# DeepDDS core data files The provenance and counts below are translated from the existing dataset notes. This path/comment audit did not recompute counts or validate the source studies. ## Data sources - **Drug synergy:** the O'Neil et al. (2016) high-throughput screen of pairwise combinations of 38 anticancer drugs across 39 human cancer cell lines. - **Synergy scores:** Loewe Additivity Scores computed with Combenefit from 4 x 4 dose-response matrices. Replicate measurements for the same combination were averaged. - **Cell-line expression:** CCLE RNA-seq TPM data released in 2018 (`CCLE_RNAseq_rsem_transcripts_tpm_20180929`), describing baseline expression under standard culture conditions without drug treatment. ## Binary labels | Condition | Label | Meaning | |---|---|---| | Loewe score > 10 | 1 | Synergistic | | Loewe score < 0 | 0 | Antagonistic | | 0 <= score <= 10 | Excluded | Ambiguous range excluded from modeling | The original notes describe the stricter positive threshold as intended to reduce near-zero noise and obtain roughly balanced classes. ## Previously recorded dataset size - 12,415 unique drug-pair/cell-line combinations. - 36 anticancer drugs and 31 human cancer cell lines. - 6,116 positive and 6,299 negative samples. ## File descriptions ### `new_labels_0_10.csv`: binary training labels The recorded layout is 12,415 records plus one header row. | Column | Type | Description | |---|---|---| | drug1 | string | SMILES representation of drug A | | drug2 | string | SMILES representation of drug B | | cell | string | Cell-line name, such as A375 or HCT116 | | label | int | 0 = antagonistic (score < 0); 1 = synergistic (score > 10) | ### `score_0_10.csv`: continuous synergy scores The existing notes record 12,415 records plus a header, corresponding to `new_labels_0_10.csv`. | Column | Type | Description | |---|---|---| | drug1 | string | Drug A name, such as 5-FU | | drug2 | string | Drug B name, such as ABT-888 | | cell | string | Cell-line name | | score | float | Continuous Loewe Additivity Score | | nor | float | Normalized score | | index | int | Sample index | | true_label | int | Binary ground-truth label | | predict_label | int | Legacy predicted label | | predict_score | float | Legacy prediction probability | | smile1 | string | Drug A SMILES | | smile2 | string | Drug B SMILES | The stored prediction columns are historical reference results, not results from the current BioMod checkout. ### `new_cell_features_954.csv`: cell-line expression - Recorded dimensions: 31 cell lines x 954 genes. - Source: CCLE baseline RNA-seq TPM, without drug treatment. - Row 1: gene IDs (ENSG). - Row 2: transcript IDs (ENST). - Rows 3-33: expression values, with the cell-line name in the first column. - Purpose: cell features encoded by fully connected layers and combined with drug graph features. ### `cell_gene954.csv`: gene identifier table - Recorded layout: 954 records plus a header. - Columns: `gene_id` (ENSG), `transcript_id` (ENST). - Purpose: identify the genes in `new_cell_features_954.csv`. ### `smiles_0_10.csv`: drug SMILES table - The legacy notes record 36 headerless rows; counts may differ from the parser-based summary in the parent directory. - Columns: drug name and SMILES representation. - Purpose: map drug names to SMILES used for molecular-graph construction. ### `drugs_0_10.csv`: drug table - Recorded layout: 36 headerless rows. - Columns: drug name, SMILES, numeric encoding. - Purpose: enumerate drugs and their encoded indices. ### `cell_line_0_10.csv`: cell-line table - Recorded layout: 31 headerless rows. - Columns: cell-line name and cancer type, such as melanoma, ovarian, or colon. - Purpose: associate cell lines with cancer types. ## DeepDDS inputs ```text Drug A SMILES -> Molecular graph -> GAT/GCN -> Drug A embedding --+ | Drug B SMILES -> Molecular graph -> GAT/GCN -> Drug B embedding --+-> Concatenate -> FC layers -> Synergy prediction | Cell line (954 genes, TPM) -> FC layers -> Cell embedding -------+ ``` ## References retained from the original notes - Wang et al. "DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations." Briefings in Bioinformatics, 2022. - O'Neil et al. "An unbiased oncology compound screen to identify novel combination strategies." Molecular Cancer Therapeutics, 2016. - CCLE: Cancer Cell Line Encyclopedia, Broad Institute.