Adaptive One-Class Gaussian Processes (Prioritization of Oncology Drug Targets)

Python implementation of Adaptive One-Class Gaussian Processes, starting from OCGP of Kemmler et al. [1].

Since the OCC problem does not allow the automatic selection of GP hyperparameters, we propose two kernels that use not a hyperparameter with a fixed value but an adaptive hyperparameter, which varies for each sample and is automatically selected according to the distance of the training samples.

• Adaptive Kernel the hyperparameter is based on the distance of the training sample from its k-th nearest training sample
• Scaled Kernel combines the distance between the samples with the average distance of the samples from their nearest.

The proposed kernels are compared with an implementation of the best known method for the hyperparameter selection for OCC of Xiao et al. [2]. The main problem addressed is the selection and prioritization of drug targets.

Example of usage

ocgp = OCGP.OCGP()

Preprocessing of dataset

• Possible preprocessing: "minmax", "zscore"

• Optional PCA

    scaleType = "minmax"
    pca = True
    X_train, X_test = ocgp.preprocessing(X_train, X_test, scaleType , pca)
  

Kernels

Squared Exponential Kernel

• ls (lenght-scale hyperpameter)

• signal variance (OPTIONAL)

   ls = 0.3
   ocgp.adaptiveKernel(X_train, X_test, ls)
  

Adaptive Kernel

• p (number of neighbors considered to determine adaptive hyperparameters)

• signal variance (OPTIONAL)

    p = 30
    ls = ocgp.adaptiveHyper(X_train,p)
    ocgp.adaptiveKernel(X_train, X_test, ls)
  

Scaled Kernel

• v (usually in [0.3, 0.8])

• N (number of neighbors considered in the average)

• signal variance (OPTIONAL)

    v = 0.8
    N = 4
    meanDist_xn, meanDist_yn = ocgp.scaledHyper(X_train, X_test, N)
    ocgp.scaledKernel(X_train, X_test, v, meanDist_xn, meanDist_yn)
  

Get Scores

• Possible scores: "mean", "var", "pred", "ratio".

    scoreType = "mean"
    scores = ocgp.getGPRscore(scoreType)
  

UCI Benchmars

The following table shows AUC scores for mean and negative variance on UCI datasets (downloaded from url), comparing the proposed kernels and implementation of the hyperparameter selection of Xiao et al. [2].

Dataset	Xiao(mean)	Xiao(-var)	Adapt.(mean)	Adapt.(-var)	Scaled(mean)	Scaled(-var)
Abalone	0,7894	0,7897	0,7745	0,7428	0,7742	0,7092
Balance	0,8366	0,8735	0,9468	0,9682	0,8657	0,9402
Biomed	0,8998	0,9036	0,9028	0,8960	0,9073	0,9117
Heart	0,8339	0,8379	0,8093	0,7925	0,8408	0,8135
Hepatitis	0,8378	0,8379	0,8006	0,7794	0,8242	0,7963
Housing	0,7917	0,7874	0,8677	0,8680	0,8107	0,8492
Ionosphere	0,9265	0,9504	0,9550	0,9649	0,9697	0,9712
Vehicle	0,5183	0,5714	0,7965	0,8656	0,6855	0,8187
Waveform	0,7497	0,8004	0,7808	0,8167	0,8024	0,7998
AVERAGE	0,7982	0,8169	0,8482	0,8549	0,8312	0,8455

Cite:

[1] Antonio de Falco, Zoltan Dezso, Francesco Ceccarelli, Angelo Ciaramella, and Michele Ceccarelli: "Adaptive One-Class Gaussian Processes Allow Accurate Prioritization of Oncology Drug Targets", 2020 (To appear)

[2] Zoltan Dezso and Michele Ceccarelli: "Machine learning prediction of oncology drug targets based on protein and network properties", BMCBioinformatics, 2020.

[3] Michael Kemmler and Erik Rodner and Joachim Denzler: "One-Class Classification with Gaussian Processes", Proceedings of the 10th Asian Conference on Computer Vision, 2010.

[4] Yingchao Xiao, Huangang Wang, and Wenli Xu: "Hyperparameter Selection for Gaussian Process One-Class Classification", IEEE Transactions on Neural Networks and Learning Systems, 2015.