Gene expression prediction using low-rank matrix completion

BMC Bioinformatics

Table 4 Comparison of the results of classification obtained using Bayesian networks learnt on low observability predicted datasets with those in which networks were learnt on original datasets

Study	Dataset	True positive rate	False positive rate	Precision	Recall	F-measure	AUROC
Lung adenocarcinoma	Original	0.944	0.057	0.944	0.944	0.944	0.988
	Low-rank prediction	0.944	0.057	0.944	0.944	0.944	0.996
	(O = 60 %)
	Sampled Uniform distribution	0.757	0.256	0.758	0.757	0.755	0.777
	(O = 60 %)
Myelodysplastic syndrome	Original	0.865	0.866	0.844	0.865	0.854	0.673
	Low-rank prediction	0.865	0.92	0.833	0.865	0.849	0.675
	(O = 40 %)
	Sampled Uniform distribution	0.85	0.868	0.842	0.85	0.846	0.425
	(O = 40 %)
Pulmonary hypertension	Original	0.638	0.121	0.633	0.638	0.635	0.854
	Low-rank prediction	0.681	0.118	0.645	0.681	0.659	0.897
	(O = 60 %)
	Sampled Uniform distribution	0.267	0.372	0.213	0.267	0.218	0.424
	(O = 60 %)
Pancreatic ductal	Original	0.782	0.218	0.784	0.782	0.782	0.886
adenocarcinoma	Low-rank prediction	0.821	0.179	0.821	0.821	0.82	0.905
	(O = 50 %)
	Sampled Uniform distribution	0.397	0.603	0.389	0.397	0.385	0.417
	(O = 50 %)
Psoriasis	Original	0.912	0.088	0.913	0.912	0.912	0.96
	Low-rank prediction	0.912	0.088	0.912	0.912	0.912	0.956
	(O = 40 %)
	Sampled Uniform distribution	0.641	0.359	0.641	0.641	0.641	0.648
	(O = 40 %)

Datasets were condensed and constituted of randomly selected 100 gene attributes. Bayesian networks were learned using a bottom-up search method known as K2 algorithm and evaluated in a 10-fold cross validation analysis. The predicted datasets were evaluated by comparing the classification results with those obtained using datasets constructed employing values sampled from a set uniform distribution instead of low-rank recovery, and the fraction of known values were the same in both cases. Notably, the performance of low-rank recovered datasets closely matched with that of the original datasets
Abbreviations: O observability, AUROC Area Under the Receiver Operating Characteristic curve deviation ratio

ISSN: 1471-2105