← Method

Method · Emotion in Matter

The shape of this molecule we already know: it's measured, deposited, certain. The work isn't meant to discover it. It's meant to perform a gesture — trying to find it again starting from the data, with the most rigorous method I can build. I know from the start that I will never get back to the molecule exactly: that's the whole point. Searching with precision, and failing. That distance — between the thing and its estimate — is the work.

known molecule (PDB)
 └─ contact graph            atoms = nodes · close in space = edges
     └─ turned into a covariance    (nearby atoms vary together)
         └─ data generated from that distribution
             └─ GraphicalLasso   →   which relations are direct?
                 └─ a threshold   →   the estimated network (ISN)
                     └─ the discrepancy  →   what the estimate can't recover

01 · the reference

The true shape: the contact graph

We start from the 3D coordinates of every heavy atom (hydrogens are left out). Two atoms become connected nodes when they're close enough in space — under a threshold of about half a nanometre (4.5 ångströms). This is the contact graph: the real shape of the molecule. All atoms are kept, not just the alpha-carbons: for a peptide of 5–11 residues, Cα atoms alone would give barely a dozen nodes — too few for any statistics.

Intent: fix the answer we already know — the yardstick against which everything else will be measured.

02 · generating the data

From the shape, a distribution — and from that, data

I have no experimental measurements of "how this molecule behaves," so I generate the data myself through simulations, based on its structure. I translate the graph into a covariance: a rule stating that atoms close in the structure tend to vary together, distant ones don't. From this distribution I draw hundreds of samples — vectors of numbers that carry the molecule's geometry inside them, without tracing it perfectly.

This isn't an arbitrary trick: that covariance coincides with the Gaussian Network Model, the way biophysics describes how a protein vibrates around its own shape. I didn't invent it — I found it again.

Intent: have data to interrogate, data that contains the molecule but doesn't hand it back whole. From here on, I pretend not to know the answer.

03 · estimating the relations

GraphicalLasso: which bonds are truly direct?

Now all I have is the data. I ask: looking at these numbers, which atoms are really bonded? It isn't enough to see two atoms varying together — they might both be moving because they're bonded to a third one: a merely apparent correlation. I want the direct relations, stripped of the indirect ones.

For this I use GraphicalLasso. It's the tool that estimates which pairs of atoms remain dependent even after accounting for all the others — which is the exact definition of a direct relation — and it does so by searching for the simplest possible explanation: few connections, the ones that matter. It's the most rigorous method I know for this question, and that's exactly why I choose it: if I have to search again for a structure I already know, I want to do it with the most exact instrument, not an approximation.

04 · the discrepancy — the work

The error sought with precision

The estimate never gets back to the molecule exactly. Two kinds of error remain — and I don't correct them, because they're the content:

False positives — the ghosts

Bonds the statistics estimate but the matter doesn't have. Projections, expectations: connections real in the data, absent in the structure.

False negatives — the shadows

Real contacts the model doesn't "see": too weak, too local to emerge. Proximity without relation.

Intent: fail with precision. I searched for the molecule with the most exact method available, and the distance that remains — the ghosts and the shadows — is what no method can recover. That distance is the work.

Chance decides, it is always the protagonist. In the search for truth, starting from a different point, the network changes: some bonds appear, others vanish. Every run produces a slightly different work — not arbitrary, but real.

We know the answer and we search for it anyway, from the data, with rigour. The loss is inevitable and intended: it's the difference between a thing and its numerical shadow.

→ Back to the works · discover why these eight molecules.