Loading…
Molecular graph foundation model for ADMET and pharmacokinetic property prediction, pretrained by contrastive learning with graph external attention.
No providers recorded yet. Browse all providers
Ask whether a candidate will survive first-pass metabolism and the evidence base turns out to be thin: the clinical pharmacokinetic records assembled for this work amount to 1,352 compounds spanning fraction unbound, volume of distribution, clearance, mean residence time and half-life. A graph neural network fitted from scratch on a few thousand molecules memorizes those scaffolds and fails on the next chemical series, so ADMET prediction is really an exercise in importing chemical knowledge from somewhere the labels are not.
MEGA-CL, from China Pharmaceutical University with Sir Run Run Hospital of Nanjing Medical University, imports it from unlabeled PubChem. It follows the contrastive recipe established by MolCLR — two randomly augmented views of the same molecular graph pulled together under an NT-Xent objective, views of different molecules pushed apart — and changes the encoder in two ways. The message-passing backbone, GCN+, adds identity residual connections, batch normalization and a dropout-regularized feed-forward block at every layer, so depth stops collapsing node representations toward each other. On top of it sits a multi-head Graph External Attention module: instead of attending within a single molecule, each graph-level embedding queries a pair of learnable external memory matrices shared across the whole corpus.
That second piece is the substantive claim: intra-graph attention can only relate atoms to atoms, while external memory routes an unfamiliar scaffold to the behavior of related molecules seen during pretraining.
The encoder uses five graph convolution layers with 300-dimensional embeddings, a 512-dimensional output feature, mean-pooling readout and dropout of 0.2. Pretraining ran for 100 epochs with the NT-Xent loss over minibatches of 512 molecules, using Adam at an initial learning rate of 5 × 10⁻⁴. The corpus was approximately 10 million unique, unannotated SMILES strings, distilled from roughly 1.7 × 10⁸ records retrieved from PubChem by filtering duplicates, invalid structures and disconnected fragments such as salts. Downstream, encoder weights are loaded and a fresh head is trained for 100 epochs at batch size 32, under 8:1:1 scaffold splits repeated over three runs.
On MoleculeNet, MEGA-CL places best or second-best on six of seven classification benchmarks and five of six regression benchmarks, reaching 93.4 ROC-AUC on ClinTox and 90.4 on BBBP and cutting FreeSolv RMSE to 1.40 from the 2.18 of the strongest baseline, D-MPNN. The authors also assembled a 21-task ADMET benchmark — 14 classification and 7 regression tasks — from PubChem BioAssay AID 1851, PharmaBench, FDA drug-induced liver injury reports and published clinical pharmacokinetic compilations, prioritizing human measurements. There it takes the highest ROC-AUC on four of the five major CYP450 isoforms and keeps more than 75% of regression predictions within 5-fold error on every dataset, rising to more than 75% within 2-fold for fraction unbound in plasma.
The intended use is triage before assay time is committed: score a series for microsomal stability, CYP inhibition, permeability, mutagenicity or liver-injury risk, and send the survivors to the bench. A command-line predictor takes a SMILES column or a single SMILES string and writes predictions for a named task, which suits medicinal chemistry groups iterating on a scaffold and pharmacokinetics teams sequencing in vitro work. Every endpoint runs through a fine-tuned head rather than a zero-shot query, so a new endpoint means fine-tuning on labeled data for it; the released checkpoint is the encoder, not a bank of task heads.
MEGA-CL is unusual among molecular property models in carrying evaluation into the laboratory. On 18 recently approved or newly indicated FDA drugs with published microsomal data, 16 predictions fell within 2-fold of measurement and the rest within 3-fold. The authors then predicted and measured hepatic clearance and CYP450 inhibition for three preclinical compounds — oroxylin A, icariin and an undisclosed synthetic candidate — obtaining clearance within 2.5-fold and 11 of 15 inhibition calls correct. Two caveats bound it: the synthetic compound's structure is withheld, so the prospective test cannot be independently reproduced, and system-level endpoints such as clearance, half-life and mean residence time show heavier error tails than structure-driven ones, reflecting physiology the molecular graph does not contain. The work is a preprint awaiting peer review, and the code and checkpoint are public but carry no stated license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.