A comparison of embedding aggregation strategies in drug-target interaction prediction

The prediction of interactions between novel drugs and biological targets is a vital step in the early stage of the drug discovery pipeline. Many deep learning approaches have been proposed over the last decade, with a substantial fraction of them sharing the same underlying two-branch architecture. Their distinction is limited to the use of different types of feature representations and branches (multi-layer perceptrons, convolutional neural networks, graph neural networks and transformers). In contrast, the strategy used to combine the outputs (embeddings) of the branches has remained mostly the same. The same general architecture has also been used extensively in the area of recommender systems, where the choice of an aggregation strategy is still an open question. In this work, we investigate the effectiveness of three different embedding aggregation strategies in the area of drug-target interaction (DTI) prediction. We formally define these strategies and prove their universal approximator capabilities. We then present experiments that compare the different strategies on benchmark datasets from the area of DTI prediction, showcasing conditions under which specific strategies could be the obvious choice.

Budget-Based Classification of Parkinson's Disease From Resting State EEG

Early detection is vital for future neuroprotective treatments of Parkinson's disease (PD). Resting state electroencephalographic (EEG) recording has shown potential as a cost-effective means to aid in detection of neurological disorders such as PD. In this study, we investigated how the number and placement of electrodes affects classifying PD patients and healthy controls using machine learning based on EEG sample entropy. We used a custom budget-based search algorithm for selecting optimized sets of channels for classification, and iterated over variable channel budgets to investigate changes in classification performance. Our data consisted of 60-channel EEG collected at three different recording sites, each of which included observations collected both eyes open (total N = 178) and eyes closed (total N = 131). Our results with the data recorded eyes open demonstrated reasonable classification performance (ACC = .76; AUC = .76) with only 5 channels placed far away from each other, the selected regions including right-frontal, left-temporal and midline-occipital sites. Comparison to randomly selected subsets of channels indicated improved classifier performance only with relatively small channel-budgets. The results with the data recorded eyes closed demonstrated consistently worse classification performance (when compared to eyes open data), and classifier performance improved more steadily as a function of number of channels. In summary, our results suggest that a small subset of electrodes of an EEG recording can suffice for detecting PD with a classification performance on par with a full set of electrodes. Furthermore our results demonstrate that separately collected EEG data sets can be used for pooled machine learning based PD detection with reasonable classification performance.

Quicksort leave-pair-out cross-validation for ROC curve analysis

Receiver Operating Characteristic (ROC) curve analysis and area under the ROC curve (AUC) are commonly used performance measures in diagnostic systems. In this work, we assume a setting, where a classifier is inferred from multivariate data to predict the diagnostic outcome for new cases. Cross-validation is a resampling method for estimating the prediction performance of a classifier on data not used for inferring it. Tournament leave-pair-out (TLPO) cross-validation has been shown to be better than other resampling methods at producing a ranking of data that can be used for estimating the ROC curves and areas under them. However, the time complexity of TLPOCV, $$O\left( n^2\right)$$ O n 2 , means that it is impractical in many applications. In this article, a method called quicksort leave-pair-out cross-validation (QLPOCV) is presented in order to decrease the time complexity of obtaining a reliable ranking of data to $$O\left( n\log n\right)$$ O n log n . The proposed method is compared with existing ones in an experimental study, demonstrating that in terms of ROC curves and AUC values QLPOCV produces as accurate performance estimation as TLPOCV, outperforming both k-fold and leave-one-out cross-validation.

Generalized vec trick for fast learning of pairwise kernel models

Pairwise learning corresponds to the supervised learning setting where the goal is to make predictions for pairs of objects. Prominent applications include predicting drug-target or protein-protein interactions, or customer-product preferences. In this work, we present a comprehensive review of pairwise kernels, that have been proposed for incorporating prior knowledge about the relationship between the objects. Specifically, we consider the standard, symmetric and anti-symmetric Kronecker product kernels, metric-learning, Cartesian, ranking, as well as linear, polynomial and Gaussian kernels. Recently, a $$O(nm+nq)$$ O ( n m + n q ) time generalized vec trick algorithm, where $$n$$ n , $$m$$ m , and $$q$$ q denote the number of pairs, drugs and targets, was introduced for training kernel methods with the Kronecker product kernel. This was a significant improvement over previous $$O(n^2)$$ O ( n 2 ) training methods, since in most real-world applications $$m,q<< n$$ m , q < < n . In this work we show how all the reviewed kernels can be expressed as sums of Kronecker products, allowing the use of generalized vec trick for speeding up their computation. In the experiments, we demonstrate how the introduced approach allows scaling pairwise kernels to much larger data sets than previously feasible, and provide an extensive comparison of the kernels on a number of biological interaction prediction tasks.

Adaptive risk prediction system with incremental and transfer learning

Currently, popular methods for prenatal risk assessment of fetal aneuploidies are based on multivariate probabilistic modelling, that are built on decades of scientific research and large-scale multi-center clinical studies. These static models that are deployed to screening labs are rarely updated or adapted to local population characteristics. In this article, we propose an adaptive risk prediction system or ARPS, which considers these changing characteristics and automatically deploys updated risk models. 8 years of real-life Down syndrome screening data was used to firstly develop a distribution shift detection method that captures significant changes in the patient population and secondly a probabilistic risk modelling system that adapts to new data when these changes are detected. Various candidate systems that utilize transfer -and incremental learning that implement different levels of plasticity were tested. Distribution shift detection using a windowed approach provides a computationally less expensive alternative to fitting models at every data block step while not sacrificing performance. This was possible when utilizing transfer learning. Deploying an ARPS to a lab requires careful consideration of the parameters regarding the distribution shift detection and model updating, as they are affected by lab throughput and the incidence of the screened rare disorder. When this is done, ARPS could be also utilized for other population screening problems. We demonstrate with a large real-life dataset that our best performing novel Incremental-Learning-Population-to-Population-Transfer-Learning design can achieve on par prediction performance without human intervention, when compared to a deployed risk screening algorithm that has been manually updated over several years.

Modeling drug combination effects via latent tensor reconstruction

Motivation

Combination therapies have emerged as a powerful treatment modality to overcome drug resistance and improve treatment efficacy. However, the number of possible drug combinations increases very rapidly with the number of individual drugs in consideration, which makes the comprehensive experimental screening infeasible in practice. Machine-learning models offer time- and cost-efficient means to aid this process by prioritizing the most effective drug combinations for further pre-clinical and clinical validation. However, the complexity of the underlying interaction patterns across multiple drug doses and in different cellular contexts poses challenges to the predictive modeling of drug combination effects.

Results

We introduce comboLTR, highly time-efficient method for learning complex, non-linear target functions for describing the responses of therapeutic agent combinations in various doses and cancer cell-contexts. The method is based on a polynomial regression via powerful latent tensor reconstruction. It uses a combination of recommender system-style features indexing the data tensor of response values in different contexts, and chemical and multi-omics features as inputs. We demonstrate that comboLTR outperforms state-of-the-art methods in terms of predictive performance and running time, and produces highly accurate results even in the challenging and practical inference scenario where full dose–response matrices are predicted for completely new drug combinations with no available combination and monotherapy response measurements in any training cell line.

Availability and implementation

comboLTR code is available at https://github.com/aalto-ics-kepaco/ComboLTR.

Supplementary information

Supplementary data are available at Bioinformatics online.

Detection of Prostate Cancer Using Biparametric Prostate MRI, Radiomics, and Kallikreins: A Retrospective Multicenter Study of Men With a Clinical Suspicion of Prostate Cancer

BACKGROUND

Accurate detection of clinically significant prostate cancer (csPCa), Gleason Grade Group ≥ 2, remains a challenge. Prostate MRI radiomics and blood kallikreins have been proposed as tools to improve the performance of biparametric MRI (bpMRI).

PURPOSE

To develop and validate radiomics and kallikrein models for the detection of csPCa.

STUDY TYPE

Retrospective.

POPULATION

A total of 543 men with a clinical suspicion of csPCa, 411 (76%, 411/543) had kallikreins available and 360 (88%, 360/411) did not take 5-alpha-reductase inhibitors. Two data splits into training, validation (split 1: single center, n = 72; split 2: random 50% of pooled datasets from all four centers), and testing (split 1: 4 centers, n = 288; split 2: remaining 50%) were evaluated.

FIELD STRENGTH/SEQUENCE

A 3 T/1.5 T, TSE T2-weighted imaging, 3x SE DWI.

ASSESSMENT

In total, 20,363 radiomic features calculated from manually delineated whole gland (WG) and bpMRI suspicion lesion masks were evaluated in addition to clinical parameters, prostate-specific antigen, four kallikreins, MRI-based qualitative (PI-RADSv2.1/IMPROD bpMRI Likert) scores.

STATISTICAL TESTS

For the detection of csPCa, area under receiver operating curve (AUC) was calculated using the DeLong's method. A multivariate analysis was conducted to determine the predictive power of combining variables. The values of P-value < 0.05 were considered significant.

RESULTS

The highest prediction performance was achieved by IMPROD bpMRI Likert and PI-RADSv2.1 score with AUC = 0.85 and 0.85 in split 1, 0.85 and 0.83 in split 2, respectively. bpMRI WG and/or kallikreins demonstrated AUCs ranging from 0.62 to 0.73 in split 1 and from 0.68 to 0.76 in split 2. AUC of bpMRI lesion-derived radiomics model was not statistically different to IMPROD bpMRI Likert score (split 1: AUC = 0.83, P-value = 0.306; split 2: AUC = 0.83, P-value = 0.488).

DATA CONCLUSION

The use of radiomics and kallikreins failed to outperform PI-RADSv2.1/IMPROD bpMRI Likert and their combination did not lead to further performance gains.

LEVEL OF EVIDENCE

1 TECHNICAL EFFICACY: Stage 2.

Synthetic minority oversampling of vital statistics data with generative adversarial networks

Objective

Minority oversampling is a standard approach used for adjusting the ratio between the classes on imbalanced data. However, established methods often provide modest improvements in classification performance when applied to data with extremely imbalanced class distribution and to mixed-type data. This is usual for vital statistics data, in which the outcome incidence dictates the amount of positive observations. In this article, we developed a novel neural network-based oversampling method called actGAN (activation-specific generative adversarial network) that can derive useful synthetic observations in terms of increasing prediction performance in this context.

Materials and Methods

From vital statistics data, the outcome of early stillbirth was chosen to be predicted based on demographics, pregnancy history, and infections. The data contained 363 560 live births and 139 early stillbirths, resulting in class imbalance of 99.96% and 0.04%. The hyperparameters of actGAN and a baseline method SMOTE-NC (Synthetic Minority Over-sampling Technique-Nominal Continuous) were tuned with Bayesian optimization, and both were compared against a cost-sensitive learning-only approach.

Results

While SMOTE-NC provided mixed results, actGAN was able to improve true positive rate at a clinically significant false positive rate and area under the curve from the receiver-operating characteristic curve consistently.

Discussion

Including an activation-specific output layer to a generator network of actGAN enables the addition of information about the underlying data structure, which overperforms the nominal mechanism of SMOTE-NC.

Conclusions

actGAN provides an improvement to the prediction performance for our learning task. Our developed method could be applied to other mixed-type data prediction tasks that are known to be afflicted by class imbalance and limited data availability.

Optimized reference spectrum for rating the façade sound insulation

Objectively determined single-number-quantities (SNQs) describing the airborne sound insulation of a façade should correspond to the subjective perception of annoyance to road traffic sounds transmitted through a façade. The reference spectra for spectrum adaptation terms C and C_tr in standard ISO 717-7 (International Organization for Standardization, 2013) are not based on psycho-acoustic evidence. The aim of this study is to develop reference spectra which result in SNQs that explain the subjective annoyance of road traffic sounds transmitted through a façade well. Data from a psycho-acoustic experiment by Hongisto, Oliva, and Rekola [J. Acoust. Soc. Am. 144(2), 1100-1112 (2018)] were used. The data included annoyance ratings for road traffic sounds (five different spectrum alternatives) attenuated by the façade (twelve different sound insulation spectrum alternatives), rated by 43 participants. The reference spectrum for each road traffic spectrum was found using mathematical optimization. The performance of the acquired SNQs was estimated with nested cross-validation. The SNQs determined with the optimized reference spectra performed better than the existing SNQs for two road traffic spectra out of five and for an aggregate of the five road traffic sound types. The results can be exploited in the development of standardized SNQs.

Effect of oat β-glucan of different molecular weights on fecal bile acids, urine metabolites and pressure in the digestive tract - A human cross over trial

While the development of oat products often requires altered molecular weight (MW) of β-glucan, the resulting health implications are currently unclear. This 3-leg crossover trial (n = 14) investigated the effects of the consumption of oat bran with High, Medium and Low MW β-glucan (average > 1000, 524 and 82 kDa respectively) with 3 consequent meals on oat-derived phenolic compounds in urine (UHPLC-MS/MS), bile acids in feces (UHPLC-QTOF), gastrointestinal conditions (ingestible capsule), and perceived gut well-being. Urine excretion of ferulic acid was higher (p < 0.001, p < 0.001), and the fecal excretion of deoxycholic (p < 0.03, p < 0.02) and chenodeoxycholic (p < 0.06, p < 0.02) acids lower after consumption of Low MW β-glucan compared with both Medium and High MW β-glucan. Duodenal pressure was higher after consumption of High MW β-glucan compared to Medium (p < 0.041) and Low (p < 0.022) MW β-glucan. The MW of β-glucan did not affect gut well-being, but the perceptions between females and males differed.

GeFeS: A generalized wrapper feature selection approach for optimizing classification performance

In this paper, we propose a generalized wrapper-based feature selection, called GeFeS, which is based on a parallel new intelligent genetic algorithm (GA). The proposed GeFeS works properly under different numerical dataset dimensions and sizes, carefully tries to avoid overfitting and significantly enhances classification accuracy. To make the GA more accurate, robust and intelligent, we have proposed a new operator for features weighting, improved the mutation and crossover operators, and integrated nested cross-validation into the GA process to properly validate the learning model. The k-nearest neighbor (kNN) classifier is utilized to evaluate the goodness of selected features. We have evaluated the efficiency of GeFeS on various datasets selected from the UCI machine learning repository. The performance is compared with state-of-the-art classification and feature selection methods. The results demonstrate that GeFeS can significantly generalize the proposed multi-population intelligent genetic algorithm under different sizes of two-class and multi-class datasets. We have achieved the average classification accuracy of 95.83%, 97.62%, 99.02%, 98.51%, and 94.28% while reducing the number of features from 56 to 28, 34 to 18, 279 to 135, 30 to 16, and 19 to 9 under lung cancer, dermatology, arrhythmia, WDBC, and hepatitis, respectively.

Developing a pain intensity prediction model using facial expression: A feasibility study with electromyography

The automatic detection of facial expressions of pain is needed to ensure accurate pain assessment of patients who are unable to self-report pain. To overcome the challenges of automatic systems for determining pain levels based on facial expressions in clinical patient monitoring, a surface electromyography method was tested for feasibility in healthy volunteers. In the current study, two types of experimental gradually increasing pain stimuli were induced in thirty-one healthy volunteers who attended the study. We used a surface electromyography method to measure the activity of five facial muscles to detect facial expressions during pain induction. Statistical tests were used to analyze the continuous electromyography data, and a supervised machine learning was applied for pain intensity prediction model. Muscle activation of corrugator supercilii was most strongly associated with self-reported pain, and the levator labii superioris and orbicularis oculi showed a statistically significant increase in muscle activation when the pain stimulus reached subjects' self -reported pain thresholds. The two strongest features associated with pain, the waveform length of the corrugator supercilii and levator labii superioris, were selected for a prediction model. The performance of the pain prediction model resulted in a c-index of 0.64. In the study results, the most detectable difference in muscle activity during the pain experience was connected to eyebrow lowering, nose wrinkling and upper lip raising. As the performance of the prediction model remains modest, yet with a statistically significant ordinal classification, we suggest testing with a larger sample size to further explore the variables that affect variation in expressiveness and subjective pain experience.

Negative Predictive Value of Biparametric Prostate Magnetic Resonance Imaging in Excluding Significant Prostate Cancer: A Pooled Data Analysis Based on Clinical Data from Four Prospective, Registered Studies

BACKGROUND

Multiparametric prostate magnetic resonance imaging (mpMRI) can be considered the gold standard in prostate magnetic resonance imaging (MRI). Biparametric prostate MRI (bpMRI) is faster and could be a feasible alternative to mpMRI.

OBJECTIVE

To determine the negative predictive value (NPV) of Improved Prostate Cancer Diagnosis (IMPROD) bpMRI as a whole and in clinical subgroups in primary diagnostics of clinically significant prostate cancer (CSPCa).

DESIGN, SETTING, AND PARTICIPANTS

This is a pooled data analysis of four prospective, registered clinical trials investigating prebiopsy IMPROD bpMRI. Men with a clinical suspicion of prostate cancer (PCa) were included.

INTERVENTION

Prebiopsy IMPROD bpMRI was performed, and an IMPROD bpMRI Likert scoring system was used. If suspicious lesions (IMPROD bpMRI Likert score 3-5) were visible, targeted biopsies in addition to systematic biopsies were taken.

OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS

Performance measures of IMPROD bpMRI in CSPCa diagnostics were evaluated. NPV was also evaluated in clinical subgroups. Gleason grade ≥3 + 4 in any biopsy core taken was defined as CSPCa.

RESULTS AND LIMITATIONS

A total of 639 men were included in the analysis. The mean age was 64 yr, mean prostate-specific antigen level was 8.9 ng/ml, and CSPCa prevalence was 48%. NPVs of IMPROD bpMRI Likert scores 3-5 and 4-5 for CSPCa were 0.932 and 0.909, respectively, and the corresponding positive predictive values were 0.589 and 0.720. Only nine of 132 (7%) men with IMPROD bpMRI Likert score 1-2 had CSPCa and none with Gleason score >7. Thus, 132 of 639 (21%) study patients could have avoided biopsies without missing a single Gleason >7 cancer in the study biopsies. In the subgroup analysis, no clear outlier was present. The limitation is uncertainty of the true CSPCa prevalence.

CONCLUSIONS

IMPROD bpMRI demonstrated a high NPV to rule out CSPCa. IMPROD bpMRI Likert score 1-2 excludes Gleason >7 PCa in the study biopsies.

PATIENT SUMMARY

We investigated the feasibility of prostate magnetic resonance imaging (MRI) with the Improved Prostate Cancer Diagnosis (IMPROD) biparametric MRI (bpMRI) protocol in excluding significant prostate cancer. In this study, highly aggressive prostate cancer was excluded using the publicly available IMPROD bpMRI protocol (http://petiv.utu.fi/multiimprod/).

Skin Conductance Response to Gradual-Increasing Experimental Pain

Patient self-reporting of pain is not always possible, in those cases automated objective pain assessment could lead to reliable pain assessment. In this context, physiological measurements have been studied and one of the promising signals is skin conductance (SC). In this study, 1Hz SC signal acquisition is performed while gradually increasing heat and electrical pain stimuli are induced. Three labeled study periods are defined based on pain stimuli presence, self-reported pain threshold and pain tolerance. Different classification and regression models are compared, together with selected SC features. The model performances are evaluated using c-index. Results show good predictability, especially for the slow tonic component decomposed from the SC signal.

Cost-effective survival prediction for patients with advanced prostate cancer using clinical trial and real-world hospital registry datasets

INTRODUCTION

Predictive survival modeling offers systematic tools for clinical decision-making and individualized tailoring of treatment strategies to improve patient outcomes while reducing overall healthcare costs. In 2015, a number of machine learning and statistical models were benchmarked in the DREAM 9.5 Prostate Cancer Challenge, based on open clinical trial data for metastatic castration resistant prostate cancer (mCRPC). However, applying these models into clinical practice poses a practical challenge due to the inclusion of a large number of model variables, some of which are not routinely monitored or are expensive to measure.

OBJECTIVES

To develop cost-specified variable selection algorithms for constructing cost-effective prognostic models of overall survival that still preserve sufficient model performance for clinical decision making.

METHODS

Penalized Cox regression models were used for the survival prediction. For the variable selection, we implemented two algorithms: (i) LASSO regularization approach; and (ii) a greedy cost-specified variable selection algorithm. The models were compared in three cohorts of mCRPC patients from randomized clinical trials (RCT), as well as in a real-world cohort (RWC) of advanced prostate cancer patients treated at the Turku University Hospital. Hospital laboratory expenses were utilized as a reference for computing the costs of introducing new variables into the models.

RESULTS

Compared to measuring the full set of clinical variables, economic costs could be reduced by half without a significant loss of model performance. The greedy algorithm outperformed the LASSO-based variable selection with the lowest tested budgets. The overall top performance was higher with the LASSO algorithm.

CONCLUSION

The cost-specified variable selection offers significant budget optimization capability for the real-world survival prediction without compromising the predictive power of the model.

Molecular Atlas of Postnatal Mouse Heart Development

Background

The molecular mechanisms mediating postnatal loss of cardiac regeneration in mammals are not fully understood. We aimed to provide an integrated resource of mRNA, protein, and metabolite changes in the neonatal heart for identification of metabolism‐related mechanisms associated with cardiac regeneration.

Methods and Results

Mouse ventricular tissue samples taken on postnatal day 1 (P01), P04, P09, and P23 were analyzed with RNA sequencing and global proteomics and metabolomics. Gene ontology analysis, KEGG pathway analysis, and fuzzy c‐means clustering were used to identify up‐ or downregulated biological processes and metabolic pathways on all 3 levels, and Ingenuity pathway analysis (Qiagen) was used to identify upstream regulators. Differential expression was observed for 8547 mRNAs and for 1199 of 2285 quantified proteins. Furthermore, 151 metabolites with significant changes were identified. Differentially regulated metabolic pathways include branched chain amino acid degradation (upregulated at P23), fatty acid metabolism (upregulated at P04 and P09; downregulated at P23) as well as the HMGCS (HMG‐CoA [hydroxymethylglutaryl‐coenzyme A] synthase)–mediated mevalonate pathway and ketogenesis (transiently activated). Pharmacological inhibition of HMGCS in primary neonatal cardiomyocytes reduced the percentage of BrdU‐positive cardiomyocytes, providing evidence that the mevalonate and ketogenesis routes may participate in regulating the cardiomyocyte cell cycle.

Conclusions

This study is the first systems‐level resource combining data from genomewide transcriptomics with global quantitative proteomics and untargeted metabolomics analyses in the mouse heart throughout the early postnatal period. These integrated data of molecular changes associated with the loss of cardiac regeneration may open up new possibilities for the development of regenerative therapies.

Prostate Cancer Risk Stratification in Men With a Clinical Suspicion of Prostate Cancer Using a Unique Biparametric MRI and Expression of 11 Genes in Apparently Benign Tissue: Evaluation Using Machine-Learning Techniques

BACKGROUND

Accurate risk stratification of men with a clinical suspicion of prostate cancer (cSPCa) remains challenging despite the increasing use of MRI.

PURPOSE

To evaluate the diagnostic accuracy of a unique biparametric MRI protocol (IMPROD bpMRI) combined with clinical and molecular markers in men with cSPCa.

STUDY TYPE

Prospective single-institutional clinical trial (NCT01864135).

SUBJECTS

Eighty men with cSPCa.

FIELD STRENGTH/SEQUENCE

3T, surface array coils. Two T₂ -weighted and three diffusion-weighted imaging (DWI) acquisitions: 1) b-values 0, 100, 200, 300, 500 s/mm² ; 2) b-values 0,1500 s/mm² ; 3) b-values 0, 2000 s/mm² .

ASSESSMENT

IMPROD bpMRI examinations were qualitatively (IMPROD bpMRI Likert score) and quantitatively (DWI-based Gleason grade score) prospectively reported. Men with IMPROD bpMRI Likert 3-5 had two targeted biopsies followed by 12-core systematic biopsies (SB); those with IMPROD bpMRI Likert 1-2 had only SB. Additionally, 2-core from normal-appearing prostate areas were obtained for the mRNA expression of ACSM1, AMACR, CACNA1D, DLX1, PCA3, PLA2G7, RHOU, SPINK1, SPON2, TMPRSS2-ERG, and TDRD1 measured by quantitative reverse-transcription polymerase chain reaction.

STATISTICAL TESTS

Univariate and multivariate analysis using regularized least-squares, feature selection and tournament leave-pair-out cross-validation (TLPOCV), as well as 10 random splits of the data in training-testing sets, were used to evaluate the mRNA, clinical and IMPROD bpMRI parameters in detecting clinically significant prostate cancer (SPCa) defined as Gleason score ≥ 3 + 4. The evaluation metric was the area under the curve (AUC).

RESULTS

IMPROD bpMRI Likert demonstrated the highest TLPOCV AUC of 0.92. The tested clinical variables had AUC 0.56-0.73, while the mRNA and additional IMPROD bpMRI parameters had AUC 0.50-0.67 and 0.65-0.89 respectively. The combination of clinical and mRNA biomarkers produced TLPOCV AUC of 0.87, the highest TLPOCV performance without including IMPROD bpMRI Likert.

DATA CONCLUSION

The qualitative IMPROD bpMRI Likert score demonstrated the highest accuracy for SPCa detection compared with the tested clinical variables and mRNA biomarkers.

LEVEL OF EVIDENCE

1 Technical Efficacy Stage: 2 J. Magn. Reson. Imaging 2020;51:1540-1553.

Prediction of biochemical recurrence in prostate cancer patients who underwent prostatectomy using routine clinical prostate multiparametric MRI and decipher genomic score

BACKGROUND

Biochemical recurrence (BCR) affects a significant proportion of patients who undergo robotic-assisted laparoscopic prostatectomy (RALP).

PURPOSE

To evaluate the performance of a routine clinical prostate multiparametric magnetic resonance imaging (mpMRI) and Decipher genomic classifier score for prediction of biochemical recurrence in patients who underwent RALP.

STUDY TYPE

Retrospective cohort study.

SUBJECTS

Ninety-one patients who underwent RALP performed by a single surgeon, had mpMRI before RALP, Decipher taken from RALP samples, and prostate specific antigen (PSA) follow-up for >3 years or BCR within 3 years, defined as PSA >0.2 mg/ml. FIELD STRENGTH/SEQUENCE: mpMRI was performed at 27 different institutions using 1.5T (n = 10) or 3T scanners and included T₂ w, diffusion-weighted imaging (DWI), or dynamic contrast-enhanced (DCE) MRI.

ASSESSMENT

All mpMRI studies were reported by one reader using Prostate Imaging Reporting and Data System v. 2.1 (PI-RADsv2.1) without knowledge of other findings. Eighteen (20%) randomly selected cases were re-reported by reader B to evaluate interreader variability.

STATISTICAL TESTS

Univariate and multivariate analysis using greedy feature selection and tournament leave-pair-out cross-validation (TLPOCV) were used to evaluate the performance of various variables for prediction of BCR, which included clinical (three), systematic biopsy (three), surgical (six: RALP Gleason Grade Group [GGG], extracapsular extension, seminal vesicle invasion, intraoperative surgical margins [PSM], final PSM, pTNM), Decipher (two: Decipher score, Decipher risk category), and mpMRI (eight: prostate volume, PSA density, PI-RADv2.1 score, MRI largest lesion size, summed MRI lesions' volume and relative volume [MRI-lesion-percentage], mpMRI ECE, mpMRI seminal vesicle invasion [SVI]) variables. The evaluation metric was the area under the curve (AUC).

RESULTS

Forty-eight (53%) patients developed BCR. The best-performing individual features with TLPOCV AUC of 0.73 (95% confidence interval [CI] 0.64-0.82) were RALP GGG, MRI-lesion-percentage followed by biopsy GGG (0.72, 0.62-0.82), and Decipher score (0.71, 0.60-0.82). The best performance was achieved by feature selection of Decipher+Surgery and MRI + Surgery variables with TLPOCV AUC of 0.82 and 0.81, respectively DATA CONCLUSION: Relative lesion volume measured on a routine clinical mpMRI failed to outperform Decipher score in BCR prediction.

LEVEL OF EVIDENCE

3 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2020;51:1075-1085.

Radiomics and machine learning of multisequence multiparametric prostate MRI: Towards improved non-invasive prostate cancer characterization

Purpose

To develop and validate a classifier system for prediction of prostate cancer (PCa) Gleason score (GS) using radiomics and texture features of T₂-weighted imaging (T₂w), diffusion weighted imaging (DWI) acquired using high b values, and T₂-mapping (T₂).

Methods

T₂w, DWI (12 b values, 0–2000 s/mm²), and T₂ data sets of 62 patients with histologically confirmed PCa were acquired at 3T using surface array coils. The DWI data sets were post-processed using monoexponential and kurtosis models, while T₂w was standardized to a common scale. Local statistics and 8 different radiomics/texture descriptors were utilized at different configurations to extract a total of 7105 unique per-tumor features. Regularized logistic regression with implicit feature selection and leave pair out cross validation was used to discriminate tumors with 3+3 vs >3+3 GS.

Results

In total, 100 PCa lesions were analysed, of those 20 and 80 had GS of 3+3 and >3+3, respectively. The best model performance was obtained by selecting the top 1% features of T₂w, ADC_m and K with ROC AUC of 0.88 (95% CI of 0.82–0.95). Features from T₂ mapping provided little added value. The most useful texture features were based on the gray-level co-occurrence matrix, Gabor transform, and Zernike moments.

Conclusion

Texture feature analysis of DWI, post-processed using monoexponential and kurtosis models, and T₂w demonstrated good classification performance for GS of PCa. In multisequence setting, the optimal radiomics based texture extraction methods and parameters differed between different image types.

Algebraic shortcuts for leave-one-out cross-validation in supervised network inference

Supervised machine learning techniques have traditionally been very successful at reconstructing biological networks, such as protein-ligand interaction, protein-protein interaction and gene regulatory networks. Many supervised techniques for network prediction use linear models on a possibly nonlinear pairwise feature representation of edges. Recently, much emphasis has been placed on the correct evaluation of such supervised models. It is vital to distinguish between using a model to either predict new interactions in a given network or to predict interactions for a new vertex not present in the original network. This distinction matters because (i) the performance might dramatically differ between the prediction settings and (ii) tuning the model hyperparameters to obtain the best possible model depends on the setting of interest. Specific cross-validation schemes need to be used to assess the performance in such different prediction settings.In this work we discuss a state-of-the-art kernel-based network inference technique called two-step kernel ridge regression. We show that this regression model can be trained efficiently, with a time complexity scaling with the number of vertices rather than the number of edges. Furthermore, this framework leads to a series of cross-validation shortcuts that allow one to rapidly estimate the model performance for any relevant network prediction setting. This allows computational biologists to fully assess the capabilities of their models. The machine learning techniques with the algebraic shortcuts are implemented in the RLScore software package: https://github.com/aatapa/RLScore.

Tournament leave-pair-out cross-validation for receiver operating characteristic analysis

Receiver operating characteristic analysis is widely used for evaluating diagnostic systems. Recent studies have shown that estimating an area under receiver operating characteristic curve with standard cross-validation methods suffers from a large bias. The leave-pair-out cross-validation has been shown to correct this bias. However, while leave-pair-out produces an almost unbiased estimate of area under receiver operating characteristic curve, it does not provide a ranking of the data needed for plotting and analyzing the receiver operating characteristic curve. In this study, we propose a new method called tournament leave-pair-out cross-validation. This method extends leave-pair-out by creating a tournament from pair comparisons to produce a ranking for the data. Tournament leave-pair-out preserves the advantage of leave-pair-out for estimating area under receiver operating characteristic curve, while it also allows performing receiver operating characteristic analyses. We have shown using both synthetic and real-world data that tournament leave-pair-out is as reliable as leave-pair-out for area under receiver operating characteristic curve estimation and confirmed the bias in leave-one-out cross-validation on low-dimensional data. As a case study on receiver operating characteristic analysis, we also evaluate how reliably sensitivity and specificity can be estimated from tournament leave-pair-out receiver operating characteristic curves.

Fast Kronecker Product Kernel Methods via Generalized Vec Trick

Kronecker product kernel provides the standard approach in the kernel methods' literature for learning from graph data, where edges are labeled and both start and end vertices have their own feature representations. The methods allow generalization to such new edges, whose start and end vertices do not appear in the training data, a setting known as zero-shot or zero-data learning. Such a setting occurs in numerous applications, including drug-target interaction prediction, collaborative filtering, and information retrieval. Efficient training algorithms based on the so-called vec trick that makes use of the special structure of the Kronecker product are known for the case where the training data are a complete bipartite graph. In this paper, we generalize these results to noncomplete training graphs. This allows us to derive a general framework for training Kronecker product kernel methods, as specific examples we implement Kronecker ridge regression and support vector machine algorithms. Experimental results demonstrate that the proposed approach leads to accurate models, while allowing order of magnitude improvements in training and prediction time.

Learning with multiple pairwise kernels for drug bioactivity prediction

Motivation

Many inference problems in bioinformatics, including drug bioactivity prediction, can be formulated as pairwise learning problems, in which one is interested in making predictions for pairs of objects, e.g. drugs and their targets. Kernel-based approaches have emerged as powerful tools for solving problems of that kind, and especially multiple kernel learning (MKL) offers promising benefits as it enables integrating various types of complex biomedical information sources in the form of kernels, along with learning their importance for the prediction task. However, the immense size of pairwise kernel spaces remains a major bottleneck, making the existing MKL algorithms computationally infeasible even for small number of input pairs.

Results

We introduce pairwiseMKL, the first method for time- and memory-efficient learning with multiple pairwise kernels. pairwiseMKL first determines the mixture weights of the input pairwise kernels, and then learns the pairwise prediction function. Both steps are performed efficiently without explicit computation of the massive pairwise matrices, therefore making the method applicable to solving large pairwise learning problems. We demonstrate the performance of pairwiseMKL in two related tasks of quantitative drug bioactivity prediction using up to 167 995 bioactivity measurements and 3120 pairwise kernels: (i) prediction of anticancer efficacy of drug compounds across a large panel of cancer cell lines; and (ii) prediction of target profiles of anticancer compounds across their kinome-wide target spaces. We show that pairwiseMKL provides accurate predictions using sparse solutions in terms of selected kernels, and therefore it automatically identifies also data sources relevant for the prediction problem.

Availability and implementation

Code is available at https://github.com/aalto-ics-kepaco.

Supplementary information

Supplementary data are available at Bioinformatics online.

A Comparative Study of Pairwise Learning Methods Based on Kernel Ridge Regression

Many machine learning problems can be formulated as predicting labels for a pair of objects. Problems of that kind are often referred to as pairwise learning, dyadic prediction, or network inference problems. During the past decade, kernel methods have played a dominant role in pairwise learning. They still obtain a state-of-the-art predictive performance, but a theoretical analysis of their behavior has been underexplored in the machine learning literature. In this work we review and unify kernel-based algorithms that are commonly used in different pairwise learning settings, ranging from matrix filtering to zero-shot learning. To this end, we focus on closed-form efficient instantiations of Kronecker kernel ridge regression. We show that independent task kernel ridge regression, two-step kernel ridge regression, and a linear matrix filter arise naturally as a special case of Kronecker kernel ridge regression, implying that all these methods implicitly minimize a squared loss. In addition, we analyze universality, consistency, and spectral filtering properties. Our theoretical results provide valuable insights into assessing the advantages and limitations of existing pairwise learning methods.

Time-Gated Raman Spectroscopy for Quantitative Determination of Solid-State Forms of Fluorescent Pharmaceuticals

Raman spectroscopy is widely used for quantitative pharmaceutical analysis, but a common obstacle to its use is sample fluorescence masking the Raman signal. Time-gating provides an instrument-based method for rejecting fluorescence through temporal resolution of the spectral signal and allows Raman spectra of fluorescent materials to be obtained. An additional practical advantage is that analysis is possible in ambient lighting. This study assesses the efficacy of time-gated Raman spectroscopy for the quantitative measurement of fluorescent pharmaceuticals. Time-gated Raman spectroscopy with a 128 × (2) × 4 CMOS SPAD detector was applied for quantitative analysis of ternary mixtures of solid-state forms of the model drug, piroxicam (PRX). Partial least-squares (PLS) regression allowed quantification, with Raman-active time domain selection (based on visual inspection) improving performance. Model performance was further improved by using kernel-based regularized least-squares (RLS) regression with greedy feature selection in which the data use in both the Raman shift and time dimensions was statistically optimized. Overall, time-gated Raman spectroscopy, especially with optimized data analysis in both the spectral and time dimensions, shows potential for sensitive and relatively routine quantitative analysis of photoluminescent pharmaceuticals during drug development and manufacturing.

Switching maintenance infliximab therapy to biosimilar infliximab in inflammatory bowel disease patients

BACKGROUND

Clinical use of biosimilar infliximab (CT-P13) in inflammatory bowel diseases (IBDs) is based on extrapolation of indication from clinical studies performed in rheumatological diseases. Only few data exist of behaviour of infliximab trough levels (TLs) and anti-drug antibodies (ADAs) during switching.

AIM

The objective of this study was to evaluate changes in TLs, ADA formation and disease activity after switching from originator infliximab to biosimilar one.

METHODS

All our IBD patients receiving maintenance infliximab therapy were switched to biosimilar infliximab. TLs and ADAs were measured before the last originator infusion and before the third biosimilar infusion. Laboratory values, disease activity indices (partial Mayo score and Harvey-Bradshaw index) and demographic data were collected from patient records.

RESULTS

A total of 62 patients were included in the final analysis (32 Crohn's disease, 30 ulcerative colitis (UC) or IBD-unclassified). No significant changes in median TLs before (5.5 mg/l) and after switching (5.5 mg/l, p = .05) occurred in the entire study group or in the Crohn's disease (CD) subgroup (5.75 and 6.5 mg/l, p = .68). However, in the subgroup of ulcerative colitis, the change in median TL was significantly different (from 5.2 to 4.25 mg/l, p = .019). Two patients developed ADAs after switching. No changes in disease activity were detected during switching and no safety concerns occurred.

CONCLUSIONS

Switching from originator to biosimilar infliximab resulted in statistically significant differences in infliximab TLs in patients with UC but not in patients with Crohn's disease. The clinical significance for this difference is doubtful and in neither group changes in disease activity occurred.

A Community Challenge for Inferring Genetic Predictors of Gene Essentialities through Analysis of a Functional Screen of Cancer Cell Lines

We report the results of a DREAM challenge designed to predict relative genetic essentialities based on a novel dataset testing 98,000 shRNAs against 149 molecularly characterized cancer cell lines. We analyzed the results of over 3,000 submissions over a period of 4 months. We found that algorithms combining essentiality data across multiple genes demonstrated increased accuracy; gene expression was the most informative molecular data type; the identity of the gene being predicted was far more important than the modeling strategy; well-predicted genes and selected molecular features showed enrichment in functional categories; and frequently selected expression features correlated with survival in primary tumors. This study establishes benchmarks for gene essentiality prediction, presents a community resource for future comparison with this benchmark, and provides insights into factors influencing the ability to predict gene essentiality from functional genetic screens. This study also demonstrates the value of releasing pre-publication data publicly to engage the community in an open research collaboration.

Assessment of metal ion concentration in water with structured feature selection

We propose a cost-effective system for the determination of metal ion concentration in water, addressing a central issue in water resources management. The system combines novel luminometric label array technology with a machine learning algorithm that selects a minimal number of array reagents (modulators) and liquid sample dilutions, such that enable accurate quantification. The algorithm is able to identify the optimal modulators and sample dilutions leading to cost reductions since less manual labour and resources are needed. Inferring the ion detector involves a unique type of a structured feature selection problem, which we formalize in this paper. We propose a novel Cartesian greedy forward feature selection algorithm for solving the problem. The novel algorithm was evaluated in the concentration assessment of five metal ions and the performance was compared to two known feature selection approaches. The results demonstrate that the proposed system can assist in lowering the costs with minimal loss in accuracy.

Circulating Cytokines Predict the Development of Insulin Resistance in a Prospective Finnish Population Cohort

CONTEXT

Metabolic inflammation contributes to the development of insulin resistance (IR), but the roles of different inflammatory and other cytokines in this process remain unclear.

OBJECTIVE

We aimed at analyzing the value of different cytokines in predicting future IR.

DESIGN, SETTING, AND PARTICIPANTS

We measured the serum concentrations of 48 cytokines from a nationwide cohort of 2200 Finns (the Cardiovascular Risk in Young Finns Study), and analyzed their role as independent risk factors for predicting the development of IR 4 years later.

MAIN OUTCOME MEASURES

We used cross-sectional regression analysis adjusted for known IR risk factors (high age, body mass index, systolic blood pressure, triglycerides, smoking, physical inactivity, and low high-density lipoprotein cholesterol), C-reactive protein and 37 cytokines to find the determinants of continuous baseline IR (defined by homeostatic model assessment). A logistic regression model adjusted for the known risk factors, baseline IR, and 37 cytokines was used to predict the future IR.

RESULTS

Several cytokines, often in a sex-dependent manner, remained as independent determinants of current IR. In men, none of the cytokines was an independent predictive risk marker of future IR. In women, in contrast, IL-17 (odds ratio, 1.42 for 1-SD change in ln-transformed IL-17) and IL-18 (odds ratio, 1.37) were independently associated with the future IR. IL-17 levels also independently predicted the development of incident future IR (odds ratio, 1.48).

CONCLUSIONS

The systemic levels of the T helper 1 cell cytokine IL-18 and the T helper 17 cell cytokine IL-17 thus may have value in predicting future insulin sensitivity in women independently of classical IR risk factors.

Fitting methods for intravoxel incoherent motion imaging of prostate cancer on region of interest level: Repeatability and gleason score prediction

PURPOSE

To evaluate different fitting methods for intravoxel incoherent motion (IVIM) imaging of prostate cancer in the terms of repeatability and Gleason score prediction.

METHODS

Eighty-one patients with histologically confirmed prostate cancer underwent two repeated 3 Tesla diffusion-weighted imaging (DWI) examinations performed using 14 b-values in the range of 0-500 s/mm² and diffusion time of 19.004 ms. Mean signal intensities of regions-of-interest were fitted using five different fitting methods for IVIM as well as monoexponential, kurtosis, and stretched exponential models. The fitting methods and models were evaluated in the terms of fitting quality [Akaike information criteria (AIC)], repeatability, and Gleason score prediction. Tumors were classified into three groups (3 + 3, 3 + 4, > 3 + 4). Machine learning algorithms were used to evaluate the performance of the combined use of the parameters. Simulation studies were performed to evaluate robustness of the fitting methods against noise.

RESULTS

Monoexponential model was preferred over IVIM based on AIC. The "pseudodiffusion" parameters demonstrated low repeatability and clinical value. Median "pseudodiffusion" fraction values were below 8.00%. Combined use of the parameters did not outperform the monoexponential model.

CONCLUSION

Monoexponential model demonstrated the highest repeatability and clinical values in the regions-of-interest based analysis of prostate cancer DWI, b-values in the range of 0-500 s/mm² . Magn Reson Med 77:1249-1264, 2017. © 2016 International Society for Magnetic Resonance in Medicine.

Diffusion-weighted imaging of prostate cancer: effect of b-value distribution on repeatability and cancer characterization

PURPOSE

To evaluate the effect of b-value distribution on the repeatability and Gleason score (GS) prediction of prostate cancer (PCa).

METHODS

Fifty PCa patients underwent two repeated 3T diffusion-weighted imaging (DWI) examinations using 12 b values in the range from 0 to 2000s/mm(2) and diffusion time of 20.3ms. Mean signal intensities of regions of interest, placed in PCa using whole mount prostatectomy sections as the reference, were fitted using monoexponential, kurtosis, stretched exponential, and biexponential models. In total, 4083 different b-value combinations consisting of 2 to 12 b values were evaluated. Repeatability was assessed by intraclass correlation coefficient, ICC(3,1), and coefficient of repeatability (CoR). Areas under receiver operating characteristic curve (AUCs) for PCa characterization were estimated while the correlation of the fitted values with GS groups (3+3, 3+4, >3+4) was evaluated by using the Spearman correlation coefficient (ρ).

RESULTS

The parameters of monoexponential, kurtosis, and stretched exponential models estimated using only 4-5, 5-7, 5-7 b values, respectively, had similar ICC(3,1), CoR, AUC, and ρ values as the parameters estimated using all 12 b values. Optimized b-value distributions demonstrated improved ICC(3,1) and CoR values but failed to improve AUC and ρ values. The parameters of biexponential model demonstrated the worst repeatability and diagnostic performance.

CONCLUSION

B-value distribution influences mainly the repeatability of DWI-derived parameters rather than the diagnostic performance.

High Performance Pattern Matching on Heterogeneous Platform

Pattern discovery is one of the fundamental tasks in bioinformatics and pattern recognition is a powerful technique for searching sequence patterns in the biological sequence databases. Fast and high performance algorithms are highly demanded in many applications in bioinformatics and computational molecular biology since the significant increase in the number of DNA and protein sequences expand the need for raising the performance of pattern matching algorithms. For this purpose, heterogeneous architectures can be a good choice due to their potential for high performance and energy efficiency. In this paper we present an efficient implementation of Aho-Corasick (AC) which is a well known exact pattern matching algorithm with linear complexity, and Parallel Failureless Aho-Corasick (PFAC) algorithm which is the massively parallelized version of AC algorithm without failure transitions, on a heterogeneous CPU/GPU architecture. We progressively redesigned the algorithms and data structures to fit on the GPU architecture. Our results on different protein sequence data sets show that the new implementation runs 15 times faster compared to the original implementation of the PFAC algorithm.

Identification of Functionally Related Enzymes by Learning-to-Rank Methods

Enzyme sequences and structures are routinely used in the biological sciences as queries to search for functionally related enzymes in online databases. To this end, one usually departs from some notion of similarity, comparing two enzymes by looking for correspondences in their sequences, structures or surfaces. For a given query, the search operation results in a ranking of the enzymes in the database, from very similar to dissimilar enzymes, while information about the biological function of annotated database enzymes is ignored. In this work, we show that rankings of that kind can be substantially improved by applying kernel-based learning algorithms. This approach enables the detection of statistical dependencies between similarities of the active cleft and the biological function of annotated enzymes. This is in contrast to search-based approaches, which do not take annotated training data into account. Similarity measures based on the active cleft are known to outperform sequence-based or structure-based measures under certain conditions. We consider the Enzyme Commission (EC) classification hierarchy for obtaining annotated enzymes during the training phase. The results of a set of sizeable experiments indicate a consistent and significant improvement for a set of similarity measures that exploit information about small cavities in the surface of enzymes.

Predicting patient acuity from electronic patient records

BACKGROUND

The ability to predict acuity (patients' care needs), would provide a powerful tool for health care managers to allocate resources. Such estimations and predictions for the care process can be produced from the vast amounts of healthcare data using information technology and computational intelligence techniques. Tactical decision-making and resource allocation may also be supported with different mathematical optimization models.

METHODS

This study was conducted with a data set comprising electronic nursing narratives and the associated Oulu Patient Classification (OPCq) acuity. A mathematical model for the automated assignment of patient acuity scores was utilized and evaluated with the pre-processed data from 23,528 electronic patient records. The methods to predict patient's acuity were based on linguistic pre-processing, vector-space text modeling, and regularized least-squares regression.

RESULTS

The experimental results show that it is possible to obtain accurate predictions about patient acuity scores for the coming day based on the assigned scores and nursing notes from the previous day. Making same-day predictions leads to even better results, as access to the nursing notes for the same day boosts the predictive performance. Furthermore, textual nursing notes allow for more accurate predictions than previous acuity scores. The best results are achieved by combining both of these information sources. The developed model achieves a concordance index of 0.821 when predicting the patient acuity scores for the following day, given the scores and text recorded on the previous day.

CONCLUSIONS

By applying language technology to electronic patient documents it is possible to accurately predict the value of the acuity scores of the coming day based on the previous daýs assigned scores and nursing notes.

Toward more realistic drug-target interaction predictions

A number of supervised machine learning models have recently been introduced for the prediction of drug-target interactions based on chemical structure and genomic sequence information. Although these models could offer improved means for many network pharmacology applications, such as repositioning of drugs for new therapeutic uses, the prediction models are often being constructed and evaluated under overly simplified settings that do not reflect the real-life problem in practical applications. Using quantitative drug-target bioactivity assays for kinase inhibitors, as well as a popular benchmarking data set of binary drug-target interactions for enzyme, ion channel, nuclear receptor and G protein-coupled receptor targets, we illustrate here the effects of four factors that may lead to dramatic differences in the prediction results: (i) problem formulation (standard binary classification or more realistic regression formulation), (ii) evaluation data set (drug and target families in the application use case), (iii) evaluation procedure (simple or nested cross-validation) and (iv) experimental setting (whether training and test sets share common drugs and targets, only drugs or targets or neither). Each of these factors should be taken into consideration to avoid reporting overoptimistic drug-target interaction prediction results. We also suggest guidelines on how to make the supervised drug-target interaction prediction studies more realistic in terms of such model formulations and evaluation setups that better address the inherent complexity of the prediction task in the practical applications, as well as novel benchmarking data sets that capture the continuous nature of the drug-target interactions for kinase inhibitors.

Genetic variants and their interactions in disease risk prediction - machine learning and network perspectives

A central challenge in systems biology and medical genetics is to understand how interactions among genetic loci contribute to complex phenotypic traits and human diseases. While most studies have so far relied on statistical modeling and association testing procedures, machine learning and predictive modeling approaches are increasingly being applied to mining genotype-phenotype relationships, also among those associations that do not necessarily meet statistical significance at the level of individual variants, yet still contributing to the combined predictive power at the level of variant panels. Network-based analysis of genetic variants and their interaction partners is another emerging trend by which to explore how sub-network level features contribute to complex disease processes and related phenotypes. In this review, we describe the basic concepts and algorithms behind machine learning-based genetic feature selection approaches, their potential benefits and limitations in genome-wide setting, and how physical or genetic interaction networks could be used as a priori information for providing improved predictive power and mechanistic insights into the disease networks. These developments are geared toward explaining a part of the missing heritability, and when combined with individual genomic profiling, such systems medicine approaches may also provide a principled means for tailoring personalized treatment strategies in the future.

Wrapper-based selection of genetic features in genome-wide association studies through fast matrix operations

BACKGROUND

Through the wealth of information contained within them, genome-wide association studies (GWAS) have the potential to provide researchers with a systematic means of associating genetic variants with a wide variety of disease phenotypes. Due to the limitations of approaches that have analyzed single variants one at a time, it has been proposed that the genetic basis of these disorders could be determined through detailed analysis of the genetic variants themselves and in conjunction with one another. The construction of models that account for these subsets of variants requires methodologies that generate predictions based on the total risk of a particular group of polymorphisms. However, due to the excessive number of variants, constructing these types of models has so far been computationally infeasible.

RESULTS

We have implemented an algorithm, known as greedy RLS, that we use to perform the first known wrapper-based feature selection on the genome-wide level. The running time of greedy RLS grows linearly in the number of training examples, the number of features in the original data set, and the number of selected features. This speed is achieved through computational short-cuts based on matrix calculus. Since the memory consumption in present-day computers can form an even tighter bottleneck than running time, we also developed a space efficient variation of greedy RLS which trades running time for memory. These approaches are then compared to traditional wrapper-based feature selection implementations based on support vector machines (SVM) to reveal the relative speed-up and to assess the feasibility of the new algorithm. As a proof of concept, we apply greedy RLS to the Hypertension - UK National Blood Service WTCCC dataset and select the most predictive variants using 3-fold external cross-validation in less than 26 minutes on a high-end desktop. On this dataset, we also show that greedy RLS has a better classification performance on independent test data than a classifier trained using features selected by a statistical p-value-based filter, which is currently the most popular approach for constructing predictive models in GWAS.

CONCLUSIONS

Greedy RLS is the first known implementation of a machine learning based method with the capability to conduct a wrapper-based feature selection on an entire GWAS containing several thousand examples and over 400,000 variants. In our experiments, greedy RLS selected a highly predictive subset of genetic variants in a fraction of the time spent by wrapper-based selection methods used together with SVM classifiers. The proposed algorithms are freely available as part of the RLScore software library at http://users.utu.fi/aatapa/RLScore/.