing to grade. The quality of the match therefore reports nothing whatever about the size of the metameric black separating the pair, because that quantity lives entirely in the directions the measurement integrates away.

and what happens when the conditions change

Change the illuminant and the projection changes with it: the arriving spectrum is the product of the reflectance and the light, so the same two reflectances, integrated against different lights, no longer give the same three numbers. The match breaks, and it breaks by an amount set by the metameric black — by how differently the two surfaces were coloured — and not by how carefully the match was made. The same happens when the observer changes, since no two people have quite the same cone fundamentals, and again when the viewing geometry or the angular size of the sample changes. The trades that live on this have made it a measured quantity rather than a hazard: the CIE defines special metamerism indices for a change in illuminant and for a change in observer, each of them a colour difference computed for a pair that matched under the reference condition. Anyone who has bought a coat and a pair of trousers under the same lamps has performed the experiment.

the remedy, which is not more of the same light

The practical discipline that follows is exact and slightly counter-intuitive. Where the same colorants can be used on the same substrate, the right answer is to match spectra and not colours, because an isomeric match survives every illuminant and every observer by construction. Where that is impossible — a dye against a pigment, a plastic trim against a woven panel, a repair against a weathered original — the match will be metameric whatever anyone does, and the industry's response is to declare it under several deliberately unlike conditions: daylight, an incandescent source, a fluorescent one, and increasingly a light-emitting diode, with the work on matches held simultaneously under multiple illuminants going back to Burns and colleagues in the late 1980s. The reasoning is dimensional. Agreement under one condition constrains three numbers; agreement under three unlike conditions constrains nine, and the kernel shrinks accordingly. Agreement under a second lamp nearly identical to the first constrains nothing new, because the directions it removes from the kernel were removed already. What buys certainty is not the number of conditions but the extent to which they fail to resemble one another.

II.The Second Known Thing — the two compounds a project calls the same

drug discovery · lead optimisation and multi-parameter trade-offs

Walk to a hall where a chemical series is being turned into a candidate. Each compound made is put through a panel: potency at the target, a counterscreen or two for selectivity, solubility, permeability, stability in liver microsomes, plasma protein binding, inhibition of the hERG potassium channel, a set of cytochrome readouts, sometimes a broad in vitro safety screen of the kind Bowes and colleagues set out in 2012 across some forty-four targets associated with adverse effects. The result is a profile: a vector of numbers, a few dozen long, standing in for a molecule that has an effectively unbounded number of properties. Multi-parameter optimisation is the business of moving that vector, and its instruments — desirability functions, property-based scores, the various drug-likeness indices — are ways of compressing it further still, to one number, so that compounds can be ranked.

where two compounds are actually declared equivalent

Ranking is not the only thing done with a profile. The stronger operation is declaring two compounds the same, and projects do it constantly: when a bioisosteric replacement is accepted as profile-neutral, when a series is triaged against a comparator, and most consequentially when a backup is nominated. A project taking one candidate into development carries a second compound behind it, and the standing logic of backup selection is that the backup should match the lead on the measured profile — otherwise it is not a backup — while differing from it structurally, ideally by belonging to a different chemotype, so that a failure arising from the lead's scaffold does not take both compounds down together. The instruction is to find the most distant molecule that measures the same. It is given in exactly those words, and it is sound as far as it goes.

and what is known about the panel itself

The panels are well characterised as objects. Large profiling matrices exist in the open: Karaman and colleagues in 2008 and Davis and colleagues in 2011 measured panels of kinase inhibitors against several hundred kinases apiece, and Klaeger and colleagues in 2017 profiled more than two hundred clinical kinase drugs by chemical proteomics. What those matrices show, consistently, is heavy correlation structure — kinases that move together, inhibitors that cluster, a spectrum of singular values that decays fast, so that a panel of two hundred columns carries nothing like two hundred independent questions. The field reads that structure as redundancy to be trimmed when designing a screening panel. What is not computed anywhere, and does not enter the nomination of a backup, is the complementary object: given this panel, what set of differences between two compounds is it constitutionally unable to report, and how large is the part of the difference between these two particular compounds that lies inside it. When an equivalence proves false in vivo the standing remedy is to add assays, and the assays added are usually close relatives of the ones already there.

III.The Bridge

no colorimetry paper counts compounds

The halls share nothing. Metamerism belongs to colorimetry and to the industries that must match dyes, paints, plastics and print; backup nomination to medicinal chemistry and to project governance. No review spans them, and the word metamerism means something else again in organic chemistry, where it is an old name for a kind of structural isomerism — a homonym, not a link. Nor is the bridge the familiar observation that assays miss things, which every chemist would grant and which predicts nothing. What the two share is narrower: an equivalence declared between two objects through a measurement of far lower dimension than the objects themselves, so that the declaration has a kernel; and in both halls the equivalence is not incidental but deliberately sought between objects chosen to be as unlike as the declaration permits.

X: an exact match reports nothing about what it integrated away

Call it metameric equivalence. Where two objects are declared the same by a measurement whose dimension is far below theirs, the declaration has a kernel, and the pair differ by an element of it. Since the match is exact by construction within the measurement, its quality carries no information at all about the size of that difference. So the divergence to expect when conditions change is set by how differently the two objects are constituted, not by how well they matched; it falls with the effective rank of the measurement rather than with its length, so that adding a correlated assay removes nothing from the kernel; and an equivalence deliberately sought between distant objects — which is precisely what insurance against a shared failure requires — is the most metameric equivalence available, buying independence in the measured directions at the price of maximal difference in the unmeasured ones. Three things follow that the chemical hall does not currently say. First, backup nomination is a trade with two priced sides and the field prices one: structural distance buys independence from a chemotype-borne failure and simultaneously buys divergence off the panel, and the same distance is the coin in both purchases, so a backup from a remote chemotype is not simply safer, it is differently exposed. Second, the redundancy already visible in the public profiling matrices is not only wasted effort, it is the size of what the panel cannot ask, and the quantity that governs an equivalence is the panel's effective rank rather than its column count — a number computable today from data already deposited. Third, the colour trades' remedy transfers intact and is not the one in use: where the objects compared are constituted differently, declare the equivalence under deliberately unlike conditions — another species, another cell background, an orthogonal readout — rather than under more assays of the kind already present, and value a proposed assay by the part of it the panel does not already span.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The chemical hall asks whether the backup matches the lead and whether it is structurally distinct, and treats both answers as good news. The colour hall has no compounds, but it learned, from matched panels that disgraced themselves in daylight, that a match declared through a projection is an assertion about the projection and that the objects most worth matching are the ones whose match will fail worst. Neither hall asks whether the two properties a project wants in a backup are in tension with one another, nor whether the tension can be quantified before anything is dosed, from the profiling matrices the field has already published.

the metameric equivalence hypothesis

Can a metamerism index be defined for a pair of compounds declared equivalent within tolerance on a given assay panel — the divergence to be expected on a condition deliberately held out of that panel — and does it predict the held-out divergence better than the pair's structural similarity or the quality of its match, with divergence rising in structural distance at fixed match tolerance and falling with the panel's effective rank rather than with the number of assays in it?

why it would matter

Because the operation the index governs is one of the most expensive decisions a programme makes on the least evidence. A backup is nominated once, carried for years, and asked to perform exactly when the lead has failed — that is, under a condition the panel did not contain, since had the panel contained it the lead would not have reached nomination. If the hypothesis holds, a project can compute, before nominating, roughly how much of the difference between its two compounds its own panel is unable to see, and can buy that visibility where it is cheapest: by adding the one assay least spanned by the others rather than the one most easily run. It would also give panel design an objective it does not have. Panels are currently assembled from targets known to cause trouble, which is a list of answers; the complementary criterion is to assemble them so that the compounds a project actually intends to call equivalent are not metamers under them, which is a property of the panel and the pair together and cannot be read off either alone.

V.How It Could Be Proven — or Broken

a candidate must be breakable, or it is merely poetry

An intercognix candidate earns nothing until it names its own test. This one needs no new chemistry and no new assay. It needs a matrix of compounds by measurements dense enough to hold out columns from, which is what the published kinase profiling sets are and what the larger public bioactivity databases approximate for well-measured targets.

the honest status

Note the narrowness. The claim concerns pairs declared equivalent within a stated tolerance on a stated panel, and one outcome — divergence on held-out measurements — chosen because it can be computed from data already public rather than because it is the thing anyone ultimately cares about. It does not claim that panels are uninformative, that structural distance in a backup is a mistake, or that the field is unaware that assays miss things. It claims something smaller and testable: that the field's two stated requirements of a backup pull against each other through a mechanism that has a name and a measure in another discipline, that the measure is computable before nomination, and that the quantity currently used to reassure a committee — the closeness of the match — is by construction silent on the risk the committee is worried about. It is also worth saying what does not transfer. A colourist can choose an isomeric match and be done; a project cannot nominate a backup identical to its lead, since identity is exactly what a backup must not have. The chemical hall therefore lives permanently in the metameric case, which is the reason the index would be worth having and also the reason it will be harder to interpret. The whole thing rests on a comparison the deposited matrices already allow: whether compounds matched on a panel diverge off it by an amount that the panel and the pair together predict, or by an amount that similarity alone predicts. If the second, this is a well-dressed restatement of what every chemist already assumes, and should be said so plainly. Until that comparison is made, this remains exactly what its stamp says: a candidate. A structural echo between two halls of the library, waiting for someone with the tools of both to walk the corridor.