A novel prediction method for protein DNA-binding residues based on neighboring residue correlations

Figures & data

Figure 1. Prediction performance (AUC) using different sliding window sizes for PDNA-62 and PDNA-224 over five-fold cross-validations..

Figure 2. An example of proposed 1-stage adjustment algorithm (a) and 2-stage adjustment algorithm (b).

Table 1. Prediction performance of individual feature and different feature combinations on PDNA-62 and PDNA-224 over five-fold cross-validations.

Dataset	Involved feature	ACC	Sen	Spe	MCC	AUC
PDNA-62	ERC	0.887	0.432	0.953	0.434	0.841
	PSS	0.838	0.436	0.896	0.313	0.746
	One-hot Encoding	0.815	0.490	0.862	0.304	0.754
	PP	0.838	0.509	0.885	0.354	0.786
	ERC + PSS	0.895	0.364	0.972	0.438	0.845
	ERC + One-hot Encoding	0.852	0.621	0.885	0.440	0.849
	ERC + PP	0.895	0.417	0.964	0.459	0.872
	ERC + PSS + One-hot Encoding	0.863	0.439	0.952	0.446	0.872
	ERC + PSS + PP	0.889	0.397	0.970	0.449	0.876
	ERC + One-hot Encoding + PP	0.878	0.493	0.943	0.462	0.881
	All Feature	0.885	0.522	0.939	0.481	0.884
PDNA-224	ERC	0.831	0.593	0.849	0.294	0.812
	PSS	0.714	0.517	0.729	0.138	0.673
	One-hot Encoding	0.723	0.620	0.731	0.197	0.728
	PP	0.858	0.388	0.894	0.217	0.762
	ERC + PSS	0.887	0.458	0.919	0.312	0.817
	ERC + One-hot Encoding	0.892	0.448	0.926	0.319	0.823
	ERC + PP	0.875	0.523	0.901	0.323	0.826
	ERC + PSS + One-hot Encoding	0.890	0.452	0.922	0.324	0.828
	ERC + PSS + PP	0.886	0.467	0.915	0.327	0.829
	ERC + One-hot Encoding + PP	0.889	0.463	0.919	0.331	0.832
	All Feature	0.899	0.421	0.938	0.336	0.837

Note: ACC, overall accuracy; Sen, sensitivity; Spe, specificity; MCC, Matthews correlation coefficient.

Figure 3. Heatmap of $P_{\mathit{NRB}-1}$ of PDNA-62 (a) and PDNA-224 (b).

Note: The x-axis and y-axis denote the 20 amino acid types, and every element denotes a specific $P_{\mathit{NRB}-1}(x,y)$.

Table 2. Prediction performance of original LightGBM outputs, 1-stage adjustment, 2-stage adjustment and higher stage adjustment algorithm on PDNA-62 and PDNA-224 over five-fold cross-validations.

Dataset	Method	Acc	Sen	Spe	MCC	AUC
PDNA-62	0-stage (original LightGBM outputs)	0.885	0.522	0.939	0.481	0.884
	1-stage	0.870	0.671	0.913	0.572	0.917
	1-stage + 2-stage	0.875	0.736	0.895	0.590	0.926
	1-stage + 2stage + 3stage	0.862	0.801	0.873	0.581	0.920
	1-srage + 2-stage + 3-stage + 4-stage	0.851	0.821	0.856	0.570	0.915
PDNA-224	0-stage (original LightGBM outputs)	0.899	0.421	0.938	0.336	0.837
	1-stage	0.889	0.541	0.916	0.359	0.861
	1-stage + 2-stage	0.878	0.558	0.904	0.371	0.866
	1-stage + 2stage + 3stage	0.866	0.591	0.886	0.356	0.862
	1-srage + 2-stage + 3-stage + 4-stage	0.858	0.621	0.876	0.349	0.859

Note: ACC, overall accuracy; Sen, sensitivity; Spe, specificity; MCC, Matthews correlation coefficient.

Figure 4. Threshold selection for 1-stage and 2-stage adjustment algorithm on PDNA-62 and PDNA-224 over five-fold cross-validations.

Table 3. Prediction performance of our proposed method and other state-of-art methods on PDNA-62 over five-fold cross-validations.

Method	Acc	Sen	Spe	MCC	AUC
Dps-pred	0.791	0.403	0.818	–	–
Dbs-pssm	0.664	0.682	0.660	–	–
BindN	0.703	0.694	0.705	-	0.752
Dp-bind	0.781	0.792	0.772	0.490	–
DP-Bind	0.772	0.764	0.766	–	–
BindN-RF	0.782	0.781	0.782	–	0.861
BindN+	0.790	0.773	0.793	0.440	0.859
PreDNA*	0.794	0.768	0.797	0.420	–
EL_PSSM-RT	0.808	0.850	0.801	0.510	0.901
Proposed Method	0.875	0.736	0.895	0.590	0.926

Note: ACC, overall accuracy; Sen, sensitivity; Spe, specificity; MCC, Matthews correlation coefficient.

* PreDNA without using structural information.

Table 4. Prediction performance of our proposed method and other state-of-art methods on PDNA-224 over five-fold cross-validations.

Method	Acc	Sen	Spe	MCC	AUC
PreDNA*	0.791	0.695	0.798	0.290	–
EL_PSSM-RT	0.781	0.796	0.780	0.340	0.865
Proposed Method	0.889	0.541	0.916	0.371	0.866

Note: ACC, overall accuracy; Sen, sensitivity; Spe, specificity; MCC, Matthews correlation coefficient.

* PreDNA without using structural information.

Figure 5. AUC comparison between our proposed method and other prediction methods on the independent testing set TS-72.

Figure 6. The prediction result of 3PVP_A of original LightGBM outputs (a), 1-stage adjustment algorithm (b) and 1-stage and 2-stage adjustment algorithm (c). The green spheres, yellow spheres and red spheres denote the true positive (TP) instances, false negative (FN) instances and false positive (FP) instances, respectively.

Data availability statement

The data that support the findings of this study are openly available at https://github.com/tlsjz/DNAbinding.