ABSTRACT
Applications of machine learning in the biomedical sciences are growing rapidly. This growth has been spurred by diverse cross-institutional and interdisciplinary collaborations, public availability of large datasets, an increase in the accessibility of analytic routines, and the availability of powerful computing resources. With this increased access and exposure to machine learning comes a responsibility for education and a deeper understanding of its bases and bounds, borne equally by data scientists seeking to ply their analytic wares in medical research and by biomedical scientists seeking to harness such methods to glean knowledge from data. This article provides an accessible and critical review of machine learning for a biomedically informed audience, as well as its applications in psychiatry. The review covers definitions and expositions of commonly used machine learning methods, and historical trends of their use in psychiatry. We also provide a set of standards, namely Guidelines for REporting Machine Learning Investigations in Neuropsychiatry (GREMLIN), for designing and reporting studies that use machine learning as a primary data-analysis approach. Lastly, we propose the establishment of the Machine Learning in Psychiatry (MLPsych) Consortium, enumerate its objectives, and identify areas of opportunity for future applications of machine learning in biological psychiatry. This review serves as a cautiously optimistic primer on machine learning for those on the precipice as they prepare to dive into the field, either as methodological practitioners or well-informed consumers.
Subject(s)
Biological Psychiatry , Machine Learning , Humans , Biological Psychiatry/methods , Psychiatry/methods , Biomedical Research/methodsABSTRACT
This study analyzed gene expression messenger RNA data, from cases with major depressive disorder (MDD) and controls, using supervised machine learning (ML). We built on the methodology of prior studies to obtain more generalizable/reproducible results. First, we obtained a classifier trained on gene expression data from the dorsolateral prefrontal cortex of post-mortem MDD cases (n = 126) and controls (n = 103). An average area-under-the-receiver-operating-characteristics-curve (AUC) from 10-fold cross-validation of 0.72 was noted, compared to an average AUC of 0.55 for a baseline classifier (p = .0048). The classifier achieved an AUC of 0.76 on a previously unused testing-set. We also performed external validation using DLPFC gene expression values from an independent cohort of matched MDD cases (n = 29) and controls (n = 29), obtained from Affymetrix microarray (vs. Illumina microarray for the original cohort) (AUC: 0.62). We highlighted gene sets differentially expressed in MDD that were enriched for genes identified by the ML algorithm. Next, we assessed the ML classification performance in blood-based microarray gene expression data from MDD cases (n = 1,581) and controls (n = 369). We observed a mean AUC of 0.64 on 10-fold cross-validation, which was significantly above baseline (p = .0020). Similar performance was observed on the testing-set (AUC: 0.61). Finally, we analyzed the classification performance in covariates subgroups. We identified an interesting interaction between smoking and recall performance in MDD case prediction (58% accurate predictions in cases who are smokers vs. 43% accurate predictions in cases who are non-smokers). Overall, our results suggest that ML in combination with gene expression data and covariates could further our understanding of the pathophysiology in MDD.
Subject(s)
Biomarkers/analysis , Brain/metabolism , Computational Biology/methods , Depressive Disorder, Major/genetics , Machine Learning , RNA, Messenger/genetics , Adolescent , Adult , Aged , Aged, 80 and over , Case-Control Studies , Cohort Studies , Depressive Disorder, Major/blood , Depressive Disorder, Major/pathology , Female , Humans , Male , Middle Aged , ROC Curve , Young AdultABSTRACT
BACKGROUND: There is a lack of reliable biomarkers for major depressive disorder (MDD) in clinical practice. However, several studies have shown an association between alterations in microRNA levels and MDD, albeit none of them has taken advantage of machine learning (ML). METHOD: Supervised and unsupervised ML were applied to blood microRNA expression profiles from a MDD case-control dataset (n = 168) to distinguish between (1) case vs control status, (2) MDD severity levels defined based on the Montgomery-Asberg Depression Rating Scale, and (3) antidepressant responders vs nonresponders. RESULTS: MDD cases were distinguishable from healthy controls with an area-under-the receiver-operating characteristic curve (AUC) of 0.97 on testing data. High- vs low-severity cases were distinguishable with an AUC of 0.63. Unsupervised clustering of patients, before supervised ML analysis of each cluster for MDD severity, improved the performance of the classifiers (AUC of 0.70 for cluster 1 and 0.76 for cluster 2). Antidepressant responders could not be successfully separated from nonresponders, even after patient stratification by unsupervised clustering. However, permutation testing of the top microRNA, identified by the ML model trained to distinguish responders vs nonresponders in each of the 2 clusters, showed an association with antidepressant response. Each of these microRNA markers was only significant when comparing responders vs nonresponders of the corresponding cluster, but not using the heterogeneous unclustered patient set. CONCLUSIONS: Supervised and unsupervised ML analysis of microRNA may lead to robust biomarkers for monitoring clinical evolution and for more timely assessment of treatment in MDD patients.
Subject(s)
Circulating MicroRNA/blood , Depressive Disorder, Major/blood , RNA-Seq , Supervised Machine Learning , Unsupervised Machine Learning , Affect/drug effects , Antidepressive Agents/therapeutic use , Biomarkers/blood , Case-Control Studies , Depressive Disorder, Major/diagnosis , Depressive Disorder, Major/drug therapy , Depressive Disorder, Major/psychology , Humans , Predictive Value of Tests , Severity of Illness Index , Treatment OutcomeABSTRACT
BACKGROUND: Machine learning (ML) algorithms and methods offer great tools to analyze large complex genomic datasets. Our goal was to compare the genomic architecture of schizophrenia (SCZ) and autism spectrum disorder (ASD) using ML. METHODS: In this paper, we used regularized gradient boosted machines to analyze whole-exome sequencing (WES) data from individuals SCZ and ASD in order to identify important distinguishing genetic features. We further demonstrated a method of gene clustering to highlight which subsets of genes identified by the ML algorithm are mutated concurrently in affected individuals and are central to each disease (i.e., ASD vs. SCZ "hub" genes). RESULTS: In summary, after correcting for population structure, we found that SCZ and ASD cases could be successfully separated based on genetic information, with 86-88% accuracy on the testing dataset. Through bioinformatic analysis, we explored if combinations of genes concurrently mutated in patients with the same condition ("hub" genes) belong to specific pathways. Several themes were found to be associated with ASD, including calcium ion transmembrane transport, immune system/inflammation, synapse organization, and retinoid metabolic process. Moreover, ion transmembrane transport, neurotransmitter transport, and microtubule/cytoskeleton processes were highlighted for SCZ. CONCLUSIONS: Our manuscript introduces a novel comparative approach for studying the genetic architecture of genetically related diseases with complex inheritance and highlights genetic similarities and differences between ASD and SCZ.
Subject(s)
Autism Spectrum Disorder , Autistic Disorder , Schizophrenia , Autism Spectrum Disorder/genetics , Autistic Disorder/genetics , Exome/genetics , Genomics , Humans , Machine Learning , Schizophrenia/genetics , Exome SequencingABSTRACT
Our hypothesis is that machine learning (ML) analysis of whole exome sequencing (WES) data can be used to identify individuals at high risk for schizophrenia (SCZ). This study applies ML to WES data from 2,545 individuals with SCZ and 2,545 unaffected individuals, accessed via the database of genotypes and phenotypes (dbGaP). Single nucleotide variants and small insertions and deletions were annotated by ANNOVAR using the reference genome hg19/GRCh37. Rare (predicted functional) variants with a minor allele frequency ≤1% and genotype quality ≥90 including missense, frameshift, stop gain, stop loss, intronic, and exonic splicing variants were selected. A file containing all cases and controls, the names of genes with variants meeting our criteria, and the number of variants per gene for each individual, was used for ML analysis. The supervised machine-learning algorithm used the patterns of variants observed in the different genes to determine which subset of genes can best predict that an individual is affected. Seventy percent of the data was used to train the algorithm and the remaining 30% of data (n = 1,526) was used to evaluate its efficiency. The supervised ML algorithm, gradient boosted trees with regularization (eXtreme Gradient Boosting implementation) was the best performing algorithm yielding promising results (accuracy: 85.7%, specificity: 86.6%, sensitivity: 84.9%, area under the receiver-operator characteristic curve: 0.95). The top 50 features (genes) of the algorithm were analyzed using bioinformatics resources for new insights about the pathophysiology of SCZ. This manuscript presents a novel predictor which could potentially enable studies exploring disease-modifying intervention in the early stages of the disease.
Subject(s)
Computational Biology/methods , Schizophrenia/genetics , Sequence Analysis, DNA/methods , Algorithms , Alleles , Case-Control Studies , Exome/genetics , Gene Frequency/genetics , Genomics , Genotype , Humans , INDEL Mutation/genetics , Machine Learning , Polymorphism, Single Nucleotide/genetics , ROC Curve , Schizophrenia/etiology , Schizophrenic Psychology , Sensitivity and Specificity , Whole Genome Sequencing/methodsABSTRACT
A literature review was conducted, using the computerized "Online Mendelian Inheritance in Man" (OMIM) and PubMed, to identify inborn errors of metabolism (IEM) in which psychosis may be a predominant feature or the initial presenting symptom. Different combinations of the following keywords were searched using OMIM: "psychosis", "schizophrenia", or "hallucinations" and "metabolic", "inborn error of metabolism", "inborn errors of metabolism", "biochemical genetics", or "metabolic genetics". The OMIM search generated 126 OMIM entries, 40 of which were well known IEM. After removing IEM lacking evidence in PubMed for an association with psychosis, 29 OMIM entries were identified. Several of these IEM are treatable. They involve different small organelles (lysosomes, peroxisomes, mitochondria), iron or copper accumulation, as well as defects in other met-abolic pathways (e.g., defects leading to hyperammonemia or homocystinemia). A clinical checklist summarizing the key features of these conditions and a guide to clinical approach are provided. The genes corresponding to each of these con-ditions were identified. Whole exome data from 2545 adult cases with schizophrenia and 2545 unrelated controls, accessed via the Database of Genotypes and Phenotypes (dbGaP), were analyzed for rare functional variants in these genes. The odds ratio of having a rare functional variant in cases versus controls was calculated for each gene. Eight genes are significantly associated with schizophrenia (p < 0.05, OR >1) using an unselected group of adult patients with schizophrenia. Increased awareness of clinical clues for these IEM will optimize referrals and timely metabolic interventions.
Subject(s)
Exome , Metabolism, Inborn Errors/genetics , Schizophrenia/genetics , Adult , Case-Control Studies , Humans , Metabolism, Inborn Errors/therapyABSTRACT
BACKGROUND AND OBJECTIVE: Alterations of the expression of a variety of genes have been reported in patients with schizophrenia (SCZ). Moreover, machine learning (ML) analysis of gene expression microarray data has shown promising preliminary results in the study of SCZ. Our objective was to evaluate the performance of ML in classifying SCZ cases and controls based on gene expression microarray data from the dorsolateral prefrontal cortex. METHODS: We apply a state-of-the-art ML algorithm (XGBoost) to train and evaluate a classification model using 201 SCZ cases and 278 controls. We utilized 10-fold cross-validation for model selection, and a held-out testing set to evaluate the model. The performance metric utilizes to evaluate classification performance was the area under the receiver-operator characteristics curve (AUC). RESULTS: We report an average AUC on 10-fold cross-validation of 0.76 and an AUC of 0.76 on testing data, not used during training. Analysis of the rolling balanced classification accuracy from high to low prediction confidence levels showed that the most certain subset of predictions ranged between 80-90%. The ML model utilized 182 gene expression probes. Further improvement to classification performance was observed when applying an automated ML strategy on the 182 features, which achieved an AUC of 0.79 on the same testing data. We found literature evidence linking all of the top ten ML ranked genes to SCZ. Furthermore, we leveraged information from the full set of microarray gene expressions available via univariate differential gene expression analysis. We then prioritized differentially expressed gene sets using the piano gene set analysis package. We augmented the ranking of the prioritized gene sets with genes from the complex multivariate ML model using hypergeometric tests to identify more robust gene sets. We identified two significant Gene Ontology molecular function gene sets: "oxidoreductase activity, acting on the CH-NH2 group of donors" and "integrin binding." Lastly, we present candidate treatments for SCZ based on findings from our study CONCLUSIONS: Overall, we observed above-chance performance from ML classification of SCZ cases and controls based on brain gene expression microarray data, and found that ML analysis of gene expressions could further our understanding of the pathophysiology of SCZ and help identify novel treatments.
Subject(s)
Schizophrenia , Brain , Case-Control Studies , Dorsolateral Prefrontal Cortex , Humans , Machine Learning , Schizophrenia/genetics , TranscriptomeABSTRACT
Phenylalanine hydroxylase deficiency is an autosomal recessive disorder that results in intolerance to the dietary intake of the essential amino acid phenylalanine. It occurs in approximately 1:15,000 individuals. Deficiency of this enzyme produces a spectrum of disorders including classic phenylketonuria, mild phenylketonuria, and mild hyperphenylalaninemia. Classic phenylketonuria is caused by a complete or near-complete deficiency of phenylalanine hydroxylase activity and without dietary restriction of phenylalanine most children will develop profound and irreversible intellectual disability. Mild phenylketonuria and mild hyperphenylalaninemia are associated with lower risk of impaired cognitive development in the absence of treatment. Phenylalanine hydroxylase deficiency can be diagnosed by newborn screening based on detection of the presence of hyperphenylalaninemia using the Guthrie microbial inhibition assay or other assays on a blood spot obtained from a heel prick. Since the introduction of newborn screening, the major neurologic consequences of hyperphenylalaninemia have been largely eradicated. Affected individuals can lead normal lives. However, recent data suggest that homeostasis is not fully restored with current therapy. Treated individuals have a higher incidence of neuropsychological problems. The mainstay of treatment for hyperphenylalaninemia involves a low-protein diet and use of a phenylalanine-free medical formula. This treatment must commence as soon as possible after birth and should continue for life. Regular monitoring of plasma phenylalanine and tyrosine concentrations is necessary. Targets of plasma phenylalanine of 120-360 µmol/L (2-6 mg/dL) in the first decade of life are essential for optimal outcome. Phenylalanine targets in adolescence and adulthood are less clear. A significant proportion of patients with phenylketonuria may benefit from adjuvant therapy with 6R-tetrahydrobiopterin stereoisomer. Special consideration must be given to adult women with hyperphenylalaninemia because of the teratogenic effects of phenylalanine. Women with phenylalanine hydroxylase deficiency considering pregnancy should follow special guidelines and assure adequate energy intake with the proper proportion of protein, fat, and carbohydrates to minimize risks to the developing fetus. Molecular genetic testing of the phenylalanine hydroxylase gene is available for genetic counseling purposes to determine carrier status of at-risk relatives and for prenatal testing.
Subject(s)
Phenylalanine Hydroxylase/deficiency , Animals , Clinical Trials as Topic , DNA Mutational Analysis , Diagnosis, Differential , Genetic Association Studies , Genetic Testing , Humans , Mutation , Phenylalanine/blood , Phenylalanine Hydroxylase/genetics , Phenylketonurias/diagnosis , Phenylketonurias/genetics , Phenylketonurias/therapyABSTRACT
BACKGROUND: We propose a phenotype-driven analysis of encrypted exome data to facilitate the widespread implementation of exome sequencing as a clinical genetic screening test.Twenty test-patients with varied syndromes were selected from the literature. For each patient, the mutation, phenotypic data, and genetic diagnosis were available. Next, control exome-files, each modified to include one of these twenty mutations, were assigned to the corresponding test-patients. These data were used by a geneticist blinded to the diagnoses to test the efficiency of our software, PhenoVar. The score assigned by PhenoVar to any genetic diagnosis listed in OMIM (Online Mendelian Inheritance in Man) took into consideration both the patient's phenotype and all variations present in the corresponding exome. The physician did not have access to the individual mutations. PhenoVar filtered the search using a cut-off phenotypic match threshold to prevent undesired discovery of incidental findings and ranked the OMIM entries according to diagnostic score. RESULTS: When assigning the same weight to all variants in the exome, PhenoVar predicted the correct diagnosis in 10/20 patients, while in 15/20 the correct diagnosis was among the 4 highest ranked diagnoses. When assigning a higher weight to variants known, or bioinformatically predicted, to cause disease, PhenoVar's yield increased to 14/20 (18/20 in top 4). No incidental findings were identified using our cut-off phenotypic threshold. CONCLUSION: The phenotype-driven approach described could render widespread use of ES more practical, ethical and clinically useful. The implications about novel disease identification, advancement of complex diseases and personalized medicine are discussed.
Subject(s)
Congenital Abnormalities/diagnosis , Congenital Abnormalities/genetics , Genomics/methods , Software , Genome, Human/genetics , Humans , Phenotype , Reproducibility of Results , SyndromeABSTRACT
BACKGROUND: The revolution in DNA sequencing technologies over the past decade has made it feasible to sequence an individual's whole genome at a relatively low cost. The potential value of the information generated by genomic technologies for medicine and society is enormous. However, in order for exome sequencing, and eventually whole genome sequencing, to be implemented clinically, a number of major challenges need to be overcome. For instance, obtaining meaningful informed-consent, managing incidental findings and the great volume of data generated (including multiple findings with uncertain clinical significance), re-interpreting the genomic data and providing additional counselling to patients as genetic knowledge evolves are issues that need to be addressed. It appears that medical genetics is shifting from the present "phenotype-first" medical model to a "data-first" model which leads to multiple complexities. DISCUSSION: This manuscript discusses the different challenges associated with integrating genomic technologies into clinical practice and describes a "phenotype-first" approach, namely, "Individualized Mutation-weighed Phenotype Search", and its benefits. The proposed approach allows for a more efficient prioritization of the genes to be tested in a clinical lab based on both the patient's phenotype and his/her entire genomic data. It simplifies "informed-consent" for clinical use of genomic technologies and helps to protect the patient's autonomy and privacy. Overall, this approach could potentially render widespread use of genomic technologies, in the immediate future, practical, ethical and clinically useful. SUMMARY: The "Individualized Mutation-weighed Phenotype Search" approach allows for an incremental integration of genomic technologies into clinical practice. It ensures that we do not over-medicalize genomic data but, rather, continue our current medical model which is based on serving the patient's concerns. Service should not be solely driven by technology but rather by the medical needs and the extent to which a technology can be safely and effectively utilized.