Deep convolutional neural networks for automatic segmentation of thoracic organs-at-risk in radiation oncology - use of non-domain transfer learning.

PURPOSE: Segmentation of organs-at-risk (OARs) is an essential component of the radiation oncology workflow. Commonly segmented thoracic OARs include the heart, esophagus, spinal cord, and lungs. This study evaluated a convolutional neural network (CNN) for automatic segmentation of these OARs. METHODS: The dataset was created retrospectively from consecutive radiotherapy plans containing all five OARs of interest, including 22,411 CT slices from 168 patients. Patients were divided into training, validation, and test datasets according to a 66%/17%/17% split. We trained a modified U-Net, applying transfer learning from a VGG16 image classification model trained on ImageNet. The Dice coefficient and 95% Hausdorff distance on the test set for each organ was compared to a commercial atlas-based segmentation model using the Wilcoxon signed-rank test. RESULTS: On the test dataset, the median Dice coefficients for the CNN model vs. the multi-atlas model were 71% vs. 67% for the spinal cord, 96% vs. 94% for the right lung, 96%vs. 94% for the left lung, 91% vs. 85% for the heart, and 63% vs. 37% for the esophagus. The median 95% Hausdorff distances were 9.5  mm vs. 25.3 mm, 5.1  mm vs. 8.1 mm, 4.0  mm vs. 8.0 mm, 9.8  mm vs. 15.8 mm, and 9.2 mm vs. 20.0 mm for the respective organs. The results all favored the CNN model (P < 0.05). CONCLUSIONS: A 2D CNN can achieve superior results to commercial atlas-based software for OAR segmentation utilizing non-domain transfer learning, which has potential utility for quality assurance and expediting patient care.

A matched-pair analysis of stereotactic body radiotherapy (SBRT) for oligometastatic lung tumors from colorectal cancer versus early stage non-small cell lung cancer.

BACKGROUND: The use of stereotactic body radiotherapy (SBRT) for early-stage primary non-small cell lung cancer (NSCLC) reported excellent local control rates. But the optimal SBRT dose for oligometastatic lung tumors (OLTs) from colorectal cancer (CRC) has not yet been determined. This study aimed to evaluate whether SBRT to a dose of 48-60 Gy in 4-5 fractions could result in similar local outcomes for OLTs from CRC as compared to early-stage NSCLC, and to examine potential dose-response relationships for OLTs from CRC. METHODS: OLTs from CRC and primary NSCLCs treated with SBRT to 48-60 Gy in 4-5 fractions at a single institution were evaluated, and a matched-pair analysis was performed. Local recurrence-free survival (LRFS) was estimated by the Kaplan-Meier method. Univariate Cox regression was performed to identify significant predictors. RESULTS: There were 72 lung lesions in 61 patients (24 OLTs from CRC in 15 patients and 48 NSCLCs in 46 patients) were analyzed with a median follow-up of 30 months. LRFS for OLTs from CRC was significantly worse than that of NSCLC when treated with 48-60 Gy/4-5 fx (p = 0.006). The 1, 3 and 5-year LRFS of OLTs from CRC vs NSCLC were 80.6% vs. 100%, 68.6% vs. 97.2%, and 68.6% vs. 81.0%, respectively. On univariate analysis, OLTs from CRC treated with higher dose (BED10 = 132 Gy) exhibited significantly better local recurrence-free survival than those treated to lower doses (BED10 ≤ 105.6 Gy) (p = 0.0022). The 1 and 3-year LRFS rates for OLTs treated to a higher dose (BED10 = 132 Gy) were 88.9% and 81.5%, vs 33.3%, and not achieved for lower doses (BED10 ≤ 105.6 Gy). CONCLUSION: The LRFS of OLTs from CRC after SBRT of 48-60 Gy/4-5 fx was significantly worse than that of primary NSCLC. Lower dose SBRT appeared to have inferior control for OLTs of CRC in this cohort. Further studies with larger sample sizes are needed.

Mapping intact protein isoforms in discovery mode using top-down proteomics.

A full description of the human proteome relies on the challenging task of detecting mature and changing forms of protein molecules in the body. Large-scale proteome analysis has routinely involved digesting intact proteins followed by inferred protein identification using mass spectrometry. This 'bottom-up' process affords a high number of identifications (not always unique to a single gene). However, complications arise from incomplete or ambiguous characterization of alternative splice forms, diverse modifications (for example, acetylation and methylation) and endogenous protein cleavages, especially when combinations of these create complex patterns of intact protein isoforms and species. 'Top-down' interrogation of whole proteins can overcome these problems for individual proteins, but has not been achieved on a proteome scale owing to the lack of intact protein fractionation methods that are well integrated with tandem mass spectrometry. Here we show, using a new four-dimensional separation system, identification of 1,043 gene products from human cells that are dispersed into more than 3,000 protein species created by post-translational modification (PTM), RNA splicing and proteolysis. The overall system produced greater than 20-fold increases in both separation power and proteome coverage, enabling the identification of proteins up to 105 kDa and those with up to 11 transmembrane helices. Many previously undetected isoforms of endogenous human proteins were mapped, including changes in multiply modified species in response to accelerated cellular ageing (senescence) induced by DNA damage. Integrated with the latest version of the Swiss-Prot database, the data provide precise correlations to individual genes and proof-of-concept for large-scale interrogation of whole protein molecules. The technology promises to improve the link between proteomics data and complex phenotypes in basic biology and disease research.

Mass Spectrometry and Antibody-Based Characterization of Blood Vessels from Brachylophosaurus canadensis.

Structures similar to blood vessels in location, morphology, flexibility, and transparency have been recovered after demineralization of multiple dinosaur cortical bone fragments from multiple specimens, some of which are as old as 80 Ma. These structures were hypothesized to be either endogenous to the bone (i.e., of vascular origin) or the result of biofilm colonizing the empty osteonal network after degradation of original organic components. Here, we test the hypothesis that these structures are endogenous and thus retain proteins in common with extant archosaur blood vessels that can be detected with high-resolution mass spectrometry and confirmed by immunofluorescence. Two lines of evidence support this hypothesis. First, peptide sequencing of Brachylophosaurus canadensis blood vessel extracts is consistent with peptides comprising extant archosaurian blood vessels and is not consistent with a bacterial, cellular slime mold, or fungal origin. Second, proteins identified by mass spectrometry can be localized to the tissues using antibodies specific to these proteins, validating their identity. Data are available via ProteomeXchange with identifier PXD001738.

A protease for 'middle-down' proteomics.

We developed a method for restricted enzymatic proteolysis using the outer membrane protease T (OmpT) to produce large peptides (>6.3 kDa on average) for mass spectrometry-based proteomics. Using this approach to analyze prefractionated high-mass HeLa proteins, we identified 3,697 unique peptides from 1,038 proteins. We demonstrated the ability of large OmpT peptides to differentiate closely related protein isoforms and to enable the detection of many post-translational modifications.

Forced evolution in silico by artificial transposons and their genetic operators: The ant navigation problem.

Modern evolutionary computation utilizes heuristic optimizations based upon concepts borrowed from the Darwinian theory of natural selection. Their demonstrated efficacy has reawakened an interest in other aspects of contemporary biology as an inspiration for new algorithms. However, amongst the many excellent candidates for study, contemporary models of biological macroevolution attract special attention. We believe that a vital direction in this field must be algorithms that model the activity of "genomic parasites", such as transposons, in biological evolution. Many evolutionary biologists posit that it is the co-evolution of populations with their genomic parasites that permits the high efficiency of evolutionary searches found in the living world. This publication is our first step in the direction of developing a minimal assortment of algorithms that simulate the role of genomic parasites. Specifically, we started in the domain of genetic algorithms (GA) and selected the Artificial Ant Problem as a test case. This navigation problem is widely known as a classical benchmark test and possesses a large body of literature. We add new objects to the standard toolkit of GA - artificial transposons and a collection of operators that operate on them. We define these artificial transposons as a fragment of an ant's code with properties that cause it to stand apart from the rest. The minimal set of operators for transposons is a transposon mutation operator, and a transposon reproduction operator that causes a transposon to multiply within the population of hosts. An analysis of the population dynamics of transposons within the course of ant evolution showed that transposons are involved in the processes of propagation and selection of blocks of ant navigation programs. During this time, the speed of evolutionary search increases significantly. We concluded that artificial transposons, analogous to real transposons, are truly capable of acting as intelligent mutators that adapt in response to an evolutionary problem in the course of co-evolution with their hosts.

Comparing label-free quantitative peptidomics approaches to characterize diurnal variation of peptides in the rat suprachiasmatic nucleus.

Mammalian circadian rhythm is maintained by the suprachiasmatic nucleus (SCN) via an intricate set of neuropeptides and other signaling molecules. In this work, peptidomic analyses from two times of day were examined to characterize variation in SCN peptides using three different label-free quantitation approaches: spectral count, spectra index and SIEVE. Of the 448 identified peptides, 207 peptides were analyzed by two label-free methods, spectral count and spectral index. There were 24 peptides with significant (adjusted p-value < 0.01) differential peptide abundances between daytime and nighttime, including multiple peptides derived from secretogranin II, cocaine and amphetamine regulated transcript, and proprotein convertase subtilisin/kexin type 1 inhibitor. Interestingly, more peptides were analyzable and had significantly different abundances between the two time points using the spectral count and spectral index methods than with a prior analysis using the SIEVE method with the same data. The results of this study reveal the importance of using the appropriate data analysis approaches for label-free relative quantitation of peptides. The detection of significant changes in so rich a set of neuropeptides reflects the dynamic nature of the SCN and the number of influences such as feeding behavior on circadian rhythm. Using spectral count and spectral index, peptide level changes are correlated to time of day, suggesting their key role in circadian function.

Quantitative peptidomics for discovery of circadian-related peptides from the rat suprachiasmatic nucleus.

In mammals the suprachiasmatic nucleus (SCN), the master circadian clock, is sensitive to light input via the optic chiasm and synchronizes many daily biological rhythms. Here we explore variations in the expression levels of neuropeptides present in the SCN of rats using a label-free quantification approach that is based on integrating peak intensities between daytime, Zeitgeber time (ZT) 6, and nighttime, ZT 18. From nine analyses comparing the levels between these two time points, 10 endogenous peptides derived from eight prohormones exhibited significant differences in their expression levels (adjusted p-value <0.05). Of these, seven peptides derived from six prohormones, including GRP, PACAP, and CART, exhibited ≥ 30% increases at ZT 18, and the VGRPEWWMDYQ peptide derived from proenkephalin A showed a >50% increase at nighttime. Several endogenous peptides showing statistically significant changes in this study have not been previously reported to alter their levels as a function of time of day, nor have they been implicated in prior functional SCN studies. This information on peptide expression changes serves as a resource for discovering unknown peptide regulators that affect circadian rhythms in the SCN.

The MMSET histone methyl transferase switches global histone methylation and alters gene expression in t(4;14) multiple myeloma cells.

The multiple myeloma SET domain (MMSET) protein is overexpressed in multiple myeloma (MM) patients with the translocation t(4;14). Although studies have shown the involvement of MMSET/Wolf-Hirschhorn syndrome candidate 1 in development, its mode of action in the pathogenesis of MM is largely unknown. We found that MMSET is a major regulator of chromatin structure and transcription in t(4;14) MM cells. High levels of MMSET correlate with an increase in lysine 36 methylation of histone H3 and a decrease in lysine 27 methylation across the genome, leading to a more open structural state of the chromatin. Loss of MMSET expression alters adhesion properties, suppresses growth, and induces apoptosis in MM cells. Consequently, genes affected by high levels of MMSET are implicated in the p53 pathway, cell cycle regulation, and integrin signaling. Regulation of many of these genes required functional histone methyl-transferase activity of MMSET. These results implicate MMSET as a major epigenetic regulator in t(4;14)+ MM.

A proteomic survey of nonribosomal peptide and polyketide biosynthesis in actinobacteria.

Actinobacteria such as streptomycetes are renowned for their ability to produce bioactive natural products including nonribosomal peptides (NRPs) and polyketides (PKs). The advent of genome sequencing has revealed an even larger genetic repertoire for secondary metabolism with most of the small molecule products of these gene clusters still unknown. Here, we employed a "protein-first" method called PrISM (Proteomic Investigation of Secondary Metabolism) to screen 26 unsequenced actinomycetes using mass spectrometry-based proteomics for the targeted detection of expressed nonribosomal peptide synthetases or polyketide synthases. Improvements to the original PrISM screening approach (Nat. Biotechnol. 2009, 27, 951-956), for example, improved de novo peptide sequencing, have enabled the discovery of 10 NRPS/PKS gene clusters from 6 strains. Taking advantage of the concurrence of biosynthetic enzymes and the secondary metabolites they generate, two natural products were associated with their previously "orphan" gene clusters. This work has demonstrated the feasibility of a proteomics-based strategy for use in screening for NRP/PK production in actinomycetes (often >8 Mbp, high GC genomes) versus the bacilli (2-4 Mbp genomes) used previously.

Endogenous peptide discovery of the rat circadian clock: a focused study of the suprachiasmatic nucleus by ultrahigh performance tandem mass spectrometry.

Understanding how a small brain region, the suprachiasmatic nucleus (SCN), can synchronize the body's circadian rhythms is an ongoing research area. This important time-keeping system requires a complex suite of peptide hormones and transmitters that remain incompletely characterized. Here, capillary liquid chromatography and FTMS have been coupled with tailored software for the analysis of endogenous peptides present in the SCN of the rat brain. After ex vivo processing of brain slices, peptide extraction, identification, and characterization from tandem FTMS data with <5-ppm mass accuracy produced a hyperconfident list of 102 endogenous peptides, including 33 previously unidentified peptides, and 12 peptides that were post-translationally modified with amidation, phosphorylation, pyroglutamylation, or acetylation. This characterization of endogenous peptides from the SCN will aid in understanding the molecular mechanisms that mediate rhythmic behaviors in mammals.

On the scalability and requirements of whole protein mass spectrometry.

Top-down proteomics has improved over the past decade despite the significant challenges presented by the analysis of large protein ions. Here, the detection of these high mass species by electrospray-based mass spectrometry (MS) is examined from a theoretical perspective to understand the mass-dependent increases in the number of charge states, isotopic peaks, and interfering species present in typical protein mass spectra. Integrating these effects into a quantitative model captures the reduced ability to detect species over 25 kDa with the speed and sensitivity characteristic of proteomics based on <3 kDa peptide ions. The model quantifies the challenge that top-down proteomics faces with respect to current MS instrumentation and projects that depletion of (13)C and (15)N isotopes can improve detection at high mass by only <2-fold at 100 kDa whereas the effect is up to 5-fold at 10 kDa. Further, we find that supercharging electrosprayed proteins to the point of producing <5 charge states at high mass would improve detection by more than 20-fold.

Identification of Phosphorylated Human Peptides by Accurate Mass Measurement Alone.

At sufficiently high mass accuracy, it is possible to distinguish phosphorylated from unmodified peptides by mass measurement alone. We examine the feasibility of that idea, tested against a library of all possible in silico tryptic digest peptides from the human proteome database. The overlaps between in silico tryptic digest phosphopeptides generated from known phosphorylated proteins (1-12 sites) and all possible unmodified human peptides are considered for assumed mass error ranges of ±10, ±50, ±100, ±1,000, and ±10,000 ppb. We find that for mass error ±50 ppb, 95% of all phosphorylated human tryptic peptides can be distinguished from nonmodified peptides by accurate mass alone through the entire nominal mass range. We discuss the prospect of on-line LC MS/MS to identify phosphopeptide precursor ions in MS1 for selected dissociation in MS2 to identify the peptide and site(s) of phosphorylation.

Discovery of protein-DNA interactions by penalized multivariate regression.

Discovering which regulatory proteins, especially transcription factors (TFs), are active under certain experimental conditions and identifying the corresponding binding motifs is essential for understanding the regulatory circuits that control cellular programs. The experimental methods used for this purpose are laborious. Computational methods have been proven extremely effective in identifying TF-binding motifs (TFBMs). In this article, we propose a novel computational method called MotifExpress for discovering active TFBMs. Unlike existing methods, which either use only DNA sequence information or integrate sequence information with a single-sample measurement of gene expression, MotifExpress integrates DNA sequence information with gene expression measured in multiple samples. By selecting TFBMs that are significantly associated with gene expression, we can identify active TFBMs under specific experimental conditions and thus provide clues for the construction of regulatory networks. Compared with existing methods, MotifExpress substantially reduces the number of spurious results. Statistically, MotifExpress uses a penalized multivariate regression approach with a composite absolute penalty, which is highly stable and can effectively find the globally optimal set of active motifs. We demonstrate the excellent performance of MotifExpress by applying it to synthetic data and real examples of Saccharomyces cerevisiae. MotifExpress is available at http://www.stat.illinois.edu/~pingma/MotifExpress.htm.

Intact mass detection, interpretation, and visualization to automate Top-Down proteomics on a large scale.

Applying high-throughput Top-Down MS to an entire proteome requires a yet-to-be-established model for data processing. Since Top-Down is becoming possible on a large scale, we report our latest software pipeline dedicated to capturing the full value of intact protein data in automated fashion. For intact mass detection, we combine algorithms for processing MS1 data from both isotopically resolved (FT) and charge-state resolved (ion trap) LC-MS data, which are then linked to their fragment ions for database searching using ProSight. Automated determination of human keratin and tubulin isoforms is one result. Optimized for the intricacies of whole proteins, new software modules visualize proteome-scale data based on the LC retention time and intensity of intact masses and enable selective detection of PTMs to automatically screen for acetylation, phosphorylation, and methylation. Software functionality was demonstrated using comparative LC-MS data from yeast strains in addition to human cells undergoing chemical stress. We further these advances as a key aspect of realizing Top-Down MS on a proteomic scale.

Size-sorting combined with improved nanocapillary liquid chromatography-mass spectrometry for identification of intact proteins up to 80 kDa.

Despite the availability of ultra-high-resolution mass spectrometers, methods for separation and detection of intact proteins for proteome-scale analyses are still in a developmental phase. Here we report robust protocols for online LC-MS to drive high-throughput top-down proteomics in a fashion similar to that of bottom-up proteomics. Comparative work on protein standards showed that a polymeric stationary phase led to superior sensitivity over a silica-based medium in reversed-phase nanocapillary LC, with detection of proteins >50 kDa routinely accomplished in the linear ion trap of a hybrid Fourier transform mass spectrometer. Protein identification was enabled by nozzle-skimmer dissociation and detection of fragment ions with <10 ppm mass accuracy for highly specific database searching using tailored software. This overall approach led to identification of proteins up to 80 kDa, with 10-60 proteins identified in single LC-MS runs of samples from yeast and human cell lines prefractionated by their molecular mass using a gel-based sieving system.

ProSight PTM 2.0: improved protein identification and characterization for top down mass spectrometry.

ProSight PTM 2.0 (http://prosightptm2.scs.uiuc.edu) is the next generation of the ProSight PTM web-based system for the identification and characterization of proteins using top down tandem mass spectrometry. It introduces an entirely new data-driven interface, integrated Sequence Gazer for protein characterization, support for fixed modifications, terminal modifications and improved support for multiple precursor ions (multiplexing). Furthermore, it supports data import and export for local analysis and collaboration.

Autosegmentation for thoracic radiation treatment planning: A grand challenge at AAPM 2017.

PURPOSE: This report presents the methods and results of the Thoracic Auto-Segmentation Challenge organized at the 2017 Annual Meeting of American Association of Physicists in Medicine. The purpose of the challenge was to provide a benchmark dataset and platform for evaluating performance of autosegmentation methods of organs at risk (OARs) in thoracic CT images. METHODS: Sixty thoracic CT scans provided by three different institutions were separated into 36 training, 12 offline testing, and 12 online testing scans. Eleven participants completed the offline challenge, and seven completed the online challenge. The OARs were left and right lungs, heart, esophagus, and spinal cord. Clinical contours used for treatment planning were quality checked and edited to adhere to the RTOG 1106 contouring guidelines. Algorithms were evaluated using the Dice coefficient, Hausdorff distance, and mean surface distance. A consolidated score was computed by normalizing the metrics against interrater variability and averaging over all patients and structures. RESULTS: The interrater study revealed highest variability in Dice for the esophagus and spinal cord, and in surface distances for lungs and heart. Five out of seven algorithms that participated in the online challenge employed deep-learning methods. Although the top three participants using deep learning produced the best segmentation for all structures, there was no significant difference in the performance among them. The fourth place participant used a multi-atlas-based approach. The highest Dice scores were produced for lungs, with averages ranging from 0.95 to 0.98, while the lowest Dice scores were produced for esophagus, with a range of 0.55-0.72. CONCLUSION: The results of the challenge showed that the lungs and heart can be segmented fairly accurately by various algorithms, while deep-learning methods performed better on the esophagus. Our dataset together with the manual contours for all training cases continues to be available publicly as an ongoing benchmarking resource.

Review of advanced catheter technologies in radiation oncology brachytherapy procedures.

The development of new catheter and applicator technologies in recent years has significantly improved treatment accuracy, efficiency, and outcomes in brachytherapy. In this paper, we review these advances, focusing on the performance of catheter imaging and reconstruction techniques in brachytherapy procedures using magnetic resonance images and electromagnetic tracking. The accuracy of catheter reconstruction, imaging artifacts, and other notable properties of plastic and titanium applicators in gynecologic treatments are reviewed. The accuracy, noise performance, and limitations of electromagnetic tracking for catheter reconstruction are discussed. Several newly developed applicators for accelerated partial breast irradiation and gynecologic treatments are also reviewed. New hypofractionated high dose rate treatment schemes in prostate cancer and accelerated partial breast irradiation are presented.

Differential lysine acetylation profiles of Erwinia amylovora strains revealed by proteomics.

Protein lysine acetylation (LysAc) has recently been demonstrated to be widespread in E. coli and Salmonella, and to broadly regulate bacterial physiology and metabolism. However, LysAc in plant pathogenic bacteria is largely unknown. Here we first report the lysine acetylome of Erwinia amylovora, an enterobacterium causing serious fire blight disease of apples and pears. Immunoblots using generic anti-lysine acetylation antibodies demonstrated that growth conditions strongly affected the LysAc profiles in E. amylovora. Differential LysAc profiles were also observed for two E. amylovora strains, known to have differential virulence in plants, indicating translational modification of proteins may be important in determining virulence of bacterial strains. Proteomic analysis of LysAc in two E. amylovora strains identified 141 LysAc sites in 96 proteins that function in a wide range of biological pathways. Consistent with previous reports, 44% of the proteins are involved in metabolic processes, including central metabolism, lipopolysaccharide, nucleotide and amino acid metabolism. Interestingly, for the first time, several proteins involved in E. amylovora virulence, including exopolysaccharide amylovoran biosynthesis- and type III secretion-associated proteins, were found to be lysine acetylated, suggesting that LysAc may play a major role in bacterial virulence. Comparative analysis of LysAc sites in E. amylovora and E. coli further revealed the sequence and structural commonality for LysAc in the two organisms. Collectively, these results reinforce the notion that LysAc of proteins is widespread in bacterial metabolism and virulence.