BACKGROUND: Histopathology images of tumor biopsies present unique challenges for applying machine learning to the diagnosis and treatment of cancer. The pathology slides are high resolution, often exceeding 1GB, have non-uniform dimensions, and often contain multiple tissue slices of varying sizes surrounded by large empty regions. The locations of abnormal or cancerous cells, which may constitute a small portion of any given tissue sample, are not annotated. Cancer image datasets are also extremely imbalanced, with most slides being associated with relatively common cancers. Since deep representations trained on natural photographs are unlikely to be optimal for classifying pathology slide images, which have different spectral ranges and spatial structure, we here describe an approach for learning features and inferring representations of cancer pathology slides based on sparse coding. RESULTS: We show that conventional transfer learning using a state-of-the-art deep learning architecture pre-trained on ImageNet (RESNET) and fine tuned for a binary tumor/no-tumor classification task achieved between 85% and 86% accuracy. However, when all layers up to the last convolutional layer in RESNET are replaced with a single feature map inferred via a sparse coding using a dictionary optimized for sparse reconstruction of unlabeled pathology slides, classification performance improves to over 93%, corresponding to a 54% error reduction. CONCLUSIONS: We conclude that a feature dictionary optimized for biomedical imagery may in general support better classification performance than does conventional transfer learning using a dictionary pre-trained on natural images.

BACKGROUND: Primer design for highly variable DNA sequences is difficult, and experimental success requires attention to many interacting constraints. The advent of next-generation sequencing methods allows the investigation of rare variants otherwise hidden deep in large populations, but requires attention to population diversity and primer localization in relatively conserved regions, in addition to recognized constraints typically considered in primer design. RESULTS: Design constraints include degenerate sites to maximize population coverage, matching of melting temperatures, optimizing de novo sequence length, finding optimal bio-barcodes to allow efficient downstream analyses, and minimizing risk of dimerization. To facilitate primer design addressing these and other constraints, we created a novel computer program (PrimerDesign) that automates this complex procedure. We show its powers and limitations and give examples of successful designs for the analysis of HIV-1 populations. CONCLUSIONS: PrimerDesign is useful for researchers who want to design DNA primers and probes for analyzing highly variable DNA populations. It can be used to design primers for PCR, RT-PCR, Sanger sequencing, next-generation sequencing, and other experimental protocols targeting highly variable DNA samples.

Abstract Summary: We present a suite of on-line tools to design candidate vaccine proteins, and to assess antigen potential, using coverage of k-mers (as proxies for potential T-cell epitopes) as a metric. The vaccine design tool uses the recently published ‘mosaic’ method to generate protein sequences optimized for coverage of high-frequency k-mers; the coverage-assessment tools facilitate coverage comparisons for any potential antigens. To demonstrate these tools, we designed mosaic protein sets for B-clade HIV-1 Gag, Pol and Nef, and compared them to antigens used in a recent human vaccine trial. Availability: http://hiv.lanl.gov/content/sequence/MOSAIC/ Contact: [email protected] Supplementary information: Supplementary data are available at ftp://ftp-t10.lanl.gov/pub/btk/WebToolsData

Nickle et al. [1] recently reported a computational method for HIV-1 vaccine antigen design; here we compare this method with our previously published vaccine antigen design method [2], using several criteria that we believe to be important for a successful vaccine candidate [2]. The intent of both approaches is to design a set of vaccine antigens that could protect against diverse circulating strains of HIV-1 by eliciting broad T cell responses. T cells recognize short peptides of 8–12 amino acids, called epitopes, that are presented on the surface of infected cells by human HLA proteins. HIV-1 is highly variable, and both vaccine methods [1,2] attempt to provide coverage of most common variants of epitope-length fragments in HIV-1 proteins [3]. The two methods use very different computational strategies, but both achieve high coverage of potential epitopes. The COT+ method generates an antigen set consisting of one full-length synthetic sequence supplemented by a set of sequence fragments [1]. Initially, the Center of Tree (COT) sequence is computed based on a maximum likelihood phylogenetic tree inferred for a set of test sequences [4]. The “center” of the tree is identified as the point that minimizes the distances to the terminal branches, and the evolutionary model and tree topology are used to reconstruct the most likely sequence at that center point of the tree [4]. Like an inferred most recent common ancestral sequence [5], a COT sequence is an estimate of an historical entity, based on a concatenation of the most likely nucleotide at each position. The COT+ antigen design method uses the COT sequence as a foundation, and builds on it by sequential addition of protein fragments selected to provide enhanced potential epitope coverage [6]. These fragments are selected using a sliding window across all sequences; peptides are scored by the level of additional coverage they provide, and the highest-scoring fragments are added to the set, with fragments being overlapped where possible. The result is the COT+ sequence set: the full-length COT sequence plus a set of additional sequence fragments of varying lengths (Table 1, Figure 1). The summed length of all fragments are considered in units of “protein lengths”; Nickle et al. [1] illustrate the use of two protein lengths of peptides in addition to the COT protein, for a total of three gene length equivalents. They also provide an option of assembling the fragments into linear proteins (but this provided reduced population coverage relative to the basic COT+ method) [1]. Table 1 Number of Proteins or Protein Fragments Included in the Antigen Design, and the Number of Rare or Unique 9-mers in Various Protein Sets Figure 1 Relative Locations and Sequence Identities of Potential Vaccine Antigens In contrast, the mosaic method that we recently described [2] assembles a specified number of full-length protein sequences that in combination optimize coverage of epitope length peptides in a population. This method creates intact HIV proteins that have the potential for natural expression and processing of epitopes. We employed a genetic algorithm to generate sets of sequences by in silico recombination of natural protein sequences. Starting with randomly recombined natural sequences, the algorithm proceeds by a series of iterations of recombination and selection, optimizing for 9-mer coverage of the input sequences. The result is a small set of full-length protein sequences that collectively approach the upper bound on coverage attainable for a given number of proteins [2]. We designed a three-mosaic protein set using the same input data as Nickle et al. to directly compare the two methods using the same metric as Nickle et al. The mosaic set, despite being constrained to use full-length proteins, performed slightly better in terms of 9-mer coverage than three protein lengths using COT+ (Figure 2). The mosaics used three full-length protein sequences, however, compared to 38 fragments for Gag and 15 for Nef included in the optimal coverage COT+ antigen design (Figure 1, Table 1). Figure 2 Comparisons of 9-mer Coverage for the COT+ and Mosaic Antigen Designs The processing of antigens that ultimately results in the presentation of epitopes is not completely understood. Therefore, a critical issue for the COT+ method is a strategy for assembling the peptide fragments into antigens that would be practical for vaccine delivery. It is possible to simply concatenate the fragments, but this creates large numbers of spurious 9-mers at the junctions (up to eight at each junction; see Table 1), and may impact processing of embedded epitopes in the context of the assembled fragments. Some recent experimental results illustrate that problems can occur at unnatural boundaries in polyproteins. We initially designed mosaic protein sets for Env, Gag, Pol, and Nef [2], although we highlighted creating mosaics from Gag and the conserved center of Nef [2]. Nef is highly variable, but is relatively conserved near the center of the protein, where the most intense and frequent T cell responses have been observed in the setting of natural infection [7]. Therefore, we excluded the variable regions of Nef and generated a single fusion protein comprising a mosaic full-length Gag plus the relatively conserved center of Nef. The Gag+central Nef fusion gene construct was cloned into a DNA vaccine vector and transfected into 293T cells. Protein expression was evaluated by Western blot analysis using HIV-specific MAbs (241-D, Gag; 6.2 and EH1, Nef). The mosaic polyprotein was larger than a Gag protein, as expected, and Gag-specific cellular immune responses were detected by IFN-γ Elispot assay in splenocytes from mice immunized with this polyprotein construct; however, Nef-specific responses were not detected. Furthermore, using recombinant Gag and Nef protein as coating antigens in an ELISA, we detected anti-Gag but not anti-Nef antibody responses in these mice. Thus, this Gag/Nef fusion construct expresses Gag and elicits Gag-specific immune responses, but does not induce anti-Nef antibody or T cell responses in vivo. These results demonstrate that unnatural linking of polypeptides can have unexpected and undesirable consequences, and that appropriate assembly of peptides can be a nontrivial endeavor. Another important aspect of mosaic protein design is the explicit exclusion of unnatural or very rare epitope-length fragments [2]. While the excellent coverage of Mosaics illustrates that this approach captures more rare variants than other approaches (such as combining optimized natural strains or sets of consensus sequences) [1,2], unique or extremely rare variants are excluded at a user-specified rarity threshold. Including such fragments in a vaccine has the potential to elicit T cell responses that are unlikely to provide cross-protection against circulating strains, and could even divert vaccine-induced T cell responses from conserved epitopes that are relevant for protection. We therefore set a threshold such that all nine amino acid–length fragments that are incorporated into mosaic proteins are found a minimum of three times in the input data. In contrast, despite algorithmic “smoothing” [1], the COT+ protein fragments contain multiple rare and unnatural 9-mers (Table 1). An exciting potential of this type of strategy is that it may provide an approach for making a global HIV vaccine. We previously compared mosaic vaccine design at three levels: regional, within-subtype, and global [2]; in contrast, COT+ was initially applied only to the B subtype [1]. The C subtype regional comparison was based on a large South African sequence dataset [8] compared to the non–South African C subtype Los Alamos database sequences; there was no discernable advantage in creating a South African–specific mosaic versus a global C subtype mosaic vaccine in terms of coverage of South African diversity [2]. We then compared B and C subtype mosaics with mosaics designed for the entire HIV-1 M group. We found that optimizing using one HIV subtype results in a dramatic reduction in coverage of other subtypes (generally about 25% for Gag and the Nef mosaics [2], also see Figure 2E for a COT comparison). In contrast, mosaics designed using a large and representative set of M group proteins from the Los Alamos database gave only slightly lower coverage for any given subtype compared to within-subtype optimized mosaics [2]—the tradeoff for this relatively small loss of coverage is the potential for much broader (i.e., global) coverage. In Figure 2C, we show the 9-mer coverage of the M group Gag mosaics designed using the global set of M group sequences from the Los Alamos database. Coverage is comparable (within a few percent) for all subtypes tested [2] as well as for the test set from Nickle et al. [1] (Figure 2D). In conclusion, although the COT+ method is an interesting and algorithmically creative suggestion for vaccine design, the Mosaic approach has several advantages. Nickle et al. [1] note that “more computational intensive approaches such as genetic algorithm searches . . . could also be brought to bear on the problem of antigen design.” Indeed, we have already applied such an approach and were able to achieve levels of coverage approaching the achievable upper bound, without sacrificing the linear protein sequence, and with the advantage of excluding rare epitope-length variants [2]. In contrast, the antigen sets generated by the COT+ method are either fragmented pieces of protein, or a subset of those fragments joined in a linear protein sequence that provides suboptimal coverage [1]. The most important result of our mosaic vaccine study is that it may provide a tractable approach for global vaccine design: high coverage of viral variability is presented in a feasible number of intact antigens for a vaccine cocktail.