The following figure shows the pipeline of DeepQuality. Generally, modern mass spectrometers have a mass range of 0~2000 Dalton and accuracy of 0.01 Dalton, and generate centroid mass spectra each containing roughly several hundred peaks, resulting in highly sparse signals (left panel). Traditional methods used machine learning upon handcrafted features to distinguish between spectra of high and low quality. However, the handcrafted features are difficult to be optimized. Compressed Sensing (CS) theory, mathematically, can recover signals with certain sparsity from far few samples than acquired by the Nyquist rate, and thus we combine CS and deep learning for end-to-end mass spectrum quality assessment (central panel). On two publicly available datasets, DeepQuality achieved AUC of 0.96 and 0.92, significantly surpasses other software (right panel).
By virtue of mshadow library, DeepQuality can run seamlessly on both CPU and GPU.
The two datasets used here is from the paper and can be downloaded here.
The precompiled binary can be found in bin directory.
Fill in the Config.json file and run as:
CompressedSensing.exe Config.json
Four requirements are needed to compile the source code and run the software:
- rapidjson: https://github.com/miloyip/rapidjson
- mshadow: https://github.com/dmlc/mshadow
- MXNet: https://github.com/apache/incubator-mxnet
- OpenBLAS: http://www.openblas.net/