Faculty Bibliography

Primary cutaneous squamous cell carcinoma (cSCC) is responsible for ~10,000 deaths annually in the United States. Stratification of risk of poor outcome at initial biopsy would significantly impact clinical decision-making during the initial post operative period where intervention has been shown to be most effective. Using whole-slide images (WSI) from 163 patients from 3 institutions, we developed a self supervised deep-learning model to predict poor outcomes in cSCC patients from histopathological features at initial diagnosis, and validated it using WSI from 563 patients, collected from two other academic institutions. For disease-free survival prediction, the model attained a concordance index of 0.73 in the development cohort and 0.84 in the Mayo cohort. The model's interpretability revealed that features like poor differentiation and deep invasion were strongly associated with poor prognosis. Furthermore, the model is effective in stratifying risk among BWH T2a and AJCC T2, known for outcome heterogeneity.

The Hhex gene encodes a transcription factor that is important for both embryonic and post-natal development, especially of hematopoietic tissues. Hhex is one of the most common sites of retroviral integration in mouse models. We found the most common integrations in AKXD (recombinant inbred strains) T-ALLs occur 57-61kb 3' of Hhex and activate Hhex gene expression. The genomic region of murine leukemia virus (MLV) integrations has features of a developmental stage-specific cis regulatory element (CRE), as evidenced by ATAC-seq in murine progenitor cells and high H3K27 acetylation at the syntenic CRE in human hematopoietic cell lines. With ChIP-exonuclease, we describe occupancy of LIM domain binding protein 1 (LDB1), the constitutive partner of the LIM Only-2 (LMO2), GATA1, and TAL1 transcription factors at GATA sites and a composite GATA-E box within the CRE. With virtual 4C analysis, we observed looping between this +65kb CRE and the proximal intron 1 enhancer of HHEX in primary human ETP-ALLs and in normal progenitor cells. Our results show that retroviral integrations at intergenic sites can mark and take advantage of CREs. Specifically, in the case of HHEX activation, this newly described +65kb CRE is co-opted in the pathogenesis of ETP-ALL by the LMO2/LDB1 complex.

Our previous studies revealed the existence of a Universal Receptive System that regulates interactions between cells and their environment. This system is composed of DNA- and RNA-based Teazeled receptors (TezRs) found on the surface of prokaryotic and eukaryotic cells, as well as integrases and recombinases. In the current study, we aimed to provide further insight into the regulatory role of TezR and its loss in Staphylococcus aureus gene transcription. To this end, transcriptomic analysis of S. aureus MSSA VT209 was performed following the destruction of TezRs. Bacterial RNA samples were extracted from nuclease-treated and untreated S. aureus MSSA VT209. After destruction of the DNA-based-, RNA-, or combined DNA- and RNA-based TezRs of S. aureus, 103, 150, and 93 genes were significantly differently expressed, respectively. The analysis revealed differential clustering of gene expression following the loss of different TezRs, highlighting individual cellular responses following the loss of DNA- and RNA-based TezRs. KEGG pathway gene enrichment analysis revealed that the most upregulated pathways following TezR inactivation included those related to energy metabolism, cell wall metabolism, and secretion systems. Some of the genetic pathways were related to the inhibition of biofilm formation and increased antibiotic resistance, and we confirmed this at the phenotypic level using in vitro studies. The results of this study add another line of evidence that the Universal Receptive System plays an important role in cell regulation, including cell responses to the environmental factors of clinically important pathogens, and that nucleic acid-based TezRs are functionally active parts of the extrabiome.

BACKGROUND AND OBJECTIVES/OBJECTIVE:Classical biomedical data science models are trained on a single modality and aimed at one specific task. However, the exponential increase in the size and capabilities of the foundation models inside and outside medicine shows a shift toward task-agnostic models using large-scale, often internet-based, data. Recent research into smaller foundation models trained on specific literature, such as programming textbooks, demonstrated that they can display capabilities similar to or superior to large generalist models, suggesting a potential middle ground between small task-specific and large foundation models. This study attempts to introduce a domain-specific multimodal model, Congress of Neurological Surgeons (CNS)-Contrastive Language-Image Pretraining (CLIP), developed for neurosurgical applications, leveraging data exclusively from Neurosurgery Publications. METHODS:We constructed a multimodal data set of articles from Neurosurgery Publications through PDF data collection and figure-caption extraction using an artificial intelligence pipeline for quality control. Our final data set included 24 021 figure-caption pairs. We then developed a fine-tuning protocol for the OpenAI CLIP model. The model was evaluated on tasks including neurosurgical information retrieval, computed tomography imaging classification, and zero-shot ImageNet classification. RESULTS:CNS-CLIP demonstrated superior performance in neurosurgical information retrieval with a Top-1 accuracy of 24.56%, compared with 8.61% for the baseline. The average area under receiver operating characteristic across 6 neuroradiology tasks achieved by CNS-CLIP was 0.95, slightly superior to OpenAI's Contrastive Language-Image Pretraining at 0.94 and significantly outperforming a vanilla vision transformer at 0.62. In generalist classification, CNS-CLIP reached a Top-1 accuracy of 47.55%, a decrease from the baseline of 52.37%, demonstrating a catastrophic forgetting phenomenon. CONCLUSION/CONCLUSIONS:This study presents a pioneering effort in building a domain-specific multimodal model using data from a medical society publication. The results indicate that domain-specific models, while less globally versatile, can offer advantages in specialized contexts. This emphasizes the importance of using tailored data and domain-focused development in training foundation models in neurosurgery and general medicine.

Cancer progression involves genetic and epigenetic changes that disrupt chromatin 3D organization, affecting enhancer-promoter interactions and promoting growth. Here, we provide an integrative approach, combining chromatin conformation, accessibility, and transcription analysis, validated by in silico and CRISPR-interference screens, to identify relevant 3D topologies in pediatric T cell leukemia (T-ALL and ETP-ALL). We characterize 3D hubs as regulatory centers for oncogenes and disease markers, linking them to biological processes like cell division, inflammation, and stress response. Single-cell mapping reveals heterogeneous gene activation in discrete epigenetic clones, aiding in patient stratification for relapse risk after chemotherapy. Finally, we identify MYB as a 3D hub regulator in leukemia cells and show that the targeting of key regulators leads to hub dissolution, thereby providing a novel and effective anti-leukemic strategy. Overall, our work demonstrates the relevance of studying oncogenic 3D hubs to better understand cancer biology and tumor heterogeneity and to propose novel therapeutic strategies.

PURPOSE/OBJECTIVE:Necrosis quantification in the neoadjuvant setting using pathology slide review is the most important validated prognostic marker in conventional osteosarcoma. Herein, we explored three deep learning strategies on histology samples to predict outcome for OSA in the neoadjuvant setting. EXPERIMENTAL DESIGN/METHODS:Our study relies on a training cohort from New York University (New York, NY) and an external cohort from Charles university (Prague, Czechia). We trained and validated the performance of a supervised approach that integrates neural network predictions of necrosis/tumor content, and compared predicted overall survival (OS) using Kaplan-Meier curves. Furthermore, we explored morphology-based supervised and self-supervised approaches to determine whether intrinsic histomorphological features could serve as a potential marker for OS in the setting of neoadjuvant. RESULTS:Excellent correlation between the trained network and the pathologists was obtained for the quantification of necrosis content (R2=0.899, r=0.949, p < 0.0001). OS prediction cutoffs were consistent between pathologists and the neural network (22% and 30% of necrosis, respectively). Morphology-based supervised approach predicted OS with p-value=0.0028, HR=2.43 [1.10-5.38]. The self-supervised approach corroborated the findings with clusters enriched in necrosis, fibroblastic stroma, and osteoblastic morphology associating with better OS (lg2HR; -2.366; -1.164; -1.175; 95% CI=[-2.996; -0.514]). Viable/partially viable tumor and fat necrosis were associated with worse OS (lg2HR;1.287;0.822;0.828; 95% CI=[0.38-1.974]). CONCLUSIONS:Neural networks can be used to automatically estimate the necrosis to tumor ratio, a quantitative metric predictive of survival. Furthermore, we identified alternate histomorphological biomarkers specific to the necrotic and tumor regions themselves which can be used as predictors.

As efforts to study the mechanisms of melanoma metastasis and novel therapeutic approaches multiply, researchers need accurate, high-throughput methods to evaluate the effects on tumor burden resulting from specific interventions. We show that automated quantification of tumor content from whole slide images is a compelling solution to assess in vivo experiments. In order to increase the outflow of data collection from preclinical studies, we assembled a large dataset with annotations and trained a deep neural network for the quantitative analysis of melanoma tumor content on histopathological sections of murine models. After assessing its performance in segmenting these images, the tool obtained consistent results with an orthogonal method (bioluminescence) of measuring metastasis in an experimental setting. This AI-based algorithm, made freely available to academic laboratories through a web-interface called MetFinder, promises to become an asset for melanoma researchers and pathologists interested in accurate, quantitative assessment of metastasis burden.

The interaction between tumors and their microenvironment is complex and heterogeneous. Recent developments in high-dimensional multiplexed imaging have revealed the spatial organization of tumor tissues at the molecular level. However, the discovery and thorough characterization of the tumor microenvironment (TME) remains challenging due to the scale and complexity of the images. Here, we propose a self-supervised representation learning framework, CANVAS, that enables discovery of novel types of TMEs. CANVAS is a vision transformer that directly takes high-dimensional multiplexed images and is trained using self-supervised masked image modeling. In contrast to traditional spatial analysis approaches which rely on cell segmentations, CANVAS is segmentation-free, utilizes pixel-level information, and retains local morphology and biomarker distribution information. This approach allows the model to distinguish subtle morphological differences, leading to precise separation and characterization of distinct TME signatures. We applied CANVAS to a lung tumor dataset and identified and validated a monocytic signature that is associated with poor prognosis.