Human Intestinal Absorption Revisited: The 1998 Descriptors Survive, and the Method Extends to Permeability

Abstract

In 1998 we predicted percent human intestinal absorption for 86 drug and drug-like compounds from six molecular descriptors chosen by a genetic algorithm and a small neural network. That result rested on one division of the compounds into training, cross-validation and external prediction groups. Repeating the division shows it was a fortunate one. Across 250 stratified extractions, 14 match the published cross-validation error and only one matches the published external prediction error. An extraction is one draw from a wide distribution, so we repeated every experiment 250 times and report the result as a distribution. Reconstructed this way, the published six descriptors give a mean extraction error roughly twice the external prediction value we reported in 1998 (Table 1). We then applied the same six descriptors, unchanged, to a larger set of 780 absorption measurements from a public compilation. Training and extraction error come to within fifteen percent of one another, and under a random forest the six still reach R 0.63. Four datasets are modeled here: 780 small molecules with percent absorbed measured in man, 3,771 small molecules and 1,281 cyclic peptides measured across Caco-2 cell monolayers, and 7,298 cyclic peptides measured across artificial membranes. Each is its own prediction problem, and one procedure is applied to all four. A genetic algorithm searches descriptor sets whose size is free to vary, scoring each candidate by the out-of-bag error of a random forest fitted inside that extraction’s training group. The descriptors chosen most often across the 250 extractions form one consensus set for each dataset, and that set is what we deploy. For absorption, the deployed twenty-five descriptor set reaches an extraction rms of 20.47 %HIA units at R 0.76, against 25.77 at R 0.59 for the six 1998 descriptors under the 1998 network. The four models are served at predicthia.ai, where a structure can be drawn or pasted, scored against its training set for similarity, and the training data with fitted models downloaded. Each model served is its fixed consensus set fitted on every compound in its dataset; the accuracy quoted beside a prediction comes from evaluating that same set across 250 extractions (Table 3). The fixed sets hold 25 descriptors for absorption, 55 for Caco-2 small-molecule permeability, 50 for Caco-2 peptide permeability and 25 for artificial-membrane permeability. Graphical abstract The search, in the layout of the original study. Left, the genetic algorithm cycle, with fitness now the out-of-bag error of a random forest. Middle, four candidate descriptor sets drawn at their own lengths, one cell per descriptor, each scored by its own forest, and below them a crossover worked through as in 1998. Right, the distribution of extraction error across 250 extractions of the 780 absorption compounds, the right panel of Figure 1, and the deployed twenty-five descriptor model against measurement.