TDEP: Predicting Target Perturbation Response from Drug-Induced Transcriptomes

Description This record contains the processed datasets, protein representations, network data, pretrained model checkpoints, and external validation queries used in the study "TDEP: Predicting Target Perturbation Response from Drug-Induced Transcriptomes" by Jieun Sung, Sookyung Kim, and Wankyu Kim. TDEP (Target-perturbed Differential Expression Profile) is a deep-learning framework that predicts transcriptome-wide responses to target inhibition. It combines protein sequence representations generated using pretrained protein language models with graph-based propagation over a gene co-expression network. The model is trained under drug-target interaction supervision using drug-induced transcriptional profiles from the LINCS Connectivity Map (CMAP). The deposited files support model training, prediction, benchmarking, and reproduction of the analyses described in the associated manuscript. Cell-specific resources are provided for six CMAP cell lines: A375, A549, HA1E, HT29, MCF7, and PC3. Contents benchmark.tar.gz - Cell-specific drug-target interaction benchmark datasets. Each cell-line directory contains standardized training, validation, and test tables, together with the corresponding compound identifier lists. The splits are organized by drug identity to support evaluation on unseen compounds. cmap.tar.gz - Processed cell-specific drug-induced differential expression matrices derived from normalized LINCS CMAP Level 5 profiles, along with CMAP gene annotation information. The expression data cover 10,086 landmark and best-inferred genes used by TDEP. ppi.tar.gz - The Entrez Gene ID-mapped STRING co-expression network used for graph-based propagation in the Perturbed Profile Generator component of TDEP. pretrained_models.tar.gz - Cell-specific TDEP model checkpoints trained using ESM-2, ProtBERT, or Gene2Vec protein representations. Checkpoints, hyperparameter records, and evaluation summaries are included for all six cell lines. protein_embeddings.tar.gz - Precomputed UniProt protein representations generated using ESM-2, ProtBERT, and Gene2Vec, including the Gene2Vec embedding model required by the public prediction pipeline. protein_metadata.tar.gz - UniProt-to-protein-sequence and UniProt-to-Entrez-Gene-ID mapping tables used to construct protein representations and connect target proteins to the gene network. queries.tar.gz - A processed set of newly released BindingDB drug-target interactions used for prospective external validation of TDEP. README.md - Description of the expected directory layout and file-naming conventions. checksums.md5 - MD5 checksums for verifying the integrity of the downloaded archives. Usage After downloading, extract each archive into a common data/ directory while preserving the internal directory names: data/benchmark/cmap/ppi/pretrained_models/protein_embeddings/protein_metadata/queries/ The accompanying TDEP source code resolves these resources relative to the data/ directory and is available at https://github.com/Jieun-Sung/TDEP Data provenance The deposited resources consist of processed or model-ready artifacts derived from publicly available databases, including LINCS Connectivity Map, STRING, UniProt, BindingDB, BioSNAP, DAVIS, KIBA, Hetionet, STITCH, TTD, ChEMBL, and TargetRX. Please also cite the relevant original databases when using the corresponding derived resources. Scope and limitations TDEP requires drug-induced CMAP expression profiles and is therefore limited to compounds represented in CMAP. The provided models were trained on six established, primarily cancer-derived cell lines and may not directly represent primary tissues or other biological contexts. TDEP predictions are intended to support target prioritization, mechanistic hypothesis generation, and drug-repurposing research; experimental validation remains necessary.