Search

Pancreatic cancer is an aggressive disease that typically presents late with poor outcomes, indicating a pronounced need for early detection. In this study, we applied artificial intelligence methods to clinical data from 6 million patients (24,000 pancreatic cancer cases) in Denmark (Danish National Patient Registry (DNPR)) and from 3 million patients (3,900 cases) in the United States (US Veterans Affairs (US-VA)). We trained machine learning models on the sequence of disease codes in clinical histories and tested prediction of cancer occurrence within incremental time windows (CancerRiskNet). For cancer occurrence within 36 months, the performance of the best DNPR model has area under the receiver operating characteristic (AUROC) curve = 0.88 and decreases to AUROC (3m) = 0.83 when disease events within 3 months before cancer diagnosis are excluded from training, with an estimated relative risk of 59 for 1,000 highest-risk patients older than age 50 years. Cross-application of the Danish model to US-VA data had lower performance (AUROC = 0.71), and retraining was needed to improve performance (AUROC = 0.78, AUROC (3m) = 0.76). These results improve the ability to design realistic surveillance programs for patients at elevated risk, potentially benefiting lifespan and quality of life by early detection of this aggressive cancer.

PURPOSE Symptoms are vital outcomes for cancer clinical trials, observational research, and population-level surveillance. Patient-reported outcomes (PROs) are valuable for monitoring symptoms, yet there are many challenges to collecting PROs at scale. We sought to develop, test, and externally validate a deep learning model to extract symptoms from unstructured clinical notes in the electronic health record. METHODS We randomly selected 1,225 outpatient progress notes from among patients treated at the Dana-Farber Cancer Institute between January 2016 and December 2019 and used 1,125 notes as our training/validation data set and 100 notes as our test data set. We evaluated the performance of 10 deep learning models for detecting 80 symptoms included in the National Cancer Institute's Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events (PRO-CTCAE) framework. Model performance as compared with manual chart abstraction was assessed using standard metrics, and the highest performer was externally validated on a sample of 100 physician notes from a different clinical context. RESULTS In our training and test data sets, 75 of the 80 candidate symptoms were identified. The ELECTRA-small model had the highest performance for symptom identification at the token level (ie, at the individual symptom level), with an F1 of 0.87 and a processing time of 3.95 seconds per note. For the 10 most common symptoms in the test data set, the F1 score ranged from 0.98 for anxious to 0.86 for fatigue. For external validation of the same symptoms, the note-level performance ranged from F1 = 0.97 for diarrhea and dizziness to F1 = 0.73 for swelling. CONCLUSION Training a deep learning model to identify a wide range of electronic health record–documented symptoms relevant to cancer care is feasible. This approach could be used at the health system scale to complement to electronic PROs.

Deep Learning (DL) has the potential to optimize machine learning in both the scientific and clinical communities. However, greater expertise is required to develop DL algorithms, and the variability of implementations hinders their reproducibility, translation, and deployment. Here we present the community-driven Generally Nuanced Deep Learning Framework (GaNDLF), with the goal of lowering these barriers. GaNDLF makes the mechanism of DL development, training, and inference more stable, reproducible, interpretable, and scalable, without requiring an extensive technical background. GaNDLF aims to provide an end-to-end solution for all DL-related tasks in computational precision medicine. We demonstrate the ability of GaNDLF to analyze both radiology and histology images, with built-in support for k-fold cross-validation, data augmentation, multiple modalities and output classes. Our quantitative performance evaluation on numerous use cases, anatomies, and computational tasks supports GaNDLF as a robust application framework for deployment in clinical workflows.

Supplementary Figure 3. Scatterplots for metabolites with high correlations between their abundances in tissue and serum.

Full results of 10-fold cross-validation performance assessment for all ML models and prediction tasks (mean +/- SD).

Supplementary figure legends Supplementary table 2

Supplementary Figure 4. Heatmap of the top twenty metabolites as identified by permutation feature importance analysis of AutoML model.

Supplementary Figure 2. Heatmap of top differential (unadjusted p

Supplemental Figure 1. Fresh radical prostatectomy specimens were inked and transverse sliced from apex to base.

Histogram showing the distribution of alterations in KDM6A in males and females for 4 subtypes of urothelial carcinoma. Red bars are scaled to show the frequency (%) of female samples with an alteration in KDM6A; Gray bars are scaled to show the frequency of bi-allelic mutation in KDM6A in female samples; and blue bars show the frequency of KDM6A mutation in males. The numbers above each bar indicate the number of samples with each category of alteration; the numbers below show the total number of samples for each sex and UC subtype.

Kaplan-Meier Curves for Progression-Free Survival and Overall Survival by high TMB {greater than or equal to}7 mutations/Mb and PTEN alterations in Chemotherapy-Treated mTNBC. (A) Progression-free survival and (C) overall survival by tumor mutational burden

Purpose:Few patients with metastatic triple-negative breast cancer (mTNBC) benefit from immune checkpoint inhibitors (ICI). On the basis of immunotherapy response correlates in other cancers, we evaluated whether high tumor mutational burden (TMB) ≥10 nonsynonymous mutations/megabase and PTEN alterations, defined as nonsynonymous mutations or 1 or 2 copy deletions, were associated with clinical benefit to anti-PD-1/L1 therapy in mTNBC.Experimental Design:We identified patients with mTNBC, who consented to targeted DNA sequencing and were treated with ICIs on clinical trials between April 2014 and January 2019 at Dana-Farber Cancer Institute (Boston, MA). Objective response rate (ORR), progression-free survival (PFS), and overall survival (OS) were correlated with tumor genomic features.Results:Sixty-two women received anti-PD-1/L1 inhibitors alone (23%) or combined with targeted therapy (19%) or chemotherapy (58%). High TMB (18%) was associated with significantly longer PFS (12.5 vs. 3.7 months; P = 0.04), while PTEN alterations (29%) were associated with significantly lower ORR (6% vs. 48%; P = 0.01), shorter PFS (2.3 vs. 6.1 months; P = 0.01), and shorter OS (9.7 vs. 20.5 months; P = 0.02). Multivariate analyses confirmed that these associations were independent of performance status, prior lines of therapy, therapy regimen, and visceral metastases. The survival associations were additionally independent of PD-L1 in patients with known PD-L1 and were not found in mTNBC cohorts treated with chemotherapy (n = 90) and non-ICI regimens (n = 169).Conclusions:Among patients with mTNBC treated with anti-PD-1/L1 therapies, high TMB and PTEN alterations were associated with longer and shorter survival, respectively. These observations warrant validation in larger datasets.

Extended methods and results, including e-references

Supplementary Figure from Identification and Management of Pathogenic Variants in BRCA1, BRCA2, and PALB2 in a Tumor-Only Genomic Testing Program

Purpose:DNA damage response and repair (DDR) gene alterations are associated with increased tumor-infiltrating lymphocytes, higher genomic instability, and higher tumor mutational burden (TMB) in cancer. Whether DDR alterations are associated with clinical outcomes to programmed death ligand 1 [PD-(L)1] blockade in non–small cell lung cancer (NSCLC) is unknown.Experimental Design:Tumors from patients treated with PD-(L)1 inhibitors were analyzed using targeted next-generation sequencing (NGS). Cancers were categorized on the basis of the presence or absence of deleterious mutations across a panel of 53 DDR genes. Clinical outcomes to PD-(L)1 inhibitors were evaluated according to DDR mutation status.Results:Of 266 patients with successful NGS who received PD-(L)1 inhibitors, 132 (49.6%) were identified as having deleterious DDR mutations (DDR-positive). DDR-positive and DDR-negative groups were similar in terms of baseline clinicopathologic characteristics. The median TMB was significantly higher in the DDR-positive group compared with the DDR-negative group (12.1 vs. 7.6 mutations/megabase; P < 0.001). Compared with DDR-negative patients (N = 134), DDR-positive patients had a significantly higher objective response rate (30.3% vs. 17.2%; P = 0.01), longer median progression-free survival [PFS; 5.4 vs. 2.2 months; HR, 0.58 (95% confidence interval (CI), 0.45–0.76); P < 0.001], and longer median overall survival [OS; 18.8 vs. 9.9 months; HR, 0.57 (95% CI, 0.42–0.77); P < 0.001] with PD-(L)1 therapy. After adjusting for PD-L1, TMB, performance status, tobacco use, and line of therapy, DDR-positive status was associated with a significantly longer PFS [HR, 0.68 (95% CI, 0.51–0.92); P = 0.01] and OS [HR, 0.60 (95% CI, 0.43–0.85); P = 0.004] in multivariate analysis.Conclusions:Deleterious DDR mutations are frequent in NSCLC and are associated with improved clinical outcomes in patients with NSCLC treated with PD-(L)1 blockade.

Supplementary Material

Supplementary Table from Identification and Management of Pathogenic Variants in BRCA1, BRCA2, and PALB2 in a Tumor-Only Genomic Testing Program

Purpose:The purpose of this study is to characterize the mutational landscape across the spectrum of urothelial carcinoma (UC) to identify mutational features and potential therapeutic targets.Experimental Design:Using targeted exome sequencing (n = 237 genes), we analyzed the mutation spectra of 82 low-grade nonmuscle-invasive bladder cancers (LG-NMIBC), 126 high-grade (HG) NMIBC, 199 muscle-invasive bladder cancers (MIBC), 10 LG-upper tract urothelial cancers (LG-UTUC), and 55 HG-UTUC.Results:FGFR3 and KDM6A mutations were significantly more common in LG-NMIBC (72% and 44%, respectively) versus other bladder subtypes. FGFR3 alterations were also enriched in LG-UTUC versus HG-UTUC tumors (80% vs. 16%). In contrast, TP53 and RB1 mutations were significantly more frequent in all 3 HG urothelial carcinoma subtypes than in LG-NIMBC (45%–58% vs. 4%; 9%–22% vs. 0; respectively). Among LG-NMIBC tumors, KDM6A mutations were more common in women than in men (71% vs. 38%). HG-NMIBC and MIBC had higher tumor mutational burden (TMB) than LG-NMIBC (P = 0.001 and P < 0.01, respectively). DNA-damage repair (DDR) alterations were associated with a higher TMB in HG-NMIBC and MIBC tumors, and these two tumor types were also enriched for an APOBEC mutational signature compared with LG-NMIBC and HG-UTUC. Alterations in FGFR3, PIK3CA, and EP300 correlated with worse overall survival in HG-UTUC and occurred concurrently.Conclusions:Our analysis suggests that a fraction of MIBCs likely arise from precursor lesions other than LG-NMIBC. KDM6A mutations are twice as common in women with LG-NIMBC than those in men. DDR gene mutations and APOBEC mutagenesis drive mutations in HG-NMIBC and MIBC. UTUC has a distinct mutation profile from bladder cancer.

Supplementary Figure Legends

Purpose:Tumor-only genomic testing can uncover somatic and germline pathogenic variants [pathogenic/likely pathogenic (P/LP)] in cancer predisposition genes. We describe the prevalence of P/LPs in BRCA1/2 and PALB2 (B1B2P2) across malignancies and the frequency of clinical germline testing (CGT) in patients with P/LPs in B1B2P2 identified on tumor-only testing.Experimental Design:Among 7,575 patients with cancer tested between 2016 and 2018 with the OncoPanel tumor-only sequencing assay, we characterized P/LP frequencies by tumor type, receipt of CGT prior to or within 12 months after OncoPanel, and factors associated with CGT.Results:272 (3.6%) patients had OncoPanel-detected P/LPs in B1B2P2: 37.5% of P/LPs were in BRCA-related cancers; the remainder were in non-BRCA tumors. P/LPs were detected in ≥5% of breast, pancreatic, prostate, ovarian, nonmelanoma skin, endometrial, small cell lung, and colorectal cancers. 37.9% of patients with P/LPs received CGT prior to OncoPanel; an additional 10.7% underwent CGT within 12 months of OncoPanel. Among 132 with CGT, 88.6% had ≥1 clinical factor for CGT compared with 47.1% who did not undergo CGT. Patients with BRCA tumors were more likely to have CGT compared with those without (81.4% vs. 29.0%, P < 0.0001). Among patients with CGT, 70.5% (93/132) of P/LPs were germline.Conclusions:Tumor-only genomic testing identified P/LPs in B1B2P2 in 3.6% of patients. 52.9% of patients with tumor-detected P/LPs and without CGT did not meet personal or family history criteria for CGT. In addition, some patients with tumor-detected P/LPs were not referred for CGT, especially those with non-BRCA tumors. Given implications for treatment selection and familial cancer risk, processes to reliably trigger CGT from tumor-genomic findings are needed.

Purpose:Molecular mechanisms of acquired resistance to MET tyrosine kinase inhibitors (TKI) are poorly understood. We aimed to characterize the genomic mechanisms of resistance to type I and type II MET TKIs and their impact on sequential MET TKI therapy outcomes in patients with metastatic MET exon 14–mutant NSCLC.Experimental Design:Genomic alterations occurring at the time of progression on MET TKIs were studied using plasma and tissue next-generation sequencing (NGS).Results:A total of 20 patients had tissue or plasma available for analysis at the time of acquired resistance to a MET TKI. Genomic alterations known or suspected to be mechanisms of resistance were detected in 15 patients (75%). On-target acquired mechanisms of resistance, including single and polyclonal MET kinase domain mutations in codons H1094, G1163, L1195, D1228, Y1230, and high levels of amplification of the MET exon 14–mutant allele, were observed in 7 patients (35%). A number of off-target mechanisms of resistance were detected in 9 patients (45%), including KRAS mutations and amplifications in KRAS, EGFR, HER3, and BRAF; one case displayed both on- and off-target mechanisms of resistance. In 2 patients with on-target resistant mutations, switching between type I and type II MET TKIs resulted in second partial responses.Conclusions:On-target secondary mutations and activation of bypass signaling drive resistance to MET TKIs. A deeper understanding of these molecular mechanisms can support the development of sequential or combinatorial therapeutic strategies to overcome resistance.

Purpose: Tumor-only genomic testing can uncover somatic and germline pathogenic variants [pathogenic/likely pathogenic (P/LP)] in cancer predisposition genes. We describe the prevalence of P/LPs in BRCA1/2 and PALB2 (B1B2P2) across malignancies and the frequency of clinical germline testing (CGT) in patients with P/LPs in B1B2P2 identified on tumor-only testing. Experimental Design: Among 7,575 patients with cancer tested between 2016 and 2018 with the OncoPanel tumor-only sequencing assay, we characterized P/LP frequencies by tumor type, receipt of CGT prior to or within 12 months after OncoPanel, and factors associated with CGT. Results: 272 (3.6%) patients had OncoPanel-detected P/LPs in B1B2P2: 37.5% of P/LPs were in BRCA-related cancers; the remainder were in non-BRCA tumors. P/LPs were detected in ≥5% of breast, pancreatic, prostate, ovarian, nonmelanoma skin, endometrial, small cell lung, and colorectal cancers. 37.9% of patients with P/LPs received CGT prior to OncoPanel; an additional 10.7% underwent CGT within 12 months of OncoPanel. Among 132 with CGT, 88.6% had ≥1 clinical factor for CGT compared with 47.1% who did not undergo CGT. Patients with BRCA tumors were more likely to have CGT compared with those without (81.4% vs. 29.0%, P < 0.0001). Among patients with CGT, 70.5% (93/132) of P/LPs were germline. Conclusions: Tumor-only genomic testing identified P/LPs in B1B2P2 in 3.6% of patients. 52.9% of patients with tumor-detected P/LPs and without CGT did not meet personal or family history criteria for CGT. In addition, some patients with tumor-detected P/LPs were not referred for CGT, especially those with non-BRCA tumors. Given implications for treatment selection and familial cancer risk, processes to reliably trigger CGT from tumor-genomic findings are needed.

Gleason score, a measure of prostate tumor differentiation, is the strongest predictor of lethal prostate cancer at the time of diagnosis. Metabolomic profiling of tumor and of patient serum could identify biomarkers of aggressive disease and lead to the development of a less-invasive assay to perform active surveillance monitoring. Metabolomic profiling of prostate tissue and serum samples was performed. Metabolite levels and metabolite sets were compared across Gleason scores. Machine learning algorithms were trained and tuned to predict transformation or differentiation status from metabolite data. A total of 135 metabolites were significantly different (Padjusted < 0.05) in tumor versus normal tissue, and pathway analysis identified one sugar metabolism pathway (Padjusted = 0.03). Machine learning identified profiles that predicted tumor versus normal tissue (AUC of 0.82 ± 0.08). In tumor tissue, 25 metabolites were associated with Gleason score (unadjusted P < 0.05), 4 increased in high grade while the remainder were enriched in low grade. While pyroglutamine and 1,5-anhydroglucitol were correlated (0.73 and 0.72, respectively) between tissue and serum from the same patient, no metabolites were consistently associated with Gleason score in serum. Previously reported as well as novel metabolites with differing abundance were identified across tumor tissue. However, a “metabolite signature” for Gleason score was not obtained. This may be due to study design and analytic challenges that future studies should consider. Implications: Metabolic profiling can distinguish benign and neoplastic tissues. A novel unsupervised machine learning method can be utilized to achieve this distinction.

Purpose: DNA damage response and repair (DDR) gene alterations are associated with increased tumor-infiltrating lymphocytes, higher genomic instability, and higher tumor mutational burden (TMB) in cancer. Whether DDR alterations are associated with clinical outcomes to programmed death ligand 1 [PD-(L)1] blockade in non–small cell lung cancer (NSCLC) is unknown. Experimental Design: Tumors from patients treated with PD-(L)1 inhibitors were analyzed using targeted next-generation sequencing (NGS). Cancers were categorized on the basis of the presence or absence of deleterious mutations across a panel of 53 DDR genes. Clinical outcomes to PD-(L)1 inhibitors were evaluated according to DDR mutation status. Results: Of 266 patients with successful NGS who received PD-(L)1 inhibitors, 132 (49.6%) were identified as having deleterious DDR mutations (DDR-positive). DDR-positive and DDR-negative groups were similar in terms of baseline clinicopathologic characteristics. The median TMB was significantly higher in the DDR-positive group compared with the DDR-negative group (12.1 vs. 7.6 mutations/megabase; P < 0.001). Compared with DDR-negative patients (N = 134), DDR-positive patients had a significantly higher objective response rate (30.3% vs. 17.2%; P = 0.01), longer median progression-free survival [PFS; 5.4 vs. 2.2 months; HR, 0.58 (95% confidence interval (CI), 0.45–0.76); P < 0.001], and longer median overall survival [OS; 18.8 vs. 9.9 months; HR, 0.57 (95% CI, 0.42–0.77); P < 0.001] with PD-(L)1 therapy. After adjusting for PD-L1, TMB, performance status, tobacco use, and line of therapy, DDR-positive status was associated with a significantly longer PFS [HR, 0.68 (95% CI, 0.51–0.92); P = 0.01] and OS [HR, 0.60 (95% CI, 0.43–0.85); P = 0.004] in multivariate analysis. Conclusions: Deleterious DDR mutations are frequent in NSCLC and are associated with improved clinical outcomes in patients with NSCLC treated with PD-(L)1 blockade.

Programmed cell death ligand 1 (PD-L1) tumor proportion score (TPS) is the primary clinically-available biomarker of response to immunotherapy in non-small-cell lung cancer (NSCLC), but factors associated with PD-L1 expression are not well understood.Consecutive nonsquamous NSCLCs with successful PD-L1 assessment and targeted next-generation sequencing were included in this retrospective study. Clinicopathological characteristics, gene mutations, and copy number changes in gene and chromosomal arms were compared among three PD-L1 expression groups: negative (TPS1%), low (TPS 1%-49%), and high (TPS ≥ 50%). A Q-value0.25 was considered significant after multiple comparisons correction.A total of 909 nonsquamous NSCLCs were included. High PD-L1 expression compared with low and negative PD-L1 expression was associated with increased tobacco exposure (median pack-years: 25 versus 20 versus 20, respectively; P = 0.01), advanced stage at diagnosis (76% versus 67% versus 61% with advanced stage of disease, respectively; P0.001), and higher tumor mutational burden (TMB) (median 12.2 versus 10.6 versus 10.6 mutations/megabase, respectively; P0.001). Negative PD-L1 expression when compared with high PD-L1 expression was associated with: mutations in STK11 (19% versus 5%; Q0.001), EGFR (22% versus 11%; Q0.001), CTNNB1 (4.3% versus 0.4%; Q = 0.04), APC (5% versus 1%; Q = 0.17), and SMARCA4 (9% versus 4%; Q = 0.20); copy number loss of CD274 (PD-L1, 28% versus 6%; Q0.001), PDCD1LG2 (PD-L2, 28% versus 6%; Q0.001), and JAK2 genes (27% versus 7%; Q0.001), loss of chromosomal arm 9p (23% versus 10%; Q = 0.04), and gain of 1q (46% versus 21%; Q 0.001). High PD-L1 expression compared with negative PD-L1 expression was associated with copy number gain of CD274 (11% versus 3%; Q = 0.01) and PDCD1LG2 (11% versus 3%; Q = 0.01). NSCLCs with CD274 loss, compared with those without loss, had a lower response rate (23% versus 9%; P = 0.006) and shorter progression-free survival (3.3 versus 2.0 months; P = 0.002) on immunotherapy.PD-L1 expression is associated with specific genomic alterations and clinicopathologic characteristics in nonsquamous NSCLC.

Purpose:Molecular mechanisms of acquired resistance to MET tyrosine kinase inhibitors (TKI) are poorly understood. We aimed to characterize the genomic mechanisms of resistance to type I and type II MET TKIs and their impact on sequential MET TKI therapy outcomes in patients with metastatic MET exon 14–mutant NSCLC.Experimental Design:Genomic alterations occurring at the time of progression on MET TKIs were studied using plasma and tissue next-generation sequencing (NGS).Results:A total of 20 patients had tissue or plasma available for analysis at the time of acquired resistance to a MET TKI. Genomic alterations known or suspected to be mechanisms of resistance were detected in 15 patients (75%). On-target acquired mechanisms of resistance, including single and polyclonal MET kinase domain mutations in codons H1094, G1163, L1195, D1228, Y1230, and high levels of amplification of the MET exon 14–mutant allele, were observed in 7 patients (35%). A number of off-target mechanisms of resistance were detected in 9 patients (45%), including KRAS mutations and amplifications in KRAS, EGFR, HER3, and BRAF; one case displayed both on- and off-target mechanisms of resistance. In 2 patients with on-target resistant mutations, switching between type I and type II MET TKIs resulted in second partial responses.Conclusions:On-target secondary mutations and activation of bypass signaling drive resistance to MET TKIs. A deeper understanding of these molecular mechanisms can support the development of sequential or combinatorial therapeutic strategies to overcome resistance.

PURPOSE Given regulatory approval of immune checkpoint inhibitors in patients with mismatch repair–deficient (MMR-D) cancers agnostic to tumor type, it has become important to characterize occurrence of MMR-D and develop cost-effective screening approaches. Using a next-generation sequencing (NGS) panel (OncoPanel), we developed an algorithm to identify MMR-D frequency in tumor samples and applied it in a clinical setting with pathologist review. METHODS To predict MMR-D, we adapted methods described previously for use in NGS panels, which assess patterns of single base-pair insertion or deletion events occurring in homopolymer regions. Tumors assayed with OncoPanel between July 2013 and July 2018 were included. For tumors tested after June 2017, sequencing results were presented to pathologists in real time for clinical MMR determination, in the context of tumor mutation burden, other mutational signatures, and clinical data. RESULTS Of 20,301 tumors sequenced, 2.7% (553) were retrospectively classified as MMR-D by the algorithm. Of 4,404 samples with pathologist sign-out of MMR status, the algorithm classified 147 (3.3%) as MMR-D: in 116 cases, MMR-D was confirmed by a pathologist, five cases were overruled by the pathologist, and 26 were assessed as indeterminate. Overall, the highest frequencies of OncoPanel-inferred MMR-D were in endometrial (21%; 152/723), colorectal (9.7%; 169/1,744), and small bowel (9.3%; 9/97) cancers. When algorithm predictions were compared with historical MMR immunohistochemistry or polymerase chain reaction results in a set of 325 tumors sequenced before initiation of pathologist assessment, the overall sensitivity and specificity of the algorithm were 91.1% and 98.2%, respectively. CONCLUSION We show that targeted, tumor-only NGS can be leveraged to determine MMR signatures across tumor types, suggesting that broader biomarker screening approaches may have clinical value.

Machine learning has revolutionized every facet of human life, while also becoming more accessible and ubiquitous. Its prevalence has had a powerful impact in healthcare, with numerous applications and intelligent systems achieving clinical level expertise. However, building robust and generalizable systems relies on training algorithms in a centralized fashion using large, heterogeneous datasets. In medicine, these datasets are time consuming to annotate and difficult to collect centrally due to privacy concerns. Recently, Federated Learning has been proposed as a distributed learning technique to alleviate many of these privacy concerns by providing a decentralized training paradigm for models using large, distributed data. This new approach has become the defacto way of building machine learning models in multiple industries (e.g. edge computing, smartphones). Due to its strong potential, Federated Learning is also becoming a popular training method in healthcare, where patient privacy is of paramount concern. In this paper we performed an extensive literature review to identify state-of-the-art Federated Learning applications for cancer research and clinical oncology analysis. Our objective is to provide readers with an overview of the evolving Federated Learning landscape, with a focus on applications and algorithms in oncology space. Moreover, we hope that this review will help readers to identify potential needs and future directions for research and development.

Imaging datasets in cancer research are growing exponentially in both quantity and information density. These massive datasets may enable derivation of insights for cancer research and clinical care, but only if researchers are equipped with the tools to leverage advanced computational analysis approaches such as machine learning and artificial intelligence. In this work, we highlight three themes to guide development of such computational tools: scalability, standardization, and ease of use. We then apply these principles to develop PathML, a general-purpose research toolkit for computational pathology. We describe the design of the PathML framework and demonstrate applications in diverse use-cases. PathML is publicly available at www.pathml.com.

Computer vision (CV) approaches applied to digital pathology have informed biological discovery and development of tools to help inform clinical decision-making. However, batch effects in the images have the potential to introduce spurious confounders and represent a major challenge to effective analysis and interpretation of these data. Standard methods to circumvent learning such confounders include (i) application of image augmentation techniques and (ii) examination of the learning process by evaluating through external validation (e.g., unseen data coming from a comparable dataset collected at another hospital). Here, we show that the source site of a histopathology slide can be learned from the image using CV algorithms in spite of image augmentation, and we explore these source site predictions using interpretability tools. A CV model trained using Empirical Risk Minimization (ERM) risks learning this source-site signal as a spurious correlate in the weak-label regime, which we abate by using a training method with abstention. We find that a patch based classifier trained using abstention outperformed a model trained using ERM by 9.9, 10 and 19.4% F1 in the binary classification tasks of identifying tumor versus normal tissue in lung adenocarcinoma, Gleason score in prostate adenocarcinoma, and tumor tissue grade in clear cell renal cell carcinoma, respectively, at the expense of up to 80% coverage (defined as the percent of tiles not abstained on by the model). Further, by examining the areas abstained by the model, we find that the model trained using abstention is more robust to heterogeneity, artifacts and spurious correlates in the tissue. Thus, a method trained with abstention may offer novel insights into relevant areas of the tissue contributing to a particular phenotype. Together, we suggest using data augmentation methods that help mitigate a digital pathology model’s reliance on potentially spurious visual features, as well as selecting models that can identify features truly relevant for translational discovery and clinical decision support.

Imaging datasets in cancer research are growing exponentially in both quantity and information density. These massive datasets may enable derivation of insights for cancer research and clinical care, but only if researchers are equipped with the tools to leverage advanced computational analysis approaches such as machine learning and artificial intelligence. In this work, we highlight three themes to guide development of such computational tools: scalability, standardization, and ease of use. We then apply these principles to develop PathML, a general-purpose research toolkit for computational pathology. We describe the design of the PathML framework and demonstrate applications in diverse use cases. PathML is publicly available at www.pathml.com.

Machine learning has become an increasingly ubiquitous technology, as big data continues to inform and influence everyday life and decision-making. Currently in healthcare, as well as in most other industries, the two most prevalent machine learning paradigms are supervised learning and transfer learning. Both practices rely on large-scale, manually annotated datasets to train increasingly complex models. However, the requirement of data to be manually labeled leaves an excess of unused, unlabeled data available in both public and private data repositories. Self-supervised learning (SSL) is a growing area of machine learning that has the ability to take advantage of unlabeled data. Contrary to other machine learning paradigms, SSL algorithms create artificial supervisory signals from unlabeled data and pretrain algorithms on these signals. The aim of this review is two-fold: firstly, we provide a formal definition of SSL, divide SSL algorithms into their four unique subsets, and review the state-of-the-art published in each of those subsets between the years of 2014-2020. Second, this work surveys recent SSL algorithms published in healthcare, in order to provide medical experts with a clearer picture of how they can integrate SSL into their research, with the objective of leveraging unlabeled data.

SUMMARYWe exploited genetic information to assess non-genetic influences in autoimmunity. We isolated gene modules whose products physically interact with environmental exposures related to autoimmunity, and analyzed their nominal statistical evidence of association with autoimmune and non-autoimmune diseases in genome-wide association studies (GWAS) data. Epstein Barr virus (EBV) and other Herpesviruses interactomes emerged as specifically associated with multiple sclerosis (MS), possibly under common regulatory mechanisms. Analyses of MS blood and brain transcriptomes, cytofluorimetric studies of endogenous EBV-infected lymphoblastoid lines, and lesion immunohistochemistry, confirmed a dysregulation of MS-associated EBV interactors, suggesting their contribution to CD40 signaling alterations in MS. These interactors resulted enriched in modules from inherited axonopathies-causing genes, supporting a link between EBV and neurodegeneration in MS, in accord with the observed transcriptomic dysregulations in MS brains. They were also enriched with top-ranked pharmaceutical targets prioritized on a genetic basis. This study delineates a disease-specific influence of herpesviruses on MS biology.

Purpose: The purpose of this study is to characterize the mutational landscape across the spectrum of urothelial carcinoma (UC) to identify mutational features and potential therapeutic targets. Experimental Design: Using targeted exome sequencing (n = 237 genes), we analyzed the mutation spectra of 82 low-grade nonmuscle-invasive bladder cancers (LG-NMIBC), 126 high-grade (HG) NMIBC, 199 muscle-invasive bladder cancers (MIBC), 10 LG-upper tract urothelial cancers (LG-UTUC), and 55 HG-UTUC. Results: FGFR3 and KDM6A mutations were significantly more common in LG-NMIBC (72% and 44%, respectively) versus other bladder subtypes. FGFR3 alterations were also enriched in LG-UTUC versus HG-UTUC tumors (80% vs. 16%). In contrast, TP53 and RB1 mutations were significantly more frequent in all 3 HG urothelial carcinoma subtypes than in LG-NIMBC (45%–58% vs. 4%; 9%–22% vs. 0; respectively). Among LG-NMIBC tumors, KDM6A mutations were more common in women than in men (71% vs. 38%). HG-NMIBC and MIBC had higher tumor mutational burden (TMB) than LG-NMIBC (P = 0.001 and P < 0.01, respectively). DNA-damage repair (DDR) alterations were associated with a higher TMB in HG-NMIBC and MIBC tumors, and these two tumor types were also enriched for an APOBEC mutational signature compared with LG-NMIBC and HG-UTUC. Alterations in FGFR3, PIK3CA, and EP300 correlated with worse overall survival in HG-UTUC and occurred concurrently. Conclusions: Our analysis suggests that a fraction of MIBCs likely arise from precursor lesions other than LG-NMIBC. KDM6A mutations are twice as common in women with LG-NIMBC than those in men. DDR gene mutations and APOBEC mutagenesis drive mutations in HG-NMIBC and MIBC. UTUC has a distinct mutation profile from bladder cancer.

Purpose: To date no biomarker has been identified that predicts response to ICI in mTNBC. This study aimed to explore if tumor genomic alterations correlate with efficacy of PD-1/PD-L1 inhibition in patients (pts) with mTNBC. Methods: Demographic, treatment response, and long-term outcome data were collected on patients with mTNBC treated at Dana-Farber Cancer Institute (DFCI) under several clinical trials incorporating PD-1/PD-L1 inhibitors, given as monotherapy or combined with chemotherapy (CT). Pts included in this analysis had available results of targeted exon sequencing performed using Oncopanel, our institutional gene sequencing panel, on archival tumor tissue. TMB was calculated by determining the number of non-synonymous somatic mutations that occur per megabase of exonic sequence data across all genes on the panel. High TMB was defined as 310 mutations/megabase. TMB and gene alterations were correlated with objective response rate (ORR) per RECIST 1.1, progression-free (PFS) and overall survival (OS). Results: A total of 50 pts with mTNBC were included in this analysis. At baseline, the median age was 55.9 years (31.8–75.9), 60% had ECOG 0 and 40% had ECOG 1, 72% had visceral metastasis, and 46% had received 31 prior lines of systemic therapy in the metastatic setting. While 26% of pts received monotherapy [pembrolizumab (n=7, NCT02447003); atezolizumab (n=6; NCT01375842)], 74% received combination with CT [pembrolizumab plus eribulin (n=31; NCT02513472); atezolizumab plus nab-paclitaxel (n=6; NCT01633970)]. PTEN alterations were present in 30% of pts (mutations = 7; one copy number loss = 7; two copy number loss = 1). Median follow-up was 14 months (1–40). Pts with tumors harboring PTEN alterations had lower ORR (7% vs 57%; P Table 1.Multivariable analysis for PFS Hazard ratioConfidence Intervalp-valueCombination therapy0.420.16 – 1.130.009Visceral metastasis1.310.63 – 2.770.46Previous lines of therapy1.020.09 – 0.700.85ECOG 12.11.06 – 1.280.034PTEN altered3.741.65 – 8.440.002Hypermutated tumors0.850.75 – 0.970.011 Citation Format: Barroso-Sousa R, Tyekucheva S, Pernas-Simon S, Exman P, Jain E, Garrido-Castro AC, Hughes M, Bychkovsky B, Di Lascio S, Umeton R, Files J, Lindeman NI, MacConaill LE, Hodi FS, Krop IE, Dillon D, Winer EP, Wagle N, Lin NU, Mittendorf EA, Tolaney SM. PTEN alterations and tumor mutational burden (TMB) as potential predictors of resistance or response to immune checkpoint inhibitors (ICI) in metastatic triple-negative breast cancer (mTNBC) [abstract]. In: Proceedings of the 2018 San Antonio Breast Cancer Symposium; 2018 Dec 4-8; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2019;79(4 Suppl):Abstract nr P5-12-02.

Prostate cancer (PCa) is associated with several genetic alterations which play an important role in the disease heterogeneity and clinical outcome. These alterations involve gene fusion between TMPRSS2 and members of the ETS family of transcription factors like ERG, ETV1, and ETV4 together with mutations or deletions in tumor suppressors like TP53 and PTEN. The expanding wealth of digital whole slide images (WSIs) and the increasing adoption of deep learning approaches offer a unique opportunity for pathologists to streamline the detection of these alterations. Here, we used 736 haematoxylin and eosin-stained WSIs from 494 primary PCa patients to identify several key genetic alterations including ERG, ETV1, and ETV4 fusion, PTEN loss, and TP53 and SPOP mutations. Using a custom segmentation pipeline, we identified tissue regions and tiled them into high-resolution (10X magnification) patches (256X256 pixels) which were passed to our deep multiple instance learning framework. Using a pre-trained ResNet50 model, we extracted informative features which were subsequently used for training to predict slide-level labels and to detect slide regions with high diagnostic relevance. Using a 10-folds cross validation approach, we divided the data into training (80%), validation (10%) and testing (10%) sets. The training and validation data were used for training the model and hyperparameters tuning, respectively while the testing data was used to provide an unbiased evaluation of the models’ performance using the mean Area Under the Receiver Operating Characteristic (AUROC) across the ten testing folds as evaluation metric. We managed to accurately detect key molecular alterations including ERG fusion, ETV1 fusion, ETV4 fusion, and PTEN loss. Additionally, we were able to detect mutations in TP53 and SPOP together with the presence of androgen-receptor splice variant 7 (ARv7). In addition to slide-level classification, we also identified subregions with high attention score which can help pathologists identify the distinct morphological features associated with each genetic alteration. Finally, in order to examine the cellular structure associated with each genetic alteration, we used Hover-Net model to segment and classify the nuclei in the high-attention tiles. Our work highlights the utility of using WSIs to accurately identify key molecular alteration in cancer and their associated morphological and cellular features on the slide which would streamline the diagnostic process. To the best of our knowledge, this is the first study that uses routine WSIs to predict and characterize key genetic alterations in PCa. Citation Format: Mohamed Omar, Zhuoran Xu, Ryan Carelli, Jacob Rosenthal, David Brundage, Daniela C. Salles, Eddie L. Imada, Renato Umeton, Edward M. Schaeffer, Brian D. Robinson, Tamara L. Lotan, Massimo Loda, Luigi Marchionni. Using attention-based deep multiple instance learning to identify key genetic alterations in prostate cancer from whole slide images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 462.

Background: Effective biomarkers are urgently needed in the clinical settings. However, most biomarkers are currently developed from single type of omics data. The goal of this study is to identify prognostic prostate cancer signatures using transcriptomics and metabolomics profiles jointly aiming to capture wider spectrum of biological information. Methods: In this study, we included 94 tumor and 48 adjacency normal samples with both transcriptomics and metabolomics profiles from Dana-Farber/Harvard Cancer Center SPORE Prostate Cancer Cohort. There were 85 patients being followed up with median length of 2.02 years including 3 lethal and 8 progression cases. We first constructed a multi-omics covering network that contained minimal set of variable pairs but sufficiently rich to account for observed inter-patient variations. The network was built on known gene-metabolite interaction pairs from Pathway Commons as prior knowledge. Next, we used a diffusion process with each of connected gene-metabolite pairs as seeds to identify candidate signatures in the network. Signature sizes were controlled under 5 by adjusting the tree height during pruning. Hierarchical clustering and survival analyses were then performed to stratify patients into two risk groups for disease-free survival probability. Only the prognostic signatures with power higher than 0.9 from bootstrapping were kept for external validation and functional analyses. Since to our best knowledge, there is no publicly available dataset with both transcriptomics and metabolomics to validate the multi-omics signature we identified. We thus trained a rank-based kTSP classifier using gene expression data in SPORE cohort as a surrogate signature and validated it in TCGA prostate cancer cohort independently. Results: Constructed covering network consisted of 12 metabolites and 54 genes with inter-patient heterogeneities being captured efficiently. We identified one high-powered multi-omics signature (Gene: EGLN3; Metabolites: succinate, trans-4-hydroxyproline) that exhibited good prognostic value where high-risk patients had significant less time of disease-free survival (log rank test: p=0.019, power: 0.974). In addition, genomic variations were observed in different percentages of patients in high and low risk groups including NCOR1(SNV, 70.6% vs 96.3% p=0.025), NKX3.1(CNV, 29.6% vs 64.7% p=0.031) and TNFRSF10C (CNV, 29.6% vs 64.7%, p=0.031). No evidence indicated the patient grouping by the signature depend on Gleason scores (p=0.44). The surrogated gene signature contained 6 pairs of genes that can effectively classify TCGA patients into two prognostic groups (log rank test: p=0.048). Still, no evidence indicated the surrogate gene signature is associated with Gleason score (p=0.23) in TCGA dataset. Conclusions: We identified a prognostic multi-omics signature (EGLN3, succinate, trans-4-hydroxyproline) with high statistical power. Citation Format: Zhuoran Xu, Elisa Benedetti, Ryan Carelli, Jacob Rosenthal, Hubert Pakula, Mohamed Omar, Renato Umeton, David Brundage, Jan Krumsiek, Massimo Loda, Luigi Marchionni. A multi-omics signature for patients’ risk classification in prostate cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5858.

Pancreatic cancer is a leading cause of cancer-related deaths worldwide with increasing incidence. Early diagnosis of pancreatic cancer is a key challenge, as the disease is typically detected at a late stage. However, patients who present with early-stage disease can be cured by a combination of surgery, chemotherapy and radiotherapy. Thus, a better understanding of the risk factors for pancreatic cancer and detection at early stages has great potential to improve patient survival and reduce overall mortality from this aggressive malignancy. Here we exploit the power of advanced machine learning (ML) technology by focusing on the time sequence of clinical events and by predicting the risk of cancer occurrence over a multi-year time interval. This investigation was initially carried out using the Danish National Patient Registry (DNPR) and data which covers 41 years (1977 to 2018) of clinical records for 8.6 million patients, of which about 40,000 had a diagnosis of pancreatic cancer. To maximize predictive information extraction from these records we tested a range of ML methods, ranging from regression methods and ML without or with time dependence to time series methods such as GRU and Transformer. We explicitly train machine learning models on the time sequence of diseases in patient clinical histories and test the ability to predict cancer occurrence in time intervals of 3 to 60 months after risk assessment. For cancer occurrence within 36 months, the performance of the best model (AUROC=0.88; OR=47.5 for 20% recall and OR=159.0 for 10% recall), substantially exceeds that of a model without time information, even when disease events within a 3 month window before cancer diagnosis are excluded from training (AUROC[3m]=0.84). Independent training and testing on the Boston dataset reaches comparable performance (AUROC=0.87, OR=112.0 for 20% recall and OR=162.4 for 10% recall). We also extract from the AI machine an estimate of the contribution to prediction of individual disease features, e.g., obesity and diabetes. These results raise the state-of-the-art level of performance of cancer risk prediction on real-world data sets and provide support for the design of future screening trials for high-risk patients. AI on real-world clinical records has the potential to shift focus from treatment of late-stage to early-stage cancer, benefiting patients by improving lifespan and quality of life. We expect further increases in prediction accuracy with the availability of data beyond disease codes, such as prescriptions, laboratory values, and images. Citation Format: Davide Placido, Bo Yuan, Jessica X. Hjaltelin, Amalie D. Haue, Piotr J. Chmura, Chen Yuan, Jihye Kim, Renato Umeton, Gregory Antell, Alexander Chowdhury, Alexandra Franz, Lauren Brais, Elizabeth Andrews, Debora S. Marks, Aviv Regev, Peter Kraft, Brian M. Wolpin, Michael Rosenthal, Søren Brunak, Chris Sander. AI predicts risk of pancreatic cancer from disease trajectories using real-world electronic health records (EHRs) from Denmark and the USA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr LB550.