Loading…
Protein druggability scoring that concatenates frozen ESM-2 sequence embeddings with Gene Ontology annotations to rank candidate drug targets.
No providers recorded yet. Browse all providers
Drugs bind pockets in a folded protein, so 3D structure looks like the obvious substrate for judging whether a protein is druggable at all. DrugTar's authors tested that assumption head-on. They fed residue contact maps into a graph neural network, then re-ran the same network with each contact map replaced by a random graph, and again with an identity matrix. All three runs returned an AUC of 0.78. The structural input was carrying nothing the sequence embedding did not already hold.
That result reframes druggability prediction as a sequence problem, and DrugTar is the model built on it. Developed in the Department of Electrical and Computer Engineering at Isfahan University of Technology with the Regenerative Medicine Research Center at Isfahan University of Medical Sciences, it uses the frozen 650M-parameter ESM-2 checkpoint as a feature extractor: residue embeddings are mean-pooled into one protein vector, concatenated with a binary Gene Ontology annotation vector, reduced by SVM-based feature selection, and passed through a small three-hidden-layer network that emits a single druggability score between zero and one.
The gap it fills sits downstream of omics. Disease studies routinely name hundreds of implicated proteins, yet close to 90% of the human proteome is untargeted by any approved drug, and nothing in a differential-expression table says which candidates a medicinal chemist could work on. DrugTar ranks such a list. DrugPLMFormer later applied the same frozen-backbone-plus-small-head shape to the same task, benchmarking against DrugTar's dataset.
Embeddings come from esm2_t33_650M_UR50D, taking the 1,280-unit output of the 33rd layer per
residue; sequences beyond 1,024 tokens are truncated to bound memory. GO membership is encoded
as a 12,610-dimensional binary vector spanning molecular function, cellular component and
biological process. The concatenation is cut to 4,000 features by SVM feature selection, run
only on training folds so the held-out fold stays unseen. The trained head is dense layers of
128, 64 and 32 units with batch normalisation, ReLU, 0.5 dropout and L2 regularisation, ending
in a sigmoid — about 523,000 parameters in the released checkpoint, since the backbone is never
updated. Ten-fold cross-validation on ProTar-II gives AUC and AUPRC of 0.94, Cohen's kappa of
0.77 and Matthews correlation of 0.77; ProTar-II-Ind scores are comparable without refitting,
as are results on a disjoint negative set and a 2,034/4,068 imbalanced variant. On the Jamali
benchmark DrugTar reaches an AUC of 0.970 and F1 of 0.920 under cross-validation, against
0.950 and 0.914 for SPIDER.
The intended use is early target triage: take the protein list a genomics, transcriptomics or network study produced, score every entry, and decide where assay development and screening budget should go. Because the backbone is frozen and the head is tiny, embeddings can be computed once and reused, so proteome-scale scoring is cheap after the first pass, and groups without GPUs can run the same model through the web server. Sequence-only operation also extends coverage to proteins with no deposited structure, which is where pocket-detection methods stop.
DrugTar's most transferable contribution is the negative result: within the algorithms tested, protein sequence — pretrained embeddings in particular — carried more druggability signal than explicit structure, against the intuition that pocket geometry should dominate. The ChEMBL check grounds the ranking externally: of the ten highest-scoring proteins held out of training entirely, seven are associated with bioactive compounds, averaging 482 ± 314 of them, while the ten lowest-scoring have none and proteins drawn at random average 8 ± 6. The work also marks evaluation practice, showing how negative-set construction inflates reported accuracy on an established benchmark. Its limits are scope rather than method: predictions are computational, the 1,024-token truncation clips long proteins, and the authors state that the structure-versus-sequence conclusion holds only for the architectures compared here.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.