Loading…
Chemical perturbation model generating post-treatment transcriptomes for compounds and cell lines never screened, from SMILES structure and dose.
No providers recorded yet. Browse all providers
A high-throughput screen measures a fixed grid. The LINCS L1000 project logged more than a million bulk expression observations, but they cover 17,202 compounds across 82 cell lines, and a model that learns one lookup embedding per compound inherits that boundary — it interpolates inside the screen and has nothing to say about a molecule never assayed. PRnet removes the lookup. Its Perturb-adapter builds the perturbation vector out of the compound's own chemistry: RDKit turns each SMILES string into a 1024-bit functional-class fingerprint (FCFP4), the fingerprints are weighted by the base-10 logarithm of their doses and summed, and a two-layer network compresses the result to a 64-dimensional embedding. Any molecule that can be written as SMILES therefore has coordinates in that space, and because the terms are summed, a combination of compounds is the sum of their contributions.
PRnet was built by Xiaoning Qi and colleagues in Yi Zhao's group at the Institute of Computing Technology, Chinese Academy of Sciences, with the biotherapy group at West China Hospital running the cell-viability work, and published in Nature Communications in 2024. Two further components surround the adapter: a Perturb-encoder that maps the perturbation embedding together with an unperturbed transcriptional profile into a 64-dimensional latent state, and a Perturb-decoder that returns a per-gene Gaussian mean and variance from which the post-treatment profile is sampled. Because the output is a distribution rather than a point estimate, the model expresses heterogeneous responses among cells of the same line.
Neighbouring entries attack the same task differently. UniCure and DrugPT replace fingerprints with pretrained molecular and cellular foundation-model encoders, and both use PRnet as a benchmark baseline. GEARS handles the genetic-perturbation version of the problem; PRnet's scope is chemical perturbation only.
All three components are two-layer feedforward networks trained jointly under a Gaussian negative log-likelihood loss, with a noise vector concatenated into the decoder input for robustness. Parameter counts are modest — the released bulk checkpoint holds roughly 550,000 parameters and the single-cell one about 2.1 million — because RDKit supplies the molecular representation rather than a learned encoder. Training used 883,269 bulk L1000 profiles drawn from 175,549 compound-dose-cell-line conditions, spanning 17,202 compounds and 82 cell lines, plus 290,888 sci-Plex3 single-cell profiles over 188 compounds across three cancer cell lines at four doses. Evaluation split the data 6:2:2 by perturbation attribute — random, unseen compound, unseen cell line, unseen pathway — with five-fold cross-validation. On bulk data PRnet reached an average Pearson correlation of 0.8 on log fold-change for unseen compounds and improved on the next best method by more than 0.3 for unseen cell lines. On single-cell data it reached an R² of 0.969 for unseen compounds and 0.97 for unseen pathways.
The trained model screens arbitrary compound libraries against a chosen cell line, and the authors used it to build an atlas of over 25 million predicted post-perturbation profiles covering 935 FDA-approved drugs, 4,158 active compounds, 30,456 natural compounds, 29,670 drug-like compounds and 54 GTEx tissues, from which they produced ranked candidate lists for 233 diseases. Four nominations were tested at the bench: SEL120-34A and (+)-fangchinoline inhibited proliferation across six small-cell lung cancer lines at IC50 values inside the predicted concentration window, and 7-methoxyrosmanol and mulberrofuran Q moderately inhibited the viability of colorectal cancer lines.
Structure-based conditioning is what makes a perturbation model usable as a screening instrument rather than a within-atlas interpolator, and the published atlas gives that capability away as a static resource for groups without GPUs. Both trained checkpoints ship inside the code repository and the processed training data is archived on Zenodo, so later perturbation-response models benchmark against the released weights directly. The limits are stated plainly by the authors: SMILES fingerprints discard 3D geometry and conformational flexibility; the reverse-signature premise inherited from the Connectivity Map does not hold for every disease; scoring is at the gene level, with no phenotypic endpoint such as IC50 predicted directly; and the compound validations are cell-viability assays, not animal or clinical results.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.