Table 1.

The number of TCGA samples and the performance of TINDL in predicting their CDR for 14 drugs

Drug	Number of clinical samples	Number of sensitive samples	Number of resistant samples	P value
Cisplatin	303	237	66	6.36E−4
Tamoxifen	20	14	6	1.14E−3
Etoposide	84	73	11	4.00E−3
Doxorubicin	100	68	32	1.42E−2
Paclitaxel	158	111	47	2.29E−2
Vinorelbine	30	23	7	2.41E−2
Oxaliplatin	54	33	21	2.41E−2
Temozolomide	95	11	84	2.94E−2
Bleomycin	52	46	6	3.41E−2
Gemcitabine	157	75	82	4.57E−2
Cyclophosphamide	101	96	5	5.60E−2
Pemetrexed	38	18	20	2.86E−1
Irinotecan	23	6	17	3.04E−1
Docetaxel	102	67	35	7.04E−1

Note: P values were calculated by a one-sided Mann–Whitney U test to determine if TINDL can distinguish between sensitive and resistant patients. To ensure the results are not biased by the initialization of the parameters of model, TINDL was trained using ten random initializations, and the mean aggregate of its prediction was used to calculate the P values. Drugs were sorted based on their associated P values. TINDL, deep learning pipeline with tissue-informed normalization; TCGA, The Cancer Genome Atlas; CDR, cancer drug response.