The Part Nobody Was Watching

I.The First Known Thing — the ear that gives away the hand

connoisseurship · the Morellian method

Giovanni Morelli was trained in medicine at Munich and came late to pictures. Between 1874 and 1876 he published a series of articles in a German art journal under the name Ivan Lermolieff, presented as a Russian scholar, and gathered them in 1880 as a study of the Italian masters in the galleries of Munich, Dresden and Berlin. The proposal was a procedure. To decide whose hand made a painting, do not look at the face, the composition or the expression. Look at the ear. Look at the hand, the thumbnail, the fold of drapery where it meets a wrist. Morelli compiled plates of these details, master by master, and read attributions off them; the re-attribution of the Dresden Sleeping Venus to Giorgione is the best known of the results. Bernard Berenson took the method up, and it remains a working part of connoisseurship, taught as one instrument among several.

the reason it works, stated as a reason

Morelli's argument for it was not that ears are interesting. It was that a painter attends to the conspicuous parts and executes the rest from habit. The face is where intention, instruction, patron and school all bear down; every competent painter of a period is pulled towards the same solution there, and a copyist or an imitator, attending to exactly the parts that a viewer attends to, reproduces them. The ear is under no such pressure. Nobody commissions an ear, no workshop corrects one, and an imitator copying a Titian is not watching the thumb. So the ear carries the painter's unrevised habit, and it carries it precisely because it did not matter. The method's discriminating power runs inversely to the significance of the feature.

what its critics established, which is also the point

The objections are old and have been absorbed. Edgar Wind observed that the method finds the artist where the artist was least present, and Richard Wollheim, examining its logic in 1973, set out how much it assumes about the stability of an unconscious habit across a career and across a workshop. Carlo Ginzburg's 1979 essay placed Morelli beside Freud and Conan Doyle as instances of a single evidential paradigm, one that reads an identity from the involuntary residue rather than from the intended display. None of this overturned the practical finding, which has been repeated in the computational attribution work of the last two decades: the features that best separate hands are small, local and unremarked, and the features that a human describes first are the features that separate hands worst. Connoisseurship's operating rule follows. Before attributing, ask of each feature how much anyone was attending to it, and weight it accordingly — high attention, low evidential weight.

II.The Second Known Thing — the hits and the heap they came from

drug discovery · hit finding, screening, library design

Walk to a hall where compounds meet a target for the first time. A screening deck of several hundred thousand to a few million compounds is put through an assay, and a few hundred or a few thousand come back active. The deck was not built for this target. It accreted: series left over from programmes against unrelated proteins, catalogue purchases, diversity sets, the output of whatever chemistry was cheap in the decade each portion was bought. One consequence is that a deck contains, scattered through it, families of close analogues — the same scaffold with ten or thirty substituents at one position — made years ago for reasons that have nothing to do with the protein now in the plate. Every member of each family is tested in the same run, on the same plates, in the same assay.

what the screen is read for

The output is read as a list of actives. Triage clusters them by scaffold or by fingerprint similarity, removes the compounds that aggregate, interfere with the readout or appear on the published lists of frequent hitters, confirms the survivors in dose–response and in an orthogonal format, and picks a handful of series to take into chemistry. The inactive analogues are not entirely wasted — a well-known 2005 method scores a hit higher when its structural neighbours in the deck are also active, and every medicinal chemist reads a hit's neighbours for a sense of whether there is a series there — but that use is about confidence: whether the hit is real. The neighbours are read for how many of them are active, not for which directions they are indifferent in.

and the question triage cannot answer

The decision that matters more is which hits bind the same way. Series that share a binding mode should be merged and prosecuted together; series that do not must be kept apart, because structure-guided optimisation of the second against a pose derived from the first will fail. Nothing in the primary screen reports this. The field's instruments for it come later or elsewhere: co-crystallography on a few compounds, competition and mutagenesis experiments, and computational analyses of accumulated potency data — matched series analysis, whose published applications list binding-mode suggestion among its uses, and non-additivity analysis, which reads binding-mode changes off double transformation cycles. All of these read the variation in potency: the rank order of substituents, the cliffs, the cycles that fail to close. All of them need a series that has already been made. And the field has been warned how badly similarity performs here — Babaoglu and Shoichet's 2006 deconstruction of a β-lactamase inhibitor found fragments of a ligand binding in places and orientations the intact ligand does not — so it knows that the conspicuous shared features of two actives are no evidence that they occupy the same site in the same way. It does not have an instrument that answers the question from the primary screen alone.

III.The Bridge

no triage paper cites an earlobe

The halls share nothing. Art attribution and hit triage have no common literature; where connoisseurship has met the laboratory it has done so through pigment analysis and imaging, which are questions about materials, and where cheminformatics has borrowed from outside it has borrowed from information retrieval and from ecology, not from the history of art. What the two halls have in common is a situation. Both must assign objects to a common origin — one hand, one binding mode — from a set of features, most of which are shared by everything that qualifies at all. Both have a natural ordering of features by how much attention was paid to them. One of the two halls discovered that the ordering is the wrong way round, and built its procedure on the inversion.

X: what was not selected for is what identifies

Call it unattended identity. Where objects are produced under strong constraint at some of their parts and none at others, everything that satisfies the constraint converges at the constrained parts, so those parts are shared by the genuine and the imitation alike and carry almost no information about origin; the unconstrained parts are free, and what fills them is whatever the producing process happened to leave there. So identity is read most reliably from the features that mattered least, and the evidential weight of a feature can be estimated in advance from how little selection it was under — before any attribution is attempted, and without knowing the answer to a single one. Two consequences are worth stating exactly, because neither is the ordinary observation that negative data has some value. First, it inverts the ordering of evidence rather than extending it: the pharmacophoric features that every active against a target displays are, on this account, the least attributive things about it, and the positions where the site is indifferent are the most attributive, so an attribution built from similarity is built from the weakest end of the evidence. Second, it makes the indifferent positions readable in the primary screen. A no-change is not a failure of an experiment; it is the observation that a direction is free, and the directions that are free are set by the site, not by the molecule — so two hits with nothing in common to the eye may be shown to share a site by agreeing about what it does not care about.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The screening hall reads its plates for actives and its series for slopes, and treats a substitution that changed nothing as a datum with no content. The connoisseur has never seen a microtitre plate. Neither asks whether the set of directions in which a hit's activity does not move is a signature of the site it occupies, nor whether that signature is already sitting, unread, in the primary screen that produced the hit.

the unattended-identity hypothesis

Does a screening hit carry an indifference profile — the set of substitution directions, computed from the deck's own analogues that were never designed for this target and were tested in the same run, along which activity does not change beyond the assay's reproducibility — such that agreement between two hits' indifference profiles predicts a shared binding mode better than whole-molecule similarity does, and does the attributive power of a position rise as the potency variance at that position falls, so that the least consequential positions are the most identifying?

why it would matter

Because the merge-or-separate decision is made at the point where the least is known and is expensive to reverse. Two hit series merged on scaffold similarity and prosecuted against one crystal pose will, if they bind differently, generate a year of structure-guided chemistry against the wrong picture; two series kept apart because their scaffolds differ will duplicate the same campaign twice. An indifference profile would cost nothing to compute — it is a re-reading of data already on disk, from the same plates, at the same time as the hit list — and it would be available before a single compound is made, which is the one window in which matched series analysis and non-additivity analysis have nothing to say. It would also change what a library is for. A deck designed to maximise diversity maximises, by construction, the number of hits with no near-neighbours; a deck designed to make hits attributable would carry deliberate constellations of indifferent substitution around each scaffold, and that is a different and cheaper thing to buy.

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 for its first pass: full-deck primary screens, their confirmatory dose–response sets and the structural record are public, and the ground truth it needs already exists for a few dozen targets.

the honest status

Note the narrowness. The claim is about hits from primary screens of decks that contain legacy analogue families, about one statistic computed from those families, and about one downstream fact — shared binding mode — for which independent ground truth exists. It says nothing about potency, about which hits are worth pursuing, or about how to optimise anything; a compound's indifference profile is not a reason to make it. It is also worth saying what the import does not carry. A painting has one maker and a lifetime of unrevised habit behind its ears; a binding site has no habit, and what fills the free positions of a ligand is the chemistry of whoever built the series, which means the signature being read is, at best, the site's tolerance seen through another programme's synthetic convenience. What transfers is a rule of evidence — weight a feature by how little it was attended to — and not the biography that makes the rule work on a canvas. The whole thing rests on an assumption that a morning with a public screening set would confirm or destroy: that real decks contain enough matched, co-tested analogues around their hits for the word indifferent to have a referent at all. Until that count 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.