Predicting antibody binders and generating synthetic antibodies using deep learning.

The antibody drug field has continually sought improvements to methods for candidate discovery and engineering. Historically, most such methods have been laboratory-based, but informatics methods have recently started to make an impact. Deep learning, a subfield of machine learning, is rapidly gaining prominence in the biomedical research. Recent advances in microfluidics technologies and next-generation sequencing have not only revolutionized therapeutic antibody discovery, but also contributed to a vast amount of antibody repertoire sequencing data, providing opportunities for deep learning-based applications. Previously, we used microfluidics, yeast display, and deep sequencing to generate a panel of binder and non-binder antibody sequences to the cancer immunotherapy targets PD-1 and CTLA-4. Here we encoded the antibody light and heavy chain complementarity-determining regions (CDR3s) into antibody images, then built and trained convolutional neural network models to classify binders and non-binders. To improve model interpretability, we performed in silico mutagenesis to identify CDR3 residues that were important for binder classification. We further built generative deep learning models using generative adversarial network models to produce synthetic antibodies against PD-1 and CTLA-4. Our models generated variable length CDR3 sequences that resemble real sequences. Overall, our study demonstrates that deep learning methods can be leveraged to mine and learn patterns in antibody sequences, offering insights into antibody engineering, optimization, and discovery.

Single-cell transcriptomics reveals the effect of PD-L1/TGF-ß blockade on the tumor microenvironment.

BACKGROUND: The anti-tumor activity of anti-PD-1/PD-L1 therapies correlates with T cell infiltration in tumors. Thus, a major goal in oncology is to find strategies that enhance T cell infiltration and efficacy of anti-PD-1/PD-L1 therapy. TGF-ß has been shown to contribute to T cell exclusion, and anti-TGF-ß improves anti-PD-L1 efficacy in vivo. However, TGF-ß inhibition has frequently been shown to induce toxicity in the clinic, and the clinical efficacy of combination PD-L1 and TGF-ß blockade has not yet been proven. To identify strategies to overcome resistance to PD-L1 blockade, the transcriptional programs associated with PD-L1 and/or TGF-ß blockade in the tumor microenvironment should be further elucidated. RESULTS: We used single-cell RNA sequencing in a mouse model to characterize the transcriptomic effects of PD-L1 and/or TGF-ß blockade on nearly 30,000 single cells in the tumor and surrounding microenvironment. Combination treatment led to upregulation of immune response genes, including multiple chemokine genes such as CCL5, in macrophages, and downregulation of extracellular matrix genes in fibroblasts. Analysis of publicly available tumor transcriptome profiles showed that the chemokine CCL5 was strongly associated with immune cell infiltration in various human cancers. Further investigation with in vivo models showed that intratumorally administered CCL5 enhanced cytotoxic lymphocytes and the anti-tumor activity of anti-PD-L1. CONCLUSIONS: Taken together, our data could be leveraged translationally to complement or find alternatives to anti-PD-L1 plus anti-TGF-ß combination therapy, for example through companion biomarkers, and/or to identify novel targets that could be modulated to overcome resistance.

Affinity maturation of antibodies by combinatorial codon mutagenesis versus error-prone PCR.

IN VITRO: affinity maturation of therapeutic monoclonal antibodies is commonly applied to achieve desired properties, such as improved binding kinetics and affinity. Currently there are no universally accepted protocols for generation of variegated antibody libraries or selection thereof. Here, we performed affinity maturation using a yeast-based single-chain variable fragment (scFv) expression system to compare two mutagenesis methods: random mutagenesis across the entire V(D)J region by error-prone PCR, and a novel combinatorial mutagenesis process limited to the complementarity-determining regions (CDRs). We applied both methods of mutagenesis to four human antibodies against well-known immuno-oncology target proteins. Detailed sequence analysis showed an even mutational distribution across the entire length of the scFv for the error-prone PCR method and an almost exclusive targeting of the CDRs for the combinatorial method. Though there were distinct mutagenesis profiles for each target antibody and mutagenesis method, we found that both methods improved scFv affinity with similar efficiency. When a subset of the affinity-matured antibodies was expressed as full-length immunoglobulin, the measured affinity constants were mostly comparable to those of the respective scFv, but the full-length antibodies were inferior to their scFv counterparts for one of the targets. Furthermore, we found that improved affinity for the full-length antibody did not always translate into enhanced binding to cell-surface expressed antigen or improved immune checkpoint blocking ability, suggesting that screening with full-length antibody or antigen-binding fragment formats might be advantageous and the subject of a future study.

Massively parallel interrogation and mining of natively paired human TCRαß repertoires.

T cells engineered to express antigen-specific T cell receptors (TCRs) are potent therapies for viral infections and cancer. However, efficient identification of clinical candidate TCRs is complicated by the size and complexity of T cell repertoires and the challenges of working with primary T cells. Here we present a high-throughput method to identify TCRs with high functional avidity from diverse human T cell repertoires. The approach used massively parallel microfluidics to generate libraries of natively paired, full-length TCRαß clones, from millions of primary T cells, which were then expressed in Jurkat cells. The TCRαß-Jurkat libraries enabled repeated screening and panning for antigen-reactive TCRs using peptide major histocompatibility complex binding and cellular activation. We captured more than 2.9 million natively paired TCRαß clonotypes from six healthy human donors and identified rare (<0.001% frequency) viral-antigen-reactive TCRs. We also mined a tumor-infiltrating lymphocyte sample from a patient with melanoma and identified several tumor-specific TCRs, which, after expression in primary T cells, led to tumor cell killing.

Rare, high-affinity anti-pathogen antibodies from human repertoires, discovered using microfluidics and molecular genomics.

Affinity-matured, functional anti-pathogen antibodies are present at low frequencies in natural human repertoires. These antibodies are often excellent candidates for therapeutic monoclonal antibodies. However, mining natural human antibody repertoires is a challenge. In this study, we demonstrate a new method that uses microfluidics, yeast display, and deep sequencing to identify 247 natively paired anti-pathogen single-chain variable fragments (scFvs), which were initially as rare as 1 in 100,000 in the human repertoires. Influenza A vaccination increased the frequency of influenza A antigen-binding scFv within the peripheral B cell repertoire from <0.1% in non-vaccinated donors to 0.3-0.4% in vaccinated donors, whereas pneumococcus vaccination did not increase the frequency of antigen-binding scFv. However, the pneumococcus scFv binders from the vaccinated library had higher heavy and light chain Replacement/Silent mutation (R/S) ratios, a measure of affinity maturation, than the pneumococcus binders from the corresponding non-vaccinated library. Thus, pneumococcus vaccination may increase the frequency of affinity-matured antibodies in human repertoires. We synthesized 10 anti-influenza A and nine anti-pneumococcus full-length antibodies that were highly abundant among antigen-binding scFv. All 10 anti-influenza A antibodies bound the appropriate antigen at KD<10 nM and neutralized virus in cellular assays. All nine anti-pneumococcus full-length antibodies bound at least one polysaccharide serotype, and 71% of the anti-pneumococcus antibodies that we tested were functional in cell killing assays. Our approach has future application in a variety of fields, including the development of therapeutic antibodies for emerging viral diseases, autoimmune disorders, and cancer.

An integrative analysis of colon cancer identifies an essential function for PRPF6 in tumor growth.

The spliceosome machinery is composed of multimeric protein complexes that generate a diverse repertoire of mRNA through coordinated splicing of heteronuclear RNAs. While somatic mutations in spliceosome components have been discovered in several cancer types, the molecular bases and consequences of spliceosome aberrations in cancer are poorly understood. Here we report for the first time that PRPF6, a member of the tri-snRNP (small ribonucleoprotein) spliceosome complex, drives cancer proliferation by preferential splicing of genes associated with growth regulation. Inhibition of PRPF6 and other tri-snRNP complex proteins, but not other snRNP spliceosome complexes, selectively abrogated growth in cancer cells with high tri-snRNP levels. High-resolution transcriptome analyses revealed that reduced PRPF6 alters the constitutive and alternative splicing of a discrete number of genes, including an oncogenic isoform of the ZAK kinase. These findings implicate an essential role for PRPF6 in cancer via splicing of distinct growth-related gene products.

Monoclonal antibody humanness score and its applications.

BACKGROUND: Monoclonal antibody therapeutics are rapidly gaining in popularity for the treatment of a myriad of diseases, ranging from cancer to autoimmune diseases and neurological diseases. Multiple forms of antibody therapeutics are in use today that differ in the amount of human sequence present in both the constant and variable regions, where antibodies that are more human-like usually have reduced immunogenicity in clinical trials. RESULTS: Here we present a method to quantify the humanness of the variable region of monoclonal antibodies and show that this method is able to clearly distinguish human and non-human antibodies with excellent specificity. After creating and analyzing a database of human antibody sequences, we conducted an in-depth analysis of the humanness of therapeutic antibodies, and found that increased humanness score is correlated with decreased immunogenicity of antibodies. We further discovered a surprisingly similarity in the immunogenicity of fully human antibodies and humanized antibodies that are more human-like based on their humanness score. CONCLUSIONS: Our results reveal that in most cases humanizing an antibody and confirming the humanness of the final form may be sufficient to eliminate immunogenicity issues to the same extent as using fully human antibodies. We created a public website to calculate the humanness score of any input antibody sequence based on our human antibody database. This tool will be of great value during the preclinical drug development process for new monoclonal antibody therapeutics.

Lactate dehydrogenase B is required for the growth of KRAS-dependent lung adenocarcinomas.

PURPOSE: This study is aimed to identify genes within the KRAS genomic amplicon that are both coupregulated and essential for cell proliferation when KRAS is amplified in lung cancer. EXPERIMENTAL DESIGN: We used an integrated genomic approach to identify genes that are coamplified with KRAS in lung adenocarcinomas and subsequently preformed an RNA interference (RNAi) screen to uncover functionally relevant genes. The role of lactate dehydrogenase B (LDHB) was subsequently investigated both in vitro and in vivo by siRNA and short hairpin RNA (shRNA)-mediated knockdown in a panel of lung adenocarcinoma cells lines. LDHB expression was also investigated in patient tumors using microarray and immunohistochemistry analyses. RESULTS: RNAi-mediated depletion of LDHB abrogated cell proliferation both in vitro and in xenografted tumors in vivo. We find that LDHB expression correlates to both KRAS genomic copy number gain and KRAS mutation in lung cancer cell lines and adenocarcinomas. This correlation between LDHB expression and KRAS status is specific for lung cancers and not other tumor types that harbor KRAS mutations. Consistent with a role for LDHB in glycolysis and tumor metabolism, KRAS-mutant lung tumors exhibit elevated expression of a glycolysis gene signature and are more dependent on glycolysis for proliferation compared with KRAS wild-type lung tumors. Finally, high LDHB expression was a significant predictor of shorter survival in patients with lung adenocarcinomas. CONCLUSION: This study identifies LDHB as a regulator of cell proliferation in a subset of lung adenocarcinoma and may provide a novel therapeutic approach for treating lung cancer.

An integrated genomic screen identifies LDHB as an essential gene for triple-negative breast cancer.

Breast cancer has been redefined into three clinically relevant subclasses: (i) estrogen/progesterone receptor positive (ER+/PR+), (ii) HER2/ERRB2 positive, and (iii) those lacking expression of all three markers (triple negative or basal-like). While targeted therapies for ER+/PR+ and HER2+ tumors have revolutionized patient treatment and increased lifespan, an urgent need exists for identifying novel targets for triple-negative breast cancers. Here, we used integrative genomic analysis to identify candidate oncogenes in triple-negative breast tumors and assess their function through loss of function screening. Using this approach, we identify lactate dehydrogenase B (LDHB), a component of glycolytic metabolism, as an essential gene in triple-negative breast cancer. Loss of LDHB abrogated cell proliferation in vitro and arrested tumor growth in fully formed tumors in vivo. We find that LDHB and other related glycolysis genes are specifically upregulated in basal-like/triple-negative breast cancers as compared with other subtypes, suggesting that these tumors are distinctly glycolytic. Consistent with this, triple-negative breast cancer cell lines were more dependent on glycolysis for growth than luminal cell lines. Finally, we find that patients with breast cancer and high LDHB expression in their tumors had a poor clinical outcome. While previous studies have focused on the ubiquitous role of LDHA in tumor metabolism and growth, our data reveal that LDHB is upregulated and required only in certain cancer genotypes. These findings suggest that targeting LDHB or other components of lactate metabolism would be of clinical benefit in triple-negative breast cancer.

The transcription factor ZNF217 is a prognostic biomarker and therapeutic target during breast cancer progression.

UNLABELLED: The transcription factor ZNF217 is a candidate oncogene in the amplicon on chromosome 20q13 that occurs in 20% to 30% of primary human breast cancers and that correlates with poor prognosis. We show that Znf217 overexpression drives aberrant differentiation and signaling events, promotes increased self-renewal capacity, mesenchymal marker expression, motility, and metastasis, and represses an adult tissue stem cell gene signature downregulated in cancers. By in silico screening, we identified candidate therapeutics that at low concentrations inhibit growth of cancer cells expressing high ZNF217. We show that the nucleoside analogue triciribine inhibits ZNF217-induced tumor growth and chemotherapy resistance and inhibits signaling events [e.g., phospho-AKT, phospho-mitogen-activated protein kinase (MAPK)] in vivo. Our data suggest that ZNF217 is a biomarker of poor prognosis and a therapeutic target in patients with breast cancer and that triciribine may be part of a personalized treatment strategy in patients overexpressing ZNF217. Because ZNF217 is amplified in numerous cancers, these results have implications for other cancers. SIGNIFICANCE: This study finds that ZNF217 is a poor prognostic indicator and therapeutic target in patients with breast cancer and may be a strong biomarker of triciribine treatment efficacy in patients. Because previous clinical trials for triciribine did not include biomarkers of treatment efficacy, this study provides a rationale for revisiting triciribine in the clinical setting as a therapy for patients with breast cancer who overexpress ZNF217.

CDK8 maintains tumor dedifferentiation and embryonic stem cell pluripotency.

CDK8 is a cyclin-dependent kinase that mediates transcriptional control of pathways linked to both cancer and stem cells. In this study, we show that CDK8 is required for both tumor growth and maintenance of tumor dedifferentiation in vivo and uncover a common role for CDK8 in controlling cancer and stem cell function. Acute CDK8 loss in vivo strongly inhibited tumor growth and promoted differentiation. Transcriptional profiling identified a set of embryonic stem cell-related genes that are activated by CDK8 in cancer. Consistent with this, we found that CDK8 expression correlated to the embryonic stem cell pluripotency state and loss of CDK8 caused embryonic stem cells to differentiate. This effect was, at least partially, mediated by the ability of CDK8 to regulate MYC protein and downstream MYC target gene expression. Similar regulation of MYC target genes by CDK8 was observed in colon tumor cells, and increased expression of a CDK8-regulated, embryonic stem cell MYC target gene signature was associated with loss of differentiation and poor outcome in primary human colon cancers. Together, these observations reveal that CDK8 acts, at least in part, through MYC to maintain both tumors and embryonic stem cells in an undifferentiated state. This raises the intriguing possibility that targeting CDK8 therapeutically may specifically inhibit the stem-like properties of cancer cells.

G1 arrest and differentiation can occur independently of Rb family function.

The ability of progenitor cells to exit the cell cycle is essential for proper embryonic development and homeostasis, but the mechanisms governing cell cycle exit are still not fully understood. Here, we tested the requirement for the retinoblastoma (Rb) protein and its family members p107 and p130 in G0/G1 arrest and differentiation in mammalian cells. We found that Rb family triple knockout (TKO) mouse embryos survive until days 9-11 of gestation. Strikingly, some TKO cells, including in epithelial and neural lineages, are able to exit the cell cycle in G0/G1 and differentiate in teratomas and in culture. This ability of TKO cells to arrest in G0/G1 is associated with the repression of key E2F target genes. Thus, G1 arrest is not always dependent on Rb family members, which illustrates the robustness of cell cycle regulatory networks during differentiation and allows for the identification of candidate pathways to inhibit the expansion of cancer cells with mutations in the Rb pathway.

The histone demethylase UTX enables RB-dependent cell fate control.

Trimethylation of histone H3 on Lys 27 (H3K27me3) is key for cell fate regulation. The H3K27me3 demethylase UTX functions in development and tumor suppression with undefined mechanisms. Here, genome-wide chromatin occupancy analysis of UTX and associated histone modifications reveals distinct classes of UTX target genes, including genes encoding Retinoblastoma (RB)-binding proteins. UTX removes H3K27me3 and maintains expression of several RB-binding proteins, enabling cell cycle arrest. Genetic interactions in mammalian cells and Caenorhabditis elegans show that UTX regulates cell fates via RB-dependent pathways. Thus, UTX defines an evolutionarily conserved mechanism to enable coordinate transcription of a RB network in cell fate control.

SIRT6 links histone H3 lysine 9 deacetylation to NF-kappaB-dependent gene expression and organismal life span.

Members of the sirtuin (SIRT) family of NAD-dependent deacetylases promote longevity in multiple organisms. Deficiency of mammalian SIRT6 leads to shortened life span and an aging-like phenotype in mice, but the underlying molecular mechanisms are unclear. Here we show that SIRT6 functions at chromatin to attenuate NF-kappaB signaling. SIRT6 interacts with the NF-kappaB RELA subunit and deacetylates histone H3 lysine 9 (H3K9) at NF-kappaB target gene promoters. In SIRT6-deficient cells, hyperacetylation of H3K9 at these target promoters is associated with increased RELA promoter occupancy and enhanced NF-kappaB-dependent modulation of gene expression, apoptosis, and cellular senescence. Computational genomics analyses revealed increased activity of NF-kappaB-driven gene expression programs in multiple Sirt6-deficient tissues in vivo. Moreover, haploinsufficiency of RelA rescues the early lethality and degenerative syndrome of Sirt6-deficient mice. We propose that SIRT6 attenuates NF-kappaB signaling via H3K9 deacetylation at chromatin, and hyperactive NF-kappaB signaling may contribute to premature and normal aging.

Molecular framework for response to imatinib mesylate in systemic sclerosis.

Systemic sclerosis (SSc) is an autoimmune disease in which the tyrosine kinases platelet-derived growth factor receptor (PDGFR) and Abl are hypothesized to contribute to the fibrosis and vasculopathy of the skin and internal organs. Herein we describe 2 patients with early diffuse cutaneous SSc (dcSSc) who experienced reductions in cutaneous sclerosis in response to therapy with the tyrosine kinase inhibitor imatinib mesylate. Immunohistochemical analyses of skin biopsy specimens demonstrated reductions of phosphorylated PDGFRbeta and Abl with imatinib therapy. By gene expression profiling, an imatinib-responsive signature specific to dcSSc was identified (P < 10(-8)). The response of these patients and the findings of the analyses suggest that PDGFRbeta and Abl play critical, synergistic roles in the pathogenesis of SSc, and that imatinib targets a gene expression program that is frequently dysregulated in dcSSc.

Global expression profiling in atopic eczema reveals reciprocal expression of inflammatory and lipid genes.

BACKGROUND: Atopic eczema (AE) is a common chronic inflammatory skin disorder. In order to dissect the genetic background several linkage and genetic association studies have been performed. Yet very little is known about specific genes involved in this complex skin disease, and the underlying molecular mechanisms are not fully understood. METHODOLOGY/FINDINGS: We used human DNA microarrays to identify a molecular picture of the programmed responses of the human genome to AE. The transcriptional program was analyzed in skin biopsy samples from lesional and patch-tested skin from AE patients sensitized to Malassezia sympodialis (M. sympodialis), and corresponding biopsies from healthy individuals. The most notable feature of the global gene-expression pattern observed in AE skin was a reciprocal expression of induced inflammatory genes and repressed lipid metabolism genes. The overall transcriptional response in M. sympodialis patch-tested AE skin was similar to the gene-expression signature identified in lesional AE skin. In the constellation of genes differentially expressed in AE skin compared to healthy control skin, we have identified several potential susceptibility genes that may play a critical role in the pathological condition of AE. Many of these genes, including genes with a role in immune responses, lipid homeostasis, and epidermal differentiation, are localized on chromosomal regions previously linked to AE. CONCLUSIONS/SIGNIFICANCE: Through genome-wide expression profiling, we were able to discover a distinct reciprocal expression pattern of induced inflammatory genes and repressed lipid metabolism genes in skin from AE patients. We found a significant enrichment of differentially expressed genes in AE with cytobands associated to the disease, and furthermore new chromosomal regions were found that could potentially guide future region-specific linkage mapping in AE. The full data set is available at http://microarray-pubs.stanford.edu/eczema.

A C-terminal sequence in the guanine nucleotide exchange factor Sec7 mediates Golgi association and interaction with the Rsp5 ubiquitin ligase.

Arf GTPases control vesicle formation from different intracellular membranes and are regulated by Arf guanine nucleotide exchange factors (GEFs). Outside of their conserved catalytic domains, known as Sec7 domains, little is known about Arf GEFs. Rsp5 is a yeast ubiquitin ligase that regulates numerous membrane trafficking events and carries a C2 domain that is specifically required for trans-Golgi network to vacuole transport. In a screen for proteins that interact with the Rsp5 C2 domain we identified Sec7, the GEF that acts on Golgi-associated Arfs. The Rsp5-Sec7 interaction is direct, occurs in vivo, and is conserved among mammalian Rsp5 and Sec7 homologues. A 50-amino acid region near the Sec7 C terminus is required for Rsp5 binding and for normal Sec7 localization. Binding of Sec7 to Rsp5 is dependent on the presence of the phosphoinositide 3-kinase Vps34, suggesting that phosphatidylinositol 3-phosphate (PI(3)P) plays a role in regulating this interaction. Overexpression of Sec7 significantly suppresses the growth and sorting defects of an rsp5 C2 domain point mutant. These observations identify a new functional region within the Sec7/BIG family of Arf GEFs that is required for trans-Golgi network localization.

Systematic functional characterization of cis-regulatory motifs in human core promoters.

A large number of cis-regulatory motifs involved in transcriptional control have been identified, but the regulatory context and biological processes in which many of them function are unknown. Here, we computationally identify the sets of human core promoters targeted by motifs, and systematically characterize their function by using a robust gene-set-based approach and diverse sources of biological data. We find that the target sets of most motifs contain both genes with similar function and genes that are coregulated in vivo, thereby suggesting both the biological process regulated by the motifs and the conditions in which this regulation may occur. Our analysis also identifies many motifs whose target sets are predicted to be regulated by a common microRNA, suggesting a connection between transcriptional and post-transcriptional control processes. Finally, we predict novel roles for uncharacterized motifs in the regulation of specific biological processes and certain types of human cancer, and experimentally validate four such predictions, suggesting regulatory roles for four uncharacterized motifs in cell cycle progression. Our analysis thus provides a concrete framework for uncovering the biological function of cis-regulatory motifs genome wide.

Revealing targeted therapy for human cancer by gene module maps.

A major goal of cancer research is to match specific therapies to molecular targets in cancer. Genome-scale expression profiling has identified new subtypes of cancer based on consistent patterns of variation in gene expression, leading to improved prognostic predictions. However, how these new genetic subtypes of cancers should be treated is unknown. Here, we show that a gene module map can guide the prospective identification of targeted therapies for genetic subtypes of cancer. By visualizing genome-scale gene expression in cancer as combinations of activated and deactivated functional modules, gene module maps can reveal specific functional pathways associated with each subtype that might be susceptible to targeted therapies. We show that in human breast cancers, activation of a poor-prognosis "wound signature" is strongly associated with induction of both a mitochondria gene module and a proteasome gene module. We found that 3-bromopyruvic acid, which inhibits glycolysis, selectively killed breast cells expressing the mitochondria and wound signatures. In addition, inhibition of proteasome activity by bortezomib, a drug approved for human use in multiple myeloma, abrogated wound signature expression and selectively killed breast cells expressing the wound signature. Thus, gene module maps may enable rapid translation of complex genomic signatures in human disease to targeted therapeutic strategies.

CSN5 isopeptidase activity links COP9 signalosome activation to breast cancer progression.

CSN5 has been implicated as a candidate oncogene in human breast cancers by genetic linkage with activation of the poor-prognosis, wound response gene expression signature. CSN5 is a subunit of the eight-protein COP9 signalosome, a signaling complex with multiple biochemical activities; the mechanism of CSN5 action in cancer development remains poorly understood. Here, we show that CSN5 isopeptidase activity is essential for breast epithelial transformation and progression. Amplification of CSN5 is required for transformation of primary human breast epithelial cells by defined oncogenes. The transforming effects of CSN5 require CSN subunits for assembly of the full COP9 signalosome and the isopeptidase activity of CSN5, which potentiates the transcriptional activity of MYC. Transgenic inhibition of CSN5 isopeptidase activity blocks breast cancer progression evoked by MYC and RAS in vivo. These results highlight CSN5 isopeptidase activity in breast cancer progression, suggesting it as a therapeutic target in aggressive human breast cancers.