Cyclic peptide permeability

Draw a cyclic peptide, or paste a SMILES and it will be drawn. Two predictions come back, one per assay, side by side. There is no third number.

The two arms are never pooled. The Caco-2 arm is fitted to peptides measured on a human intestinal cell line. The passive arm is fitted to peptides measured on an artificial membrane. On the 585 peptides both arms contain, the two assays correlate 0.557 and sit 1.15 log units apart. That is a systematic offset between two different measurements, not noise to be averaged away, so this page offers no combined figure and no conversion control.

Worked examples.

Each one loads into the editor and scores itself on both arms. The measured values shown are from the training compilation and are there so you can see the model beside the data rather than on its own.

SMILES

Both arms are scored from the same structure in one request. A small drug-like molecule belongs on the absorption page instead.

What oxytocin shows

Oxytocin is a cyclic peptide, so it looks like a fair question to ask. It is not one this model can answer. The training set is drawn from a compilation of synthetic macrocycles, mostly head to tail cyclized and often N methylated. Oxytocin is a natural nonapeptide closed by a disulfide bridge, and nothing of that shape is in the data.

Both arms therefore flag it as far outside the training set, at a distance of 0.831 on the Caco-2 arm against a threshold of 0.228. A number is still returned, because hiding it would be worse than showing it with the flag attached, but the flag is the answer and the number is not. A peptide that reaches this page from a real program should be checked against that band before anything is done with the prediction.

Cyclosporine A is the opposite case, and it is the one worth trusting. It sits well inside the covered region, and the Caco-2 arm never saw it: it fell in that arm's held out set, so its agreement with the measured value is an honest test rather than a memory.

How to read the two cards

Each card carries its own bar, its own interval and its own applicability band, because each is a separate model fitted to a separate population. The Caco-2 arm holds 1,281 peptides and makes a held-out error of 0.644 log units. The passive arm holds 7,298 peptides and makes a held-out error of 0.824 log units. A predictor that answers with the training mean scores 0.834 on the Caco-2 rows and 1.120 on the passive rows. Those two are the floor, not results: they are printed so the model errors can be read against something.

The passive arm has more data and a larger error. That is not a contradiction. It covers a wider and more varied population of peptides, so there is more spread in it to explain. Comparing the two errors to each other is the same mistake as pooling the two arms.

A flagged card means the peptide sits in a thinly covered or distant region of that arm's training set. The prediction is still shown, and the flag sits above it in an amber box rather than under it in a footnote. What each arm was fitted to is set out on Data, and the descriptors the algorithm chose for them are on How it works.