DBindR: Prediction of DNA-binding residues in proteins from amino
acid sequences using a random forest model with a hybrid feature
1. Dataset.
Supplementary Table 1 Protein chain IDs of datasets (PDB ID_chain).
|
|
1A0A_A 1D2I_A 1GTW_A 1IU3_C 1LB2_B 1OZJ_A 1RZ9_A 1XSD_A 2C62_A 2HZV_A
1A35_A 1D5Y_A 1GXP_A 1IV6_A 1LLM_C 1P4E_A 1RZR_L 1Y1W_D 2C6Y_A 2I0Q_B
1A36_A 1D66_A 1H0M_A 1IXY_A 1LQ1_A 1P7D_A 1S40_A 1Y1W_G 2C9L_Y 2I13_A
1A6Y_A 1D8Y_A 1H6F_A 1J1V_A 1LRR_A 1P8K_Z 1SA3_A 1Y6F_A 2CCZ_A 2IIE_A 1A73_A 1DC1_A 1H88_C 1J4W_A 1LWS_A 1PER_L 1SFU_A 1YNW_B 2CGP_A 2ISZ_A 1AKH_A 1DCT_A 1H89_A 1J5N_A 1M06_F 1PH1_B 1SKN_P 1YO5_C 2D45_A 2IX1_A 1AKH_B 1DEW_A 1H8A_C 1J5O_A 1M06_G 1PP8_F 1SKR_B 1YZ8_P 2D5V_A 2JEA_A 1AM9_A 1DIZ_A 1H9D_A 1J75_A 1M07_A 1PUE_E 1SVC_P 1Z1B_A 2D7D_A 2JEA_B 1AN4_A 1DMU_A 1H9T_A 1JB7_A 1M5X_A 1PUF_A 1T2K_D 1Z63_A 2DDG_A 2JEA_I 1AOI_A 1DNK_A 1HAO_H 1JB7_B 1MDM_B 1PUF_B 1T38_A 1Z9C_A 2DGC_A 2NLL_A 1AOI_B 1DP7_P 1HBX_A 1JE8_A 1MDY_A 1PV4_A 1TAU_A 1ZGW_A 2DPI_A 2NOB_A 1AOI_D 1DSZ_A 1HBX_G 1JEY_A 1MJE_A 1PYI_A 1TEZ_A 1ZLK_A 2DRP_A 2NP2_A 1AV6_A 1DUX_C 1HHT_P 1JEY_B 1MJE_B 1Q0T_A 1TF3_A 1ZME_C 2DWL_A 2NRA_C 1AWC_A 1ECR_A 1HJB_A 1JFI_A 1MM8_A 1Q9Y_A 1TF6_A 1ZQ3_P 2ER8_A 2NTC_A 1AWC_B 1EMH_A 1HLO_A 1JFI_B 1MNM_A 1QAI_A 1TQE_P 1ZS4_A 2ES2_A 2O4I_A 1AZP_A 1EO3_A 1HLV_A 1JMC_A 1MNM_C 1QBJ_A 1TRO_A 1ZTG_A 2ETW_A 2OAA_A 1B2M_A 1EQZ_C 1HWT_C 1JNM_A 1MOW_A 1QN3_A 1TSR_A 1ZX4_A 2EX5_A 2ODI_A 1B3T_A 1EQZ_D 1I3J_A 1JT0_A 1MSE_C 1QPI_A 1TTU_A 1ZZI_A 2EZV_A 2OFI_A 1B72_A 1EWN_A 1I6H_A 1K61_A 1MTL_A 1QRV_A 1U1K_A 2A07_F 2F8N_K 2OH2_A 1B72_B 1EXI_A 1I6H_B 1K6O_B 1MVM_A 1QUM_A 1U3E_M 2A0I_A 2F8X_K 2OST_A 1B8I_A 1EYG_A 1I6H_C 1K78_A 1MW8_X 1QZG_A 1U78_A 2A1R_A 2F8X_M 2OWO_A 1B8I_B 1EYU_A 1I6H_E 1K8G_A 1NFK_A 1R0N_B 1U8B_A 2A3V_A 2FD8_A 2PJR_A 1BDH_A 1F0V_A 1I6H_F 1KB2_A 1NG9_A 1R0O_A 1U8R_A 2A66_A 2FIO_A 2PJR_B 1BF5_A 1F2I_G 1I6H_H 1KBU_A 1NGM_B 1R4I_A 1UBD_C 2A6O_A 2FKC_A 2STT_A 1BG1_A 1F4K_A 1I6H_I 1KC6_A 1NH2_C 1R4O_A 1V14_A 2ACJ_A 2FO1_D 3CRO_L 1BHM_A 1F4S_P 1I6H_J 1KDH_A 1NK2_P 1R71_A 1VAS_A 2AJQ_A 2FO1_E 3HTS_B 1BRN_L 1F5T_A 1I6H_K 1KQQ_A 1NKP_A 1R7M_A 1VFC_A 2AOQ_A 2FQZ_A 3KTQ_A 1BVO_A 1FIU_A 1I6H_L 1KSY_A 1NLW_A 1R8D_A 1VRR_A 2AQ4_A 2G1P_A 3ORC_A 1C7Y_A 1FJL_A 1I7D_A 1KU7_A 1NOP_A 1RC7_A 1W36_B 2ASD_A 2GAT_A 3PJR_A 1C9B_A 1FOK_A 1IAW_A 1L1M_A 1NOY_A 1RC8_A 1W36_C 2AYB_A 2GLI_A 4GAT_A 1CEZ_A 1FOS_E 1IC8_A 1L1T_A 1NWQ_A 1REP_C 1W36_D 2B9S_A 2GXA_A 6PAX_A 1CF7_A 1FZP_B 1ID3_C 1L3S_A 1O4X_A 1RIO_A 1WVL_A 2B9S_B 2GZK_A 10MH_A 1CF7_B 1G38_A 1ID3_D 1L9Z_A 1O4X_B 1RM1_A 1X9N_A 2BGW_A 2H1O_E 1CIT_A 1G4D_A 1IF1_A 1L9Z_C 1ODG_A 1RM1_B 1XF2_B 2BOP_A 2H27_A 1CKQ_A 1GCC_A 1IG4_A 1L9Z_D 1ODH_A 1RM1_C 1XHZ_A 2BSQ_A 2H7G_X 1CMA_A 1GD2_E 1IGN_A 1L9Z_E 1ORN_A 1RRQ_A 1XJV_A 2BSQ_E 2H8C_A 1CW0_A 1GDT_A 1IHF_A 1L9Z_H 1OSB_A 1RTD_B 1XPX_A 2BZF_A 2HDC_A 1D02_A 1GM5_A 1IO4_D 1LAU_E 1OUP_A 1RXV_A 1XS9_A 2C5R_A 2HVR_A |
|
PDNA-62 |
|
1A02_F 1BF5_A 1CMA_A 1GCC_A 1HWT_D 1MDY_A 1PAR_B 1REP_C 1UBD_C 2DRP_D
1A02_J 1BHM_A 1D02_A 1GDT_A 1IF1_A 1MEY_F 1PDN_C 1SRS_A 1XBR_A 2GLI_A
1A02_N 1BL0_A 1D66_A 1HCQ_A 1IGN_A 1MHD_A 1PER_L 1SVC_P 1YRN_A 2HDC_A
1A74_A 1C0W_B 1DP7_P 1HCR_A 1IHF_A 1MNM_A 1PNR_A 1TC3_C 1YRN_B 3CRO_L 1AAY_A 1CDW_A 1ECR_A 1HDD_C 1IHF_B 1MNM_C 1PUE_E 1TF3_A 1YSA_C 1AZQ_A 1CF7_A 1FJL_A 1HLO_A 1J59_A 1MSE_C 1PVI_B 1TRO_A 1YUI_A 1B3T_A 1CJG_A 1GAT_A 1HRY_A 1LMB_4 1OCT_C 1PYI_A 1TSR_A 2BOP_A |
|
TS75 |
|
1A6Y_A 1CIT_A 1EMH_A 1HHT_P 1L9Z_H 1O4X_B 1RXV_A 1YO5_C 2C6Y_A 2PJR_B
1AKH_A 1CKQ_A 1EO3_A 1I6H_E 1LAU_E 1QN3_A 1RZ9_A 1ZGW_A 2EZV_A 3PJR_A
1AKH_B 1CW0_A 1F0V_A 1IAW_A 1M06_G 1QZG_A 1RZR_L 2A1R_A 2FO1_E 4GAT_A
1AOI_B 1D2I_A 1FIU_A 1J4W_A 1MJE_B 1R0N_B 1SKN_P 2A66_A 2GZK_A
1AOI_D 1DC1_A 1FOK_A 1J5N_A 1NGM_B 1RC8_A 1TQE_P 2ASD_A 2H7G_X 1AWC_A 1DNK_A 1GD2_E 1JNM_A 1NK2_P 1RM1_B 1U3E_M 2B9S_A 2I0Q_B 1B2M_A 1DSZ_A 1HAO_H 1JT0_A 1NKP_A 1RM1_C 1Y1W_D 2B9S_B 2JEA_I 1C9B_A 1DUX_C 1HBX_A 1K61_A 1NLW_A 1RRQ_A 1Y1W_G 2BSQ_E 2OFI_A |
DISIS (http://cubic.bioc.columbia.edu/services/disis)(Ofran, et al., 2007) predicts DNA-binding sites (6.0 A is designated as the cutoff distance in the definition of them) in a DNA-binding protein from its amino acid sequence through neural networks (NN) and SVM relying on sequence environment, evolutionary profiles and predicted structural features (secondary structure, solvent accessibility, globularity). Putative DNA-binding residues in the test dataset TS75 were predicted by DISIS with the default values for the optional parameters. We also trained RF models with the exactly same strategy as that for DP-Bind, except that the different cutoff distance (6.0 A) in the definition of a binding residue. Then, the RF models were used to predict putative DNA-binding residues in the test dataset TS75. The overall accuracy is 81.59% with a low SE 7.65% for the DISIS predictor, and the low SE value may be due to DISIS’s more focus on the accuracy of predicted positive samples (Supplementary Table 2). The total accuracy is 78.24% for the RF predictor (Supplementary Table 2).
Supplementary Table 2 . Performance comparisons with DISIS.
Classifiers |
ACC(%) |
SE (%) |
PR (%) |
SP (%) |
MCC |
DISIS |
81.59 |
7.65 |
70.37 |
99.23 |
0.190 |
RF |
78.24 |
51.38 |
44.25 |
84.62 |
0.341 |
To demonstrate that the RF-based model is a useful tool for understanding protein-nucleic acid interactions, we have applied it to predict DNA-binding residues in the archaeal TATA-box-binding protein (TBP) in the structure of TBP/TFB/promoter complex (PDB ID: 1D3U) by visualizing them in the format of three-dimensional structures. The TBP contains 181 residues and was not used for training the RF classifier. Its only homologue in the DBP-374 dataset is the wild-type TATA box-binding protein (TBP) in the structure of TBP–TATA box complex (PDB ID: 1QN3) with 38% sequence identity. The actual DNA-contacting residues were verified by Otis Littlefield et al. (Littlefield, et al., 1999) and these are E12,N13,V15,K37,F43,P44,I47,H49,L58,F60,S62,V66,T68,Q103,N104,V106,F134,P135,R140,V147,L149,F151,S153,V157 and S159. As shown in Supplementary Fig.1., twenty out of twenty-five DNA-binding residues (80.00%) are correctly identified and highlighted in red. The five residues in orange are false negatives (DNA-binding residues but predicted as negatives). For the non-binding residues, 104 of 146 (71.23%) are predicted correctly (residues in blue) and 42 are wrongly predicted (residues in yellow). The total 10 residues located at the N-terminal and C-terminal of the TBP were not used for reporting prediction performance by the RF model and shown in light pink. The results show that DNA-binding residues were predicted by the RF model at 72.51% overall accuracy with Matthew’s correlation coefficient (MCC) of 0.377, and with a sensitivity of 80.00% and a specificity of 71.23%. The overall accuracy is 79.56% and 65.14% with the low sensitivity 24.00% and 36.00% for the BindN and Ho et al predictors, and 83.42%, 84.53% and 76.24% for the SVM, KLR and PLR predictors in DP-Bind, respectively.
Supplementary Fig.1. Prediction performance of residues within the archaeal TATA-box-binding protein (TBP) in the structure of TBP/TFB/promoter complex (PDB ID: 1D3U) and its presentation in the format of three-dimensional structures. The correctly identified binding residues (true positives,TPs) are in red space fill; the correctly identified non-binding residues (true negatives,TNs) are in blue space fill; the binding residues with negative predictions (false negatives,FNs) are in orange space fill; the non-binding residues but wrongly predicted as positives (false positives,FPs) are in yellow space fill; the total 10 residues located in the N-terminal and C-terminal of the TBP protein were not used in reporting prediction performance by our model and shown in light pink space fill. The DNA molecule is indicated in green wire frame. The picture is generated with PyMOL (http://www.pymol.org). Prediction performance: ACC 72.51%, SE 80.00%, SP 71.23%, MCC 0.377
The length of the structural motifs ranges from 7 to ~20 residues in the protein-DNA complexs. Comparison of prediction performances is studied on considering different lengths of a data instance from 7 to 19 residues (Supplementary Fig.2). Compared with other window sizes, the RF classifiers constructed withβ＝11 presented the best performance.
