A Multimodal Generative AI Copilot for Human Pathology.

The field of computational pathology[1,2] has witnessed remarkable progress in the development of both task-specific predictive models and task-agnostic self-supervised vision encoders[3,4]. However, despite the explosive growth of generative artificial intelligence (AI), there has been limited study on building general purpose, multimodal AI assistants and copilots[5] tailored to pathology. Here we present PathChat, a vision-language generalist AI assistant for human pathology. We build PathChat by adapting a foundational vision encoder for pathology, combining it with a pretrained large language model and finetuning the whole system on over 456,000 diverse visual language instructions consisting of 999,202 question-answer turns. We compare PathChat against several multimodal vision language AI assistants and GPT4V, which powers the commercially available multimodal general purpose AI assistant ChatGPT-4[7]. PathChat achieved state-of-the-art performance on multiple-choice diagnostic questions from cases of diverse tissue origins and disease models. Furthermore, using open-ended questions and human expert evaluation, we found that overall PathChat produced more accurate and pathologist-preferable responses to diverse queries related to pathology. As an interactive and general vision-language AI Copilot that can flexibly handle both visual and natural language inputs, PathChat can potentially find impactful applications in pathology education, research, and human-in-the-loop clinical decision making.

A visual-language foundation model for computational pathology.

The accelerated adoption of digital pathology and advances in deep learning have enabled the development of robust models for various pathology tasks across a diverse array of diseases and patient cohorts. However, model training is often difficult due to label scarcity in the medical domain, and a model's usage is limited by the specific task and disease for which it is trained. Additionally, most models in histopathology leverage only image data, a stark contrast to how humans teach each other and reason about histopathologic entities. We introduce CONtrastive learning from Captions for Histopathology (CONCH), a visual-language foundation model developed using diverse sources of histopathology images, biomedical text and, notably, over 1.17 million image-caption pairs through task-agnostic pretraining. Evaluated on a suite of 14 diverse benchmarks, CONCH can be transferred to a wide range of downstream tasks involving histopathology images and/or text, achieving state-of-the-art performance on histology image classification, segmentation, captioning, and text-to-image and image-to-text retrieval. CONCH represents a substantial leap over concurrent visual-language pretrained systems for histopathology, with the potential to directly facilitate a wide array of machine learning-based workflows requiring minimal or no further supervised fine-tuning.

Towards a general-purpose foundation model for computational pathology.

Quantitative evaluation of tissue images is crucial for computational pathology (CPath) tasks, requiring the objective characterization of histopathological entities from whole-slide images (WSIs). The high resolution of WSIs and the variability of morphological features present significant challenges, complicating the large-scale annotation of data for high-performance applications. To address this challenge, current efforts have proposed the use of pretrained image encoders through transfer learning from natural image datasets or self-supervised learning on publicly available histopathology datasets, but have not been extensively developed and evaluated across diverse tissue types at scale. We introduce UNI, a general-purpose self-supervised model for pathology, pretrained using more than 100 million images from over 100,000 diagnostic H&E-stained WSIs (>77 TB of data) across 20 major tissue types. The model was evaluated on 34 representative CPath tasks of varying diagnostic difficulty. In addition to outperforming previous state-of-the-art models, we demonstrate new modeling capabilities in CPath such as resolution-agnostic tissue classification, slide classification using few-shot class prototypes, and disease subtyping generalization in classifying up to 108 cancer types in the OncoTree classification system. UNI advances unsupervised representation learning at scale in CPath in terms of both pretraining data and downstream evaluation, enabling data-efficient artificial intelligence models that can generalize and transfer to a wide range of diagnostically challenging tasks and clinical workflows in anatomic pathology.

A General-Purpose Self-Supervised Model for Computational Pathology.

Tissue phenotyping is a fundamental computational pathology (CPath) task in learning objective characterizations of histopathologic biomarkers in anatomic pathology. However, whole-slide imaging (WSI) poses a complex computer vision problem in which the large-scale image resolutions of WSIs and the enormous diversity of morphological phenotypes preclude large-scale data annotation. Current efforts have proposed using pretrained image encoders with either transfer learning from natural image datasets or self-supervised pretraining on publicly-available histopathology datasets, but have not been extensively developed and evaluated across diverse tissue types at scale. We introduce UNI, a general-purpose self-supervised model for pathology, pretrained using over 100 million tissue patches from over 100,000 diagnostic haematoxylin and eosin-stained WSIs across 20 major tissue types, and evaluated on 33 representative CPath clinical tasks in CPath of varying diagnostic difficulties. In addition to outperforming previous state-of-the-art models, we demonstrate new modeling capabilities in CPath such as resolution-agnostic tissue classification, slide classification using few-shot class prototypes, and disease subtyping generalization in classifying up to 108 cancer types in the OncoTree code classification system. UNI advances unsupervised representation learning at scale in CPath in terms of both pretraining data and downstream evaluation, enabling data-efficient AI models that can generalize and transfer to a gamut of diagnostically-challenging tasks and clinical workflows in anatomic pathology.

Digital media archive for gross pathology images based on open-source tools and Fast Healthcare Interoperability Resources (FHIR).

Macroscopic examination of surgical pathology and autopsy cases is a fundamental component of anatomic pathology. The photographic documentation of such clinical specimens is essential, and it may be required in certain instances. Our department began using consumer-grade digital cameras in 2005 to improve the practice of gross photography. However, the lack of an application to correctly catalog the photographs resulted in thousands of digital image files scattered across shared network drives, with limited case and patient metadata, making image retrieval a cumbersome and sometimes impossible task. Thirteen years later, we examined the legacy method of acquiring and accessing gross photographs in our department and determined the need for a web-based digital media archive to capture images with structured metadata. Using several open-source tools, including MediaWiki, we developed a flexible platform for building our digital media archive with a data schema based on the Fast Healthcare Interoperability Resources standard. Following a short pilot, we replaced the legacy method of handling gross pathology images with a new acquisition workflow and digital media archive. Through March 2021, 233 distinct users have accessed the system, 58 of which have uploaded 21,024 images. Of those images, 13,684 (65.1%) correspond to surgical pathology images, 4045 (19.2%) belong to neuropathology cases, and 3295 (15.7%) originate from autopsies. We demonstrate the design and implementation of a customizable anatomic pathology digital media archive solution in an academic pathology department setting using a modern standard for exchanging healthcare information electronically. The system's efficiency and scalability for our current operation will enable us to integrate with other applications and pathology informatics initiatives in the future.

Recurrent PROC and novel PROS1 mutations in Vietnamese patients diagnosed with idiopathic deep venous thrombosis.

INTRODUCTION: Genetic mutations of PROC and PROS1 are well-known risk factors for deep venous thrombosis (DVT) in the Asian population. However, the genetic profile of Vietnamese patients with DVT remains elusive. This study aimed to investigate the spectrum of genetic mutations of these two genes in Vietnamese patients diagnosed with idiopathic DVT. MATERIALS AND METHODS: A total of 50 Vietnamese patients diagnosed with idiopathic DVT were recruited in this study. The entire coding regions of the protein C and protein S genes were amplified and directly sequenced to determine genetic alterations. RESULTS: Four and six genetic mutations were detected in protein C and protein S genes, respectively, in 24 Vietnamese DVT patients. PROC c.565C > T (p.R189W) was the most common mutation found in 13 out of 50 patients, while the mutations of PROS1 comprised three missense and three nonsense variants which diffuse along the gene. CONCLUSIONS: This study shows that mutations of protein C and protein S genes are prevalent in Vietnamese patients diagnosed with idiopathic DVT, and PROC c.565C > T (p.R189W) was the most common genetic alteration.

VarGrouper: A Bioinformatic Tool for Local Haplotyping of Deletion-Insertion Variants from Next-Generation Sequencing Data after Variant Calling.

Accurate genetic variant representation through nomenclature and annotation is essential for understanding functional consequence and properly noting the presence of variants across time, assays, and laboratories. Current variant calling algorithms detect single deletion-insertion variants as multiple indel and/or substitution variants from next-generation sequencing data. Consequently, these variants are separately annotated in bioinformatics pipelines, leading to inaccurate variant representation. We developed a bioinformatic solution to this problem-VarGrouper-that automatically recognizes individual variants that arise from a deletion-insertion variant and aggregates them into a single variant that can be properly annotated. This tool has been integrated into our routine clinical molecular diagnostics workflow for DNA sequencing of solid tumors. Over an 11-month period, VarGrouper variants were reported by all attending molecular pathologists involved in interpretation and represented 4.1% of all variants reported; 10.9% of cases with reportable variants contained at least one VarGrouper variant. VarGrouper improves the practice of molecular diagnostics by increasing the accuracy and consistency of variant annotation. VarGrouper is freely available for use by the molecular diagnostic community.

Clinicopathological and molecular features of SF3B1-mutated myeloproliferative neoplasms.

The introduction of next-generation sequencing has broadened the genetic landscape of myeloproliferative neoplasms (MPNs) beyond JAK2, MPL, and CALR. However, the biological role and clinical impact of most other mutations are not well defined. We interrogated 101 genes in 143 BCR-ABL1-negative MPNs in chronic phase from 2 large institutions. We detected SF3B1 mutations in 15 cases (10%) and set to investigate the clinical, morphologic, and molecular features of SF3B1 mutated (SF3B1+) MPNs in comparison to SF3B1 wild-type (SF3B1-) cases and to identify distinctive features with myelodysplastic/myeloproliferative neoplasms with ring sideroblasts (RS) and thrombocytosis, which can show partial clinical and morphological overlap with MPNs. SF3B1+ cases were enriched in primary myelofibrosis in both prefibrotic and fibrotic stage, but mutations of SF3B1 seem to occur only as a late event in the fibrotic phase of essential thrombocythemia and polycythemia vera. SF3B1+ MPNs showed borderline lower hemoglobin but no other clinical or molecular differences compared to SF3B1- MPNs. Of note, RS were present only in a subset of SF3B1+ cases (4/10) without any other feature of erythroid or granulocytic dysplasia. Our results suggest that mutations in SF3B1 are not a rare event in MPNs, especially in primary myelofibrosis and during late fibrotic stages of essential thrombocythemia and polycythemia vera, but are not associated with myelodysplastic progression. Careful examination of bone marrow and peripheral blood for morphologic dysplasia is crucial to reach the correct diagnosis and avoid a misdiagnosis of myelodysplastic/myeloproliferative neoplasms with RS and thrombocytosis, a pitfall with potential prognostic and therapeutic implications.

Expressed Gene Fusions as Frequent Drivers of Poor Outcomes in Hormone Receptor-Positive Breast Cancer.

We sought to uncover genetic drivers of hormone receptor-positive (HR+) breast cancer, using a targeted next-generation sequencing approach for detecting expressed gene rearrangements without prior knowledge of the fusion partners. We identified intergenic fusions involving driver genes, including PIK3CA, AKT3, RAF1, and ESR1, in 14% (24/173) of unselected patients with advanced HR+ breast cancer. FISH confirmed the corresponding chromosomal rearrangements in both primary and metastatic tumors. Expression of novel kinase fusions in nontransformed cells deregulates phosphoprotein signaling, cell proliferation, and survival in three-dimensional culture, whereas expression in HR+ breast cancer models modulates estrogen-dependent growth and confers hormonal therapy resistance in vitro and in vivo Strikingly, shorter overall survival was observed in patients with rearrangement-positive versus rearrangement-negative tumors. Correspondingly, fusions were uncommon (<5%) among 300 patients presenting with primary HR+ breast cancer. Collectively, our findings identify expressed gene fusions as frequent and potentially actionable drivers in HR+ breast cancer.Significance: By using a powerful clinical molecular diagnostic assay, we identified expressed intergenic fusions as frequent contributors to treatment resistance and poor survival in advanced HR+ breast cancer. The prevalence and biological and prognostic significance of these alterations suggests that their detection may alter clinical management and bring to light new therapeutic opportunities. Cancer Discov; 8(3); 336-53. ©2017 AACR.See related commentary by Natrajan et al., p. 272See related article by Liu et al., p. 354This article is highlighted in the In This Issue feature, p. 253.

Feasibility of Ultra-High-Throughput Functional Screening of Melanoma Biopsies for Discovery of Novel Cancer Drug Combinations.

Purpose: Successful development of targeted therapy combinations for cancer patients depends on first discovering such combinations in predictive preclinical models. Stable cell lines and mouse xenograft models can have genetic and phenotypic drift and may take too long to generate to be useful as a personalized medicine tool.Experimental Design: To overcome these limitations, we have used a platform of ultra-high-throughput functional screening of primary biopsies preserving both cancer and stroma cell populations from melanoma patients to nominate such novel combinations from a library of thousands of drug combinations in a patient-specific manner within days of biopsy. In parallel, patient-derived xenograft (PDX) mouse models were created and novel combinations tested for their ability to shrink matched PDXs.Results: The screening method identifies specific drug combinations in tumor cells with patterns that are distinct from those obtained from stable cell lines. Screening results were highly specific to individual patients. For patients with matched PDX models, we confirmed that individualized novel targeted therapy combinations could inhibit tumor growth. In particular, a combination of multi-kinase and PI3K/Akt inhibitors was effective in some BRAF-wild-type melanomas, and the addition of cediranib to the BRAF inhibitor PLX4720 was effective in a PDX model with BRAF mutation.Conclusions: This proof-of-concept study demonstrates the feasibility of using primary biopsies directly for combinatorial drug discovery, complementing stable cell lines and xenografts, but with much greater speed and efficiency. This process could potentially be used in a clinical setting to rapidly identify therapeutic strategies for individual patients. Clin Cancer Res; 23(16); 4680-92. ©2017 AACR.

KIF13B-NRG1 Gene Fusion and KRAS Amplification in a Case of Natural Progression of Lung Cancer.

An 85-year-old woman with a history of several primary lung cancers presented with liver metastases. The primary lung cancers were all managed surgically and the patient did not receive adjuvant or neoadjuvant therapy prior to presenting with metastases. Comparison of molecular testing results from the most recent primary and the liver metastases demonstrated ( a) a clonal relationship between the 2 cancers and ( 2) the presence of a KIF13B-NRG1 fusion and KRAS amplification unique to the metastases. When integrated, the molecular surgical pathology findings in this case illustrate the extent of "oncogenic drive" in preterminal lung cancer.

Colorectal cancer in Crohn's colitis is comparable to sporadic colorectal cancer.

PURPOSE: It is now recognized that Crohn's disease (CD), similar to ulcerative colitis (UC), carries an up to 20-fold higher cancer risk, and the development of colorectal carcinoma (CRC) is a major long-term complication. Once CRC is present, molecular profiling is one of the components in selecting appropriate treatment strategies; however, in contrast to UC, genetic alterations in Crohn's colitis-associated CRC are poorly understood. METHODS: In a series of 227 patients with Crohn's colitis, we identified 33 cases of CRC (~14 %) and performed targeted mutational analysis of BRAF/KRAS/NRAS and determined microsatellite status as well as immunophenotype of the tumors. RESULTS: In the CRC cohort, the median age at time of cancer diagnosis was 58 (range 34-77 vs. 59.5 in sporadic; P = 0.81) and the median CD duration was 29 years (range 6-45). As a group, CRC complicating Crohn's colitis is BRAF (97 %) and NRAS (100 %) wild type and the vast majority is microsatellite stable (94 %); KRAS-mutations were found in six cases (18 %). Stage grouping, anatomic distribution, and overall survival were similar to sporadic CRC; however, long-standing CD (≥25 years) as well as gastric-immunophenotype (MUC5AC+) was associated with significantly shorter overall survival (P = 0.0029; P = 0.036, respectively). CONCLUSION: In summary, the clinicopathological and molecular profile of CD-associated CRC is similar to that observed in sporadic CRC.

Design of a Genomics Curriculum: Competencies for Practicing Pathologists.

CONTEXT: The field of genomics is rapidly impacting medical care across specialties. To help guide test utilization and interpretation, pathologists must be knowledgeable about genomic techniques and their clinical utility. The technology allowing timely generation of genomic data is relatively new to patient care and the clinical laboratory, and therefore, many currently practicing pathologists have been trained without any molecular or genomics exposure. Furthermore, the exposure that current and recent trainees receive in this field remains inconsistent. OBJECTIVE: To assess pathologists' learning needs in genomics and to develop a curriculum to address these educational needs. DESIGN: A working group formed by the College of American Pathologists developed an initial list of genomics competencies (knowledge and skills statements) that a practicing pathologist needs to be successful. Experts in genomics were then surveyed to rate the importance of each competency. These data were used to create a final list of prioritized competencies. A subset of the working group defined subtopics and tasks for each competency. Appropriate delivery methods for the educational material were also proposed. RESULTS: A final list of 32 genomics competency statements was developed. A prioritized curriculum was created with designated subtopics and tasks associated with each competency. CONCLUSIONS: We present a genomics curriculum designed as a first step toward providing practicing pathologists with the competencies needed to practice successfully.

Anchored multiplex PCR for targeted next-generation sequencing.

We describe a rapid target enrichment method for next-generation sequencing, termed anchored multiplex PCR (AMP), that is compatible with low nucleic acid input from formalin-fixed paraffin-embedded (FFPE) specimens. AMP is effective in detecting gene rearrangements (without prior knowledge of the fusion partners), single nucleotide variants, insertions, deletions and copy number changes. Validation of a gene rearrangement panel using 319 FFPE samples showed 100% sensitivity (95% confidence limit: 96.5-100%) and 100% specificity (95% confidence limit: 99.3-100%) compared with reference assays. On the basis of our experience with performing AMP on 986 clinical FFPE samples, we show its potential as both a robust clinical assay and a powerful discovery tool, which we used to identify new therapeutically important gene fusions: ARHGEF2-NTRK1 and CHTOP-NTRK1 in glioblastoma, MSN-ROS1, TRIM4-BRAF, VAMP2-NRG1, TPM3-NTRK1 and RUFY2-RET in lung cancer, FGFR2-CREB5 in cholangiocarcinoma and PPL-NTRK1 in thyroid carcinoma. AMP is a scalable and efficient next-generation sequencing target enrichment method for research and clinical applications.

Crizotinib in ROS1-rearranged non-small-cell lung cancer.

BACKGROUND: Chromosomal rearrangements of the gene encoding ROS1 proto-oncogene receptor tyrosine kinase (ROS1) define a distinct molecular subgroup of non-small-cell lung cancers (NSCLCs) that may be susceptible to therapeutic ROS1 kinase inhibition. Crizotinib is a small-molecule tyrosine kinase inhibitor of anaplastic lymphoma kinase (ALK), ROS1, and another proto-oncogene receptor tyrosine kinase, MET. METHODS: We enrolled 50 patients with advanced NSCLC who tested positive for ROS1 rearrangement in an expansion cohort of the phase 1 study of crizotinib. Patients were treated with crizotinib at the standard oral dose of 250 mg twice daily and assessed for safety, pharmacokinetics, and response to therapy. ROS1 fusion partners were identified with the use of next-generation sequencing or reverse-transcriptase-polymerase-chain-reaction assays. RESULTS: The objective response rate was 72% (95% confidence interval [CI], 58 to 84), with 3 complete responses and 33 partial responses. The median duration of response was 17.6 months (95% CI, 14.5 to not reached). Median progression-free survival was 19.2 months (95% CI, 14.4 to not reached), with 25 patients (50%) still in follow-up for progression. Among 30 tumors that were tested, we identified 7 ROS1 fusion partners: 5 known and 2 novel partner genes. No correlation was observed between the type of ROS1 rearrangement and the clinical response to crizotinib. The safety profile of crizotinib was similar to that seen in patients with ALK-rearranged NSCLC. CONCLUSIONS: In this study, crizotinib showed marked antitumor activity in patients with advanced ROS1-rearranged NSCLC. ROS1 rearrangement defines a second molecular subgroup of NSCLC for which crizotinib is highly active. (Funded by Pfizer and others; ClinicalTrials.gov number, NCT00585195.).

BRAF V600E status adds incremental value to current risk classification systems in predicting papillary thyroid carcinoma recurrence.

BACKGROUND: Papillary thyroid cancer (PTC) recurrence risk is difficult to predict. No current risk classification system incorporates BRAF mutational status. Here, we assess the incremental value of BRAF mutational status in predicting PTC recurrence relative to existing recurrence risk algorithms. METHODS: Serial data were collected for a historical cohort having undergone total thyroidectomy for papillary thyroid carcinoma (PTC) during a 5-year period. Corresponding BRAF(V600E) testing was performed and Cox proportional hazard regression modeling, with and without BRAF status, was used to evaluate existing recurrence risk algorithms. RESULTS: The 5-year cumulative PTC recurrence incidence within our 356 patient cohort was 15%. A total of 205 (81%) of associated archived specimens were successfully genotyped, and 110 (54%) harbored the BRAF(V600E) mutation. The 5-year cumulative recurrence incidence among BRAF(V600E) patients was 20% versus 8% among BRAF wild type. BRAF(V600E) was significantly associated with time to recurrence when added to the following algorithms: AMES (hazard ratio [HR] 2.43 [confidence interval 1.08-5.49]), MACIS category (HR 2.46 [1.09-5.54]), AJCC-TNM (HR 2.51 [1.11-5.66]), and ATA recurrence-risk category (HR 2.44 [1.08-5.50]), and model discrimination improved (incremental c-index range 0.046-0.109). CONCLUSION: The addition of BRAF mutational status to established risk algorithms improves the discrimination of risk recurrence in patients undergoing total thyroidectomy for PTC.

EZH2 codon 641 mutations are common in BCL2-rearranged germinal center B cell lymphomas.

Mutations at codon 641 of EZH2 are recurrent in germinal center B cell lymphomas, and the most common variants lead to altered EZH2 enzymatic activity and enhanced tri-methylation of histone H3 at lysine 27, a repressive chromatin modification. As an initial step toward screening patients for cancer genotype-directed therapy, we developed a screening assay for EZH2 codon 641 mutations amenable for testing formalin-fixed clinical specimens, based on the sensitive SNaPshot single nucleotide extension technology. We detected EZH2 mutations in 12/55 (22%) follicular lymphomas (FL), 5/35 (14%) diffuse large B cell lymphomas with a germinal center immunophenotype (GCB-DLBCL), and 2/11 (18%) high grade B cell lymphomas with concurrent rearrangements of BCL2 and MYC. No EZH2 mutations were detected in cases of Burkitt lymphoma (0/23). EZH2 mutations were frequently associated with the presence of BCL2 rearrangement (BCL2-R) in both the FL (28% of BCL-R cases versus 0% of BCL2-WT cases, p<0.05) and GCB-DLBCL groups (33% of BCL2-R cases versus 4% of BCL2-WT cases, p<0.04), and across all lymphoma types excluding BL (27% of BCL2-R cases versus 3% of BCL2-WT cases, p<0.003). We confirmed gain-of-function activity for all previously reported EZH2 codon 641 mutation variants. Our findings suggest that EZH2 mutations constitute an additional genetic "hit" in many BCL2-rearranged germinal center B cell lymphomas. Our work may be helpful in the selection of lymphoma patients for future trials of pharmacologic agents targeting EZH2 and EZH2-regulated pathways.

Recurrent chromosomal copy number alterations in sporadic chordomas.

The molecular events in chordoma pathogenesis have not been fully delineated, particularly with respect to copy number changes. Understanding copy number alterations in chordoma may reveal critical disease mechanisms that could be exploited for tumor classification and therapy. We report the copy number analysis of 21 sporadic chordomas using array comparative genomic hybridization (CGH). Recurrent copy changes were further evaluated with immunohistochemistry, methylation specific PCR, and quantitative real-time PCR. Similar to previous findings, large copy number losses, involving chromosomes 1p, 3, 4, 9, 10, 13, 14, and 18, were more common than copy number gains. Loss of CDKN2A with or without loss of CDKN2B on 9p21.3 was observed in 16/20 (80%) unique cases of which six (30%) showed homozygous deletions ranging from 76 kilobases to 4.7 megabases. One copy loss of the 10q23.31 region which encodes PTEN was found in 16/20 (80%) cases. Loss of CDKN2A and PTEN expression in the majority of cases was not attributed to promoter methylation. Our sporadic chordoma cases did not show hotspot point mutations in some common cancer gene targets. Moreover, most of these sporadic tumors are not associated with T (brachyury) duplication or amplification. Deficiency of CDKN2A and PTEN expression, although shared across many other different types of tumors, likely represents a key aspect of chordoma pathogenesis. Sporadic chordomas may rely on mechanisms other than copy number gain if they indeed exploit T/brachyury for proliferation.