Human-centered explainability for life sciences, healthcare, and medical informatics.

Rapid advances in artificial intelligence (AI) and availability of biological, medical, and healthcare data have enabled the development of a wide variety of models. Significant success has been achieved in a wide range of fields, such as genomics, protein folding, disease diagnosis, imaging, and clinical tasks. Although widely used, the inherent opacity of deep AI models has brought criticism from the research field and little adoption in clinical practice. Concurrently, there has been a significant amount of research focused on making such methods more interpretable, reviewed here, but inherent critiques of such explainability in AI (XAI), its requirements, and concerns with fairness/robustness have hampered their real-world adoption. We here discuss how user-driven XAI can be made more useful for different healthcare stakeholders through the definition of three key personas-data scientists, clinical researchers, and clinicians-and present an overview of how different XAI approaches can address their needs. For illustration, we also walk through several research and clinical examples that take advantage of XAI open-source tools, including those that help enhance the explanation of the results through visualization. This perspective thus aims to provide a guidance tool for developing explainability solutions for healthcare by empowering both subject matter experts, providing them with a survey of available tools, and explainability developers, by providing examples of how such methods can influence in practice adoption of solutions.

Applying Artificial Intelligence to Address the Knowledge Gaps in Cancer Care.

BACKGROUND: Rapid advances in science challenge the timely adoption of evidence-based care in community settings. To bridge the gap between what is possible and what is practiced, we researched approaches to developing an artificial intelligence (AI) application that can provide real-time patient-specific decision support. MATERIALS AND METHODS: The Oncology Expert Advisor (OEA) was designed to simulate peer-to-peer consultation with three core functions: patient history summarization, treatment options recommendation, and management advisory. Machine-learning algorithms were trained to construct a dynamic summary of patients cancer history and to suggest approved therapy or investigative trial options. All patient data used were retrospectively accrued. Ground truth was established for approximately 1,000 unique patients. The full Medline database of more than 23 million published abstracts was used as the literature corpus. RESULTS: OEA's accuracies of searching disparate sources within electronic medical records to extract complex clinical concepts from unstructured text documents varied, with F1 scores of 90%-96% for non-time-dependent concepts (e.g., diagnosis) and F1 scores of 63%-65% for time-dependent concepts (e.g., therapy history timeline). Based on constructed patient profiles, OEA suggests approved therapy options linked to supporting evidence (99.9% recall; 88% precision), and screens for eligible clinical trials on ClinicalTrials.gov (97.9% recall; 96.9% precision). CONCLUSION: Our results demonstrated technical feasibility of an AI-powered application to construct longitudinal patient profiles in context and to suggest evidence-based treatment and trial options. Our experience highlighted the necessity of collaboration across clinical and AI domains, and the requirement of clinical expertise throughout the process, from design to training to testing. IMPLICATIONS FOR PRACTICE: Artificial intelligence (AI)-powered digital advisors such as the Oncology Expert Advisor have the potential to augment the capacity and update the knowledge base of practicing oncologists. By constructing dynamic patient profiles from disparate data sources and organizing and vetting vast literature for relevance to a specific patient, such AI applications could empower oncologists to consider all therapy options based on the latest scientific evidence for their patients, and help them spend less time on information "hunting and gathering" and more time with the patients. However, realization of this will require not only AI technology maturation but also active participation and leadership by clincial experts.

Selective inhibition of histone deacetylases sensitizes malignant cells to death receptor ligands.

Evasion of death receptor ligand-induced apoptosis represents an important contributor to cancer development and progression. Therefore, molecules that restore sensitivity to death receptor stimuli would be important tools to better understand this biological pathway and potential leads for therapeutic adjuncts. Previously, the small-molecule 4-(4-chloro-2-methylphenoxy)-N-hydroxybutanamide (that we propose be named droxinostat) was identified as a chemical sensitizer to death receptor stimuli, decreasing the expression of the caspase-8 inhibitor FLIP. However, the direct targets of droxinostat were unknown. To better understand the mechanism of action of droxinostat and highlight new strategies to restore sensitivity to death receptor ligands, we analyzed changes in gene expression using the Connectivity Map after treating cells with droxinostat. Changes in gene expression after droxinostat treatment resembled changes observed after treatment with histone deacetylase (HDAC) inhibitors. Therefore, we examined the effects of droxinostat on HDAC activity and showed that it selectively inhibited HDAC3, HDAC6, and HDAC8 and that inhibition of these HDACs was functionally important for its ability to sensitize cells to death ligands. Thus, we have identified a selective HDAC inhibitor and showed that selective HDAC inhibition sensitizes cells to death ligands, thereby highlighting a new mechanism to overcome resistance to death receptor ligands.

An upstream insulator regulates DLK1 imprinting in AML.

DLK1 is an imprinted gene on chromosome 14. Using informative coding single nucleotide polymorphisms, we found DLK1 expression to be monoallelic in normal bone marrow, whereas it was biallelic in 76% of acute myeloid leukemia (AML) overexpressing DLK1 (61% of all AML). Quantitative methylation analysis of 7 cytosine-phosphate-guanosine-rich areas (3 upstream of or within DLK1, the putative intergenic-differentially methylated region and 3 upstream of or within MEG3) revealed a strong association between biallelic DLK1 expression and hypermethylation of a cytosine-phosphate-guanosine-rich region 18 kb upstream of DLK1. Allele-specific methylation analysis of this region revealed the alleles to be differentially methylated in normal bone marrow and monoallelic DLK1 AML, whereas there was increased methylation of both alleles in AML with biallelic expression. Moreover, chromatin immunoprecipitation analysis revealed that CCTC-binding factor binds to this region in monoallelic but not biallelic expression samples. Taken together, our data indicate that an insulator located 18 kb upstream of DLK1 plays an important role in regulating DLK1 imprinting.

Wilms' tumor 1 suppressor gene mediates antiestrogen resistance via down-regulation of estrogen receptor-alpha expression in breast cancer cells.

The antiestrogen tamoxifen has been used in the treatment of hormone-responsive breast cancer for over a decade. The loss of estrogen receptor (ER) expression is the most common mechanism for de novo antiestrogen resistance. Wilms' tumor 1 suppressor gene (WT1) is a clinically useful marker that is associated with poor prognosis in breast cancer patients; its high level expression is frequently observed in cases of breast cancer that are estrogen and progesterone receptor negative. The lack of expression of these receptors is characteristic of tumor cells that are not responsive to hormonal manipulation. To determine whether there is a linkage between WT1 expression and antiestrogen resistance in breast cancer cells, we studied the effect of WT1 on tamoxifen responsiveness in ERalpha-positive MCF-7 cells. We found that overexpression of WT1 in MCF-7 markedly abrogated tamoxifen-induced cell apoptosis and 17beta-estradiol (E(2))-mediated cell proliferation. The expressions of ERalpha and its downstream target genes were significantly repressed following overexpression of WT1, whereas the down-regulation of WT1 by WT1 shRNA could enhance ERalpha expression and the sensitivity to tamoxifen treatment in ERalpha-negative MDA468 and HCC1954 cells that express high levels of WT1. Furthermore, we have confirmed that the WT1 protein can bind to endogenous WT1 consensus sites in the proximal promoter of ERalpha and thus inhibit the transcriptional activity of the ERalpha promoter in a WT1 site sequence-specific manner. Our study clearly implicates WT1 as a mediator of antiestrogen resistance in breast cancer through down-regulation of ERalpha expression and supports the development of WT1 inhibitors as a potential means of restoring antiestrogen responsiveness in breast cancer therapy.

Lyl1 interacts with CREB1 and alters expression of CREB1 target genes.

The basic helix-loop-helix (bHLH) transcription factor family contains key regulators of cellular proliferation and differentiation as well as the suspected oncoproteins Tal1 and Lyl1. Tal1 and Lyl1 are aberrantly over-expressed in leukemia as a result of chromosomal translocations, or other genetic or epigenetic events. Protein-protein and protein-DNA interactions described so far are mediated by their highly homologous bHLH domains, while little is known about the function of other protein domains. Hetero-dimers of Tal1 and Lyl1 with E2A or HEB, decrease the rate of E2A or HEB homo-dimer formation and are poor activators of transcription. In vitro, these hetero-dimers also recognize different binding sites from homo-dimer complexes, which may also lead to inappropriate activation or repression of promoters in vivo. Both mechanisms are thought to contribute to the oncogenic potential of Tal1 and Lyl1. Despite their bHLH structural similarity, accumulating evidence suggests that Tal1 and Lyl1 target different genes. This raises the possibility that domains flanking the bHLH region, which are distinct in the two proteins, may participate in target recognition. Here we report that CREB1, a widely-expressed transcription factor and a suspected oncogene in acute myelogenous leukemia (AML) was identified as a binding partner for Lyl1 but not for Tal1. The interaction between Lyl1 and CREB1 involves the N terminal domain of Lyl1 and the Q2 and KID domains of CREB1. The histone acetyl-transferases p300 and CBP are recruited to these complexes in the absence of CREB1 Ser 133 phosphorylation. In the Id1 promoter, Lyl1 complexes direct transcriptional activation. We also found that in addition to Id1, over-expressed Lyl1 can activate other CREB1 target promoters such as Id3, cyclin D3, Brca1, Btg2 and Egr1. Moreover, approximately 50% of all gene promoters identified by ChIP-chip experiments were jointly occupied by CREB1 and Lyl1, further strengthening the association of Lyl1 with Cre binding sites. Given the newly recognized importance of CREB1 in AML, the ability of Lyl1 to modulate promoter responses to CREB1 suggests that it plays a role in the malignant phenotype by occupying different promoters than Tal1.