Abstract-When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.

Summary Glioblastomas are lethal cancers characterized by florid angiogenesis promoted in part by glioma stem cells (GSCs). As hypoxia regulates angiogenesis, we examined hypoxic responses in GSCs. We now demonstrate that hypoxia-inducible factor HIF2α and multiple HIF-regulated genes are preferentially expressed in GSCs in comparison to nonstem tumor cells and normal neural progenitors. In tumor specimens, HIF2α co-localizes with cancer stem cell markers. Targeting HIFs in GSCs inhibits self-renewal, proliferation and survival in vitro, and attenuates tumor initiation potential of GSCs in vivo. Analysis of a molecular database reveals that HIF2A expression correlates with poor glioma patient survival. Our results demonstrate that GSCs differentially respond to hypoxia with distinct HIF induction patterns and HIF2α may represent a promising target for anti-glioblastoma therapies. Significance Recent evidence supports the presence of cancer stem cell populations that contribute to tumor progression through preferential resistance to radiation and chemotherapy, and promotion of tumor angiogenesis, invasion, and metastasis. Therefore, the elucidation of molecular regulators of cancer stem cells may translate into improved anti-neoplastic therapies. Our work demonstrates that cancer stem cells derived from glioblastomas differentially respond to hypoxia with a distinct induction of HIF2α. We find that HIFs are critical to cancer stem cell maintenance and angiogenic drive, and that expression of HIF2α is significantly associated with poor glioma patient survival. These data further suggest that anti-angiogenic therapies can be designed to target cancer stem cell specific molecules involved in neoangiogenesis, including HIF2α and its regulated factors.

Malignant gliomas are highly lethal cancers dependent on angiogenesis. Critical tumor subpopulations within gliomas share characteristics with neural stem cells. We examined the potential of stem cell-like glioma cells (SCLGC) to support tumor angiogenesis. SCLGC isolated from human glioblastoma biopsy specimens and xenografts potently generated tumors when implanted into the brains of immunocompromised mice, whereas non-SCLGC tumor cells isolated from only a few tumors formed secondary tumors when xenotransplanted. Tumors derived from SCLGC were morphologically distinguishable from non-SCLGC tumor populations by widespread tumor angiogenesis, necrosis, and hemorrhage. To determine a potential molecular mechanism for SCLGC in angiogenesis, we measured the expression of a panel of angiogenic factors secreted by SCLGC. In comparison with matched non-SCLGC populations, SCLGC consistently secreted markedly elevated levels of vascular endothelial growth factor (VEGF), which were further induced by hypoxia. In an in vitro model of angiogenesis, SCLGC-conditioned medium significantly increased endothelial cell migration and tube formation compared with non-SCLGC tumor cell-conditioned medium. The proangiogenic effects of glioma SCLGC on endothelial cells were specifically abolished by the anti-VEGF neutralizing antibody bevacizumab, which is in clinical use for cancer therapy. Furthermore, bevacizumab displayed potent antiangiogenic efficacy in vivo and suppressed growth of xenografts derived from SCLGC but limited efficacy against xenografts derived from a matched non-SCLGC population. Together these data indicate that stem cell-like tumor cells can be a crucial source of key angiogenic factors in cancers and that targeting proangiogenic factors from stem cell-like tumor populations may be critical for patient therapy. (Cancer Res 2006; 66(16): 7843-8)

Reducing noble metal loading and increasing specific activity of oxygen evolution catalysts are omnipresent challenges in proton exchange membrane (PEM) water electrolysis, which have recently been tackled by utilizing mixed oxides of noble and non-noble elements (e.g. perovskites, IrNiO x , etc.). However, proper verification of the stability of these materials is still pending. In this work dissolution processes of various iridium-based oxides are explored by introducing a new metric, defined as the ratio between amount of evolved oxygen and dissolved iridium. The so called Stability-number is independent of loading, surface area or involved active sites and thus, provides a reasonable comparison of diverse materials with respect to stability. Furthermore it can support the clarification of dissolution mechanisms and the estimation of a catalyst's lifetime. The case study on iridium-based perovskites shows that leaching of the non-noble elements in mixed oxides leads to formation of highly active amorphous iridium oxide, the instability of which is explained by participation of activated oxygen atoms, generating short-lived vacancies that favour dissolution. These insights are considered to guide further research which should be devoted to increasing utilization of pure crystalline iridium oxide, as it is the only known structure that guarantees a high durability in acidic conditions. In case amorphous iridium oxides are used, solutions for stabilization are needed.

Long QT syndrome (LQT) is an inherited disorder that causes sudden death from cardiac arrhythmias, specifically torsade de pointes and ventricular fibrillation. We previously mapped three LQT loci: LQT1 on chromosome 11p15.5, LQT2 on 7q35-36, and LQT3 on 3p21-24. Here we report genetic linkage between LQT3 and polymorphisms within SCN5A, the cardiac sodium channel gene. Single strand conformation polymorphism and DNA sequence analyses reveal identical intragenic deletions of SCN5A in affected members of two unrelated LQT families. The deleted sequences reside in a region that is important for channel inactivation. These data suggest that mutations in SCN5A cause chromosome 3-linked LQT and indicate a likely cellular mechanism for this disorder.

When building a unified vision system or gradually adding new apabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.

No abstract

Dot1 is an evolutionarily conserved histone methyltransferase that methylates lysine-79 of histone H3 in the core domain. Unlike other histone methyltransferases, Dot1 does not contain a SET domain, and it specifically methylates nucleosomal histone H3. We have solved a 2.5 A resolution structure of the catalytic domain of human Dot1, hDOT1L, in complex with S-adenosyl-L-methionine (SAM). The structure reveals a unique organization of a mainly alpha-helical N-terminal domain and a central open alpha/beta structure, an active site consisting of a SAM binding pocket, and a potential lysine binding channel. We also show that a flexible, positively charged region at the C terminus of the catalytic domain is critical for nucleosome binding and enzymatic activity. These structural and biochemical analyses, combined with molecular modeling, provide mechanistic insights into the catalytic mechanism and nucleosomal specificity of Dot1 proteins.