Data

What each of the three models was actually fitted to, where those measurements came from, and what the algorithm chose from them. Nothing on this site is fitted to anything that is not on this page.

Absorption

The absorption model is fitted to 781 compounds carrying a continuous percent-absorbed label, drawn from a compilation of 783 compounds. Two of the 783 carry no usable continuous label and are not fitted. The label is percent of an oral dose absorbed in humans, and it runs from 0 to 100, so the model is a regression and not a two-class call.

The labels are published measurements collected from the literature, which means they carry the literature's disagreements with them. Of the 783 compounds, 262 carry more than one published value; 61 of those disagree by more than 20 points and 31 by more than 40. Methotrexate appears at 20, 59, 65, 70 and 100. Where a compound carries several values, the median is used, and the spread is not discarded: it is what sets the ceiling described on How it works.

The last row is the floor and not a result. It is what a predictor that ignores the structure and answers with the training mean scores on the same held out rows.

Peptides

The peptide measurements come from CycPeptMPDB, the public database of cyclic peptide membrane permeability. They are split into two arms by what was measured, and the split is the point.

The Caco-2 arm holds 1,281 cyclic peptides measured on a monolayer of a human intestinal cell line. The passive arm holds 7,298 cyclic peptides measured on artificial membranes of the PAMPA class. Both report a log10 apparent permeability, and that shared unit is exactly what makes pooling them tempting and wrong.

These are different assays on different populations. A cell monolayer has transporters, efflux and metabolism in it; an artificial membrane has none of those. The two populations are also not the same peptides: one arm is roughly six times the size of the other and reaches further into macrocyclic chemistry. The 585 peptides that both arms contain correlate 0.557 across the two assays and sit 1.15 log units apart. An offset of that size with a correlation of that size is a systematic difference between two measurements, not noise that averaging removes.

So there are two models, they are reported side by side, and no number on this site is ever computed from both. There is no conversion between them because none is supported by this data.

The mean predictor column is the floor and not a result, and it belongs to its own arm. The two held-out errors are also not comparable with each other: the passive arm covers a wider population, so it has more spread to explain. Comparing the two errors is the same mistake as pooling the two arms.

The consensus descriptors per model

Each model uses a small descriptor set chosen by a genetic algorithm run repeatedly on different splits. The recurrence column counts how many of those repeats picked that descriptor, and it is the only claim about importance this site makes. A descriptor picked once out of twelve is a coincidence; one picked twelve times out of twelve is the search telling you something.

Absorption, 781 compounds, 12 repeats

Charge and polar surface are the two the search insists on. They are the same physics the 1998 model reached through hydrogen bonding and geometric descriptors, out of a pool that no longer exists.

Caco-2 peptide arm, 1,281 peptides, 12 repeats

Passive peptide arm, 7,298 peptides, 5 repeats

Polar surface per heavy atom and lipophilicity carry both peptide arms. N-methylation and conformational shielding were in the pool on every repeat and were not chosen, which is reported here as what happened rather than as an argument about why.

Getting the tables

The training tables and the fitted models will be served from Download. They are not open yet, and that page says so plainly rather than listing files that do not exist.

Predict absorption Predict peptide permeability