A primer on the use of machine learning to distil knowledge from data in biological psychiatry

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.

ABI1 ‐based expression signature predicts breast cancer metastasis and survival

Despite the current standard of care, breast cancer remains one of the leading causes of mortality in women worldwide, thus emphasizing the need for better predictive and therapeutic targets. ABI1 is associated with poor survival and an aggressive breast cancer phenotype, although its role in tumorigenesis, metastasis, and the disease outcome remains to be elucidated. Here, we define the ABI1‐based seven‐gene prognostic signature that predicts survival of metastatic breast cancer patients; ABI1 is an essential component of the signature. Genetic disruption of Abi1 in primary breast cancer tumors of PyMT mice led to significant reduction of the number and size of lung metastases in a gene dose‐dependent manner. The disruption of Abi1 resulted in deregulation of the WAVE complex at the mRNA and protein levels in mouse tumors. In conclusion, ABI1 is a prognostic metastatic biomarker in breast cancer. We demonstrate, for the first time, that lung metastasis is associated with an Abi1 gene dose and specific gene expression aberrations in primary breast cancer tumors. These results indicate that targeting ABI1 may provide a therapeutic advantage in breast cancer patients.

Integrative Survival Analysis of Breast Cancer with Gene Expression and DNA Methylation Data

Motivation

Integrative multi-feature fusion analysis on biomedical data has gained much attention recently. In breast cancer, existing studies have demonstrated that combining genomic mRNA data and DNA methylation data can better stratify cancer patients with distinct prognosis than using single signature. However, those existing methods are simply combining these gene features in series and have ignored the correlations between separate omics dimensions over time.

Results

In the present study, we propose an adaptive multi-task learning method, which combines the Cox loss task with the ordinal loss task, for survival prediction of breast cancer patients using multi-modal learning instead of performing survival analysis on each feature dataset. First, we use local maximum quasi-clique merging (lmQCM) algorithm to reduce the mRNA and methylation feature dimensions and extract cluster eigengenes respectively. Then, we add an auxiliary ordinal loss to the original Cox model to improve the ability to optimize the learning process in training and regularization. The auxiliary loss helps to reduce the vanishing gradient problem for earlier layers and helps to decrease the loss of the primary task. Meanwhile, we use an adaptive weights approach to multi-task learning which weighs multiple loss functions by considering the homoscedastic uncertainty of each task. Finally, we build an ordinal cox hazards model for survival analysis and use long short-term memory (LSTM) method to predict patients’ survival risk. We use the cross-validation method and the concordance index (C-index) for assessing the prediction effect. Stringent cross-verification testing processes for the benchmark dataset and two additional datasets demonstrate that the developed approach is effective, achieving very competitive performance with existing approaches.

Availability and implementation

https://github.com/bhioswego/ML_ordCOX.

Bioimage-Based Prediction of Protein Subcellular Location in Human Tissue with Ensemble Features and Deep Networks

Prediction of protein subcellular location has currently become a hot topic because it has been proven to be useful for understanding both the disease mechanisms and novel drug design. With the rapid development of automated microscopic imaging technology in recent years, classification methods of bioimage-based protein subcellular location have attracted considerable attention for images can describe the protein distribution intuitively and in detail. In the current study, a prediction method of protein subcellular location was proposed based on multi-view image features that are extracted from three different views, including the four texture features of the original image, the global and local features of the protein extracted from the protein channel images after color segmentation, and the global features of DNA extracted from the DNA channel image. Finally, the extracted features were combined together to improve the performance of subcellular localization prediction. From the performance comparison of different combination features under the same classifier, the best ensemble features could be obtained. In this work, a classifier based on Stacked Auto-encoders and the random forest was also put forward. To improve the prediction results, the deep network was combined with the traditional statistical classification methods. Stringent cross-validation and independent validation tests on the benchmark dataset demonstrated the efficacy of the proposed method.

Abstract 3404: Loss of Abi1 abrogates lung metastasis in the PyMT mouse model of breast cancer

This study aims to elucidate the role of Abelson interactor 1 (Abi1), a key protein in the WAVE regulatory complex, in mammary carcinogenesis and metastasis. Breast cancer is the second-leading cause of mortality in women in the United States with an estimated ~200,000 new cases and over 40,000 deaths this year. Despite current treatment modalities, breast cancer patients often relapse after only a few years of treatment thus emphasizing the need for better therapeutic targets. Abi1 is an adaptor protein mainly associated with the WAVE (Wiscott-Aldrich syndrome protein family verprolin homologous) regulatory complex and Arp2/3 (Actin-related proteins 2 and 3)-mediated actin cytoskeleton remodeling. Our bioinformatic and gene expression analyses of human tumor data indicates that Abi1 is frequently upregulated in invasive breast cancers, is associated with poor survival, and may promote an aggressive breast tumor phenotype. Downregulation of Abi1 also abrogates motility and invasion of breast cancer cells, most likely through inactivation of both Src and PI3 kinase as well as certain matrix metalloproteases. We therefore hypothesize that Abi1 positively regulates breast tumor progression and invasion through dysregulation of these cell signaling pathways. To determine the role of Abi1 in breast tumor progression, we used a Cre-lox system to conditionally delete Abi1 in the mammary tissue of Polyoma Middle T (PyMT) breast cancer mice. Abi1 knockout (KO) and control mice were palpated bi-weekly to determine tumor latency and tumors were measured with a caliper to determine total tumor burden. We are currently analyzing changes in tumor kinetics as a result of conditional Abi1 knockout in the mammary epithelium of PyMT mice. Our preliminary studies of indicate slowed tumor growth in Abi1 KO PyMT mice. Western blot analyses of Abi1 KO mammary tumors indicate concomitant loss of WAVE complex proteins supporting our previous findings that WAVE complex integrity is dependent on Abi1. Most interestingly, Abi1 null PyMT mice exhibit significantly reduced incidence of lung metastasis, supporting our hypothesis that Abi1 promotes invasion by breast cancer cells. In summary, Abi1 loss leads to reduction of lung metastasis in PyMT mice, possibly through inactivation of key cell signaling and proliferation pathways such as Src and PI3 kinase. This work will establish Abi1 as a potential prognostic marker and therapeutic target in metastatic breast cancer.

Citation Format: Angelina Regua, Isabelle Bichindaritz, Tiffany Caza, Julie White, Robert Adamiecki, Mira Krendel, Gennady Bratslavsky, Leszek Kotula. Loss of Abi1 abrogates lung metastasis in the PyMT mouse model of breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 3404. doi:10.1158/1538-7445.AM2017-3404

Case Mining for Research Findings in a Case-based Reasoning System in the Biology of Aging

Scientific literature has been quickly expanding as the availability of articles in electronic form has increased rapidly. For the scientific researcher and the practitioner alike, keeping track with the advancement of the research is an ongoing challenge, and for the most part, the mass of experience recorded in the scientific literature is largely untapped. In particular, novice scientists, non researchers, and students would benefit from a system proposing recommendations for the problems they are interested in resolving. This article presents the first stages of the Digital Knowledge Finder design, a case-based reasoning system to manage experience from the scientific literature. One of the main functionality of the system is to enable both to represent the experience in a declarative and searchable form, and to reason from it through reuse - the latter being a consequence of the former. This article focuses on research findings mining and results from an aging literature dataset.

Report on the Eighteenth International Conference on Case-Based Reasoning

This article reports on the main track papers, speakers, satellite events, and other activities of the Eighteenth International Conference on Case-Based Reasoning (ICCBR), held 19-22 July 2010 in Alessandria, Italy

Case-based reasoning in the health sciences: What's next?

OBJECTIVES

This paper presents current work in case-based reasoning (CBR) in the health sciences, describes current trends and issues, and projects future directions for work in this field.

METHODS AND MATERIAL

It represents the contributions of researchers at two workshops on case-based reasoning in the health sciences. These workshops were held at the Fifth International Conference on Case-Based Reasoning (ICCBR-03) and the Seventh European Conference on Case-Based Reasoning (ECCBR-04).

RESULTS

Current research in CBR in the health sciences is marked by its richness. Highlighted trends include work in bioinformatics, support to the elderly and people with disabilities, formalization of CBR in biomedicine, and feature and case mining.

CONCLUSION

CBR systems are being better designed to account for the complexity of biomedicine, to integrate into clinical settings and to communicate and interact with diverse systems and methods.

Mémoire: A framework for semantic interoperability of case-based reasoning systems in biology and medicine

OBJECTIVE

Mémoire is a framework for sharing and distributing case bases and case-based reasoning (CBR) systems in biology and medicine.

METHODS AND MATERIAL

This paper first introduces the semantic Web approach to build a better Web where search engines, knowledge sources and servers, applications and services can live, work, and learn in cooperation. This semantic approach is particularly well suited for biomedical domains because significant ontologies have been developed there and constitute a sound basis for the standardization effort required for the semantic Web. Case-based reasoning systems in biomedicine have also benefited from these biomedical ontologies and models.

RESULTS

This article demonstrates for three such systems how a semantics infused approach in CBR gives better, more accurate results in CBR. From this previous work on a semantic approach in CBR in biomedicine, the Mémoire framework has evolved.

CONCLUSION

Mémoire proposes a unified OWL-based representation language for cases and case-based ontologies in biomedicine, where a Web Ontology Language (OWL) is a language to represent ontologies on the Web. Mémoire provides a set of tools for building case-based reasoning systems compliant with its language. Mémoire is extensible and can be adapted to different types of biomedical application domains, tasks, and environments. Mémoire will permit bridging the gap within the multiple case-based reasoning systems dedicated to a single domain, and make available to agents and Web services the case-based competency of the CBR systems adopting its interchange language. The approach could be extended to other application domains of CBR.

Physician participation in research surveys. A randomized study of inducements to return mailed research questionnaires

The authors randomly selected 400 physicians from a population of 1,545 practicing physicians providing follow-up care to patients who received bone marrow or blood stem cell transplants at the Fred Hutchinson Cancer Research Center to determine interest in receiving Internet-based transplant information. In a two-factor completely randomized factorial design, the 400 physicians were assigned to receive mailed surveys with either no compensation or a $5 check and either no follow-up call or a follow-up call 3 weeks after mailing. Overall, 51.5% of the physicians returned the mailed surveys. Comparison of logit models showed that inclusion of a $5 check in the mailer significantly (p = .016) increased the probability of returning the surveys (57.5% vs. 45.5%). In contrast, the telephone follow-up had no overall effect. The authors concluded a modest financial reward can significantly improve physician response rates to research surveys but a telephone follow-up may be inefficient and even ineffective.

CARE-PARTNER: a computerized knowledge-support system for stem-cell post-transplant long-term follow-up on the World-Wide-Web

Evidence-based practice in medicine promotes the performance of medicine based upon proven and validated practice. The CARE-PARTNER system presented here is a computerized knowledge-support system for stem-cell post-transplant long-term follow-up (LTFU) care on the WWW, which means that it monitors the quality of the knowledge both of its own knowledge-base and of its users. Its aim is to support the evidence-based practice of the LTFU clinicians and of the home-town physicians who actually care for the transplanted patients. Currently, three fundamental characteristics of CARE-PARTNER are accountable for its knowledge-support function: the quality of its knowledge-base, its availability on the WWW, and its learning from experience capability. As a matter of fact, the integration of a case-based reasoner in the reasoning framework enables the system to introspectively study its results, and to learn from its successes and failures, thus confronting the quality of the guidelines and pathways it reuses to the reality and complexity of the clinical cases.

Selective processing of eating and body words in restricting-type anorexics, binge-eating-type anorexics, bulimics, and control subjects

A French adaptation of the Stroop colour-naming task was used to investigate selective processing of information related to eating and the body in 92 female subjects: 18 with restricting-type anorexia nervosa (RAs), 25 with binge-eating-type anorexia (BAs), 20 with bulimia nervosa (BNs), and 29 controls (Cs). All participants were significantly slower in colour-naming words related to eating and the body. This suggests that eating and body Stroop effects were not diagnostic category effects. Eating Stroop effect was significantly stronger in eating disordered patients than in Cs, but very similar in RAs, BAs, and BNs. It did not correlate with the body mass index nor with Eating Disorder Inventory (EDI) scores. Body Stroop effect did not differ significantly between groups. However some of our results suggested a tendency for this effect to be higher in BNs, corroborating previous literature data. Body Stroop effect correlated with EDI score in RAs, and with the "body dissatisfaction" subscale score in BNs and BAs. It did not correlate with the body mass index. Further studies are currently being carried out to determine whether inpatient treatment for anorexia nervosa induces significant changes in eating and/or body Stroop effect, and whether such changes might be predictive of clinical outcome.

[Measuring the development of eating disorders in a group of patients with the Eating Disorders Inventory (EDI)]

The Eating Disorders Inventory (EDI I) has been applied to four groups of patients with Eating Disorders (ED): Anorexic Bulimic Subtype (AB N = 22), Anorexic Restrictive Subtype (AR N = 36); Bulimic (B N = 51) and Other Eating Disorders (OED N = 15), according to the DSM III-R criteria. We have also applied the Eating Attitude Test (EAT-40), the Beck Depression Inventory (13 items), the Hamilton Anxiety Rating Scale and the Hamilton Depression Rating Scale. During their hospital care, the patients have been treated by cognitive-behavioural therapy. The order or decreasing total scores in EDI at the admission is the B subtype, then the AR subtype, then the OED group. In total score, the B and AB improved more than the order groups, between their admission and a second assessment three months later. Body dissatisfaction, Drive for thinness and Interpersonnal distrust are the only subscales for which no group showed a statistical improvement after three months. The B group decreased more in the subscale, than the other groups, particularly in Interoceptive awareness, Ineffectiveness and bulimia.

A case-based assistant for clinical psychiatry expertise

Case-based reasoning is an artificial intelligence methodology for the processing of empirical knowledge. Recent case-based reasoning systems also use theoretic knowledge about the domain to constrain the case-based reasoning. The organization of the memory is the key issue in case-based reasoning. The case-based assistant presented here has two structures in memory: cases and concepts. These memory structures permit it to be as skilled in problem-solving tasks, such as diagnosis and treatment planning, as in interpretive tasks, such as clinical research. A prototype applied to clinical work about eating disorders in psychiatry, reasoning from the alimentary questionnaires of these patients, is presented as an example of the system abilities.