Citations

Abstract Gallbladder cancer (GBC) is one of the deadliest cancers, with a 5-year-survival-rate of less than 5 percent for late-stage disease. The response rate to chemotherapy among GBC patients is generally poor. Recent research has attempted to identify diagnostic, prognostic, and predictive biomarkers, however, currently, no biomarkers can accurately diagnose GBC and predict patients’ prognosis. Integrative analysis of molecular and clinical characterization has not been fully established, and minimal improvement has been made to the survival of these patients, in part due to the heterogeneity of GBC. Machine learning techniques have been proven to empower analysis of big data in oncology, allowing for improvement in the generation of biomarkers to predict patient outcomes. Using machine learning, we can utilize high-throughput RNA sequencing with clinicopathologic data to develop a predictive tool for GBC prognosis. Current predictive models for GBC outcomes often utilize clinical data only, with the highest C-statistic reported being 0.71. C-statistic values over 0.7 generally indicate good models, however 0.8 is the threshold for strong predictive models. We aim to build a superior algorithm to predict overall survival in GBC patients with advanced disease, using machine learning approaches to prioritize biomarkers for GBC prognosis. We have identified over 80 fresh frozen GBC tissue samples from Mayo Clinic Rochester, Dongsan Medical Center in Daegu, Korea, University of the Witwatersrand, in Johannesburg, South Africa, Lithuanian University of Health Science in Vilnius, Lithuania, and University of Calgary in Calgary, Canada, from patients enrolled between 2012 and 2021. We will perform next-generation RNA sequencing on these tissue samples. The patients’ clinical, pathologic and survival data will be abstracted from the medical record uniformly across sites. Feature engineering and dimensionality reduction will be performed. Then random forests, support vector machines, and gradient boosting machines will be applied to train the data. Variable importance will prioritize multi-omic markers. Standard 5-fold cross validation will be used to assess performance of each ML algorithm. If overall survival can be better predicted with the addition patients’ transcriptional sequencing data compared to using clinical profiles alone, we can gain a greater understanding of key biomarkers driving the tumor phenotype. Citation Format: Linsey Jackson, Loretta Allotey, Valles Kenneth, Gavin Oliver, Asha Nair, Daniel O'Brien, Rondell Graham, Mitesh Borad, Arjun Athreya, Lewis Roberts. Prognostic biomarkers for gallbladder cancer: A machine learning approach [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 1944.

Introduction: The Clinical Global Impressions-Improvement (CGI-I) scale is widely used in clinical research to assess symptoms and functioning in the context of treatment. The correlates of the CGI-I with efficacy scales for adolescent major depressive disorder are poorly understood. This study focused on benchmarking CGI-I scores with changes in the Children's Depression Rating Scale-Revised (CDRS-R) and the Quick Inventory of Depressive Symptomatology-Adolescent (17-item) Self-Report (QIDS-A17-SR). Methods: We examined three datasets with the clinician-rated CDRS-R to ascertain equivalent percent changes in total scores and CGI-I ratings. Exploratory analyses examined corresponding percentage changes in the QIDS-A17-SR and the CGI-I ratings. The CGI-I was the reference scale for nonparametric equipercentile linking with the Equate package in R. Results: CGI-I scores of 1 mapped to ≥78%–95% change in CDRS-R scores at 4–6 weeks across three datasets. CGI-I scores of 2 mapped to 56%–94% change in CDRS-R scores at 4–6 weeks across three studies. CGI-I scores of 3 mapped to 30%–68% changes in CDRS-R scores at 4–6 weeks across three studies. CGI-I scores of 4 mapped to a range of 29%–44% at 4–6 weeks across three studies. There was no significant difference ( p ≥ 0.6) between treatment groups in both the Treatment of Adolescents with Depression and Treatment of Resistant Depression in Adolescents studies, for each CGI-I score ( = 1, or = 2 or = 3, or ≥4), associated mapping of total depression severity score, or associated percent change from baseline for corresponding follow-up visits. There was no significant sex difference ( p > 0.2) in CGI-I linkages to CDRS-R total or percentage changes. Conclusions: These findings establish clear relationships among CGI-I scores and the CDRS-R and the QIDS-A17-SR. These benchmarks have utility for clinical trial study design, inter-rater reliability training, and clinical implementation.

Background: Ischemic event rates are low after percutaneous coronary intervention (PCI), conversely second generation P2Y12 inhibitors increase bleeding therefore identifying patients at risk to enable precision in prescribing dual anti-platelet therapy (DAPT) is of clinical interest. Prior machine-learning predictive analyses have been limited by use of older generation stents and lack of genetic data. We utilized clinical and genetic data of patients with new generation stents enrolled in TAILOR-PCI, a 40-center international trial to predict ischemic events after PCI. Methods: There were 4398 eligible post PCI patients with acute coronary syndromes or stable coronary artery disease who were randomly split into training (n=3299) and validation (n=1099) cohorts. The primary endpoint comprising of cardiovascular death, myocardial infarction, stroke, or stent thrombosis within 12 months of PCI was predicted with patient characteristics prior to hospital discharge. First, random forests were used for feature selection with 1,000 bootstraps comprising equal numbers of those with and without ischemic events to minimize the bias of sample imbalance. Features were selected by a consensus on variable-importance over bootstraps. Second, using these features in the training cohort, support vector machines (SVM) with a polynomial kernel was trained with repeated 5-fold cross validation to choose a model with highest area under the receiver operating characteristic curve (AUC) which was then used to predict ischemic events in the validation cohort. Results: The primary endpoint occurred in 2.4% of patients. Using 89 clinical and genetic variables in the random-forest feature selection step, variable importance-based consensus identified 50 top predictors. These 50 predictors achieved AUC=0.703 in the SVM training cohort and AUC=0.744 (specificity=0.84; sensitivity =0.43) in the SVM validation cohort. Conclusions: Machine learning methods using baseline clinical and genetic data enables prediction of infrequent ischemic events in post PCI patients receiving new generation stents. The role of such artificial intelligence algorithms in individualization of DAPT by identifying high or low risk patients needs to be explored.

Background The treatment of depression in children and adolescents is a substantial public health challenge. This study examined artificial intelligence tools for the prediction of early outcomes in depressed children and adolescents treated with fluoxetine, duloxetine, or placebo. Methods The study samples included training datasets ( N = 271) from patients with major depressive disorder (MDD) treated with fluoxetine and testing datasets from patients with MDD treated with duloxetine ( N = 255) or placebo ( N = 265). Treatment trajectories were generated using probabilistic graphical models (PGMs). Unsupervised machine learning identified specific depressive symptom profiles and related thresholds of improvement during acute treatment. Results Variation in six depressive symptoms (difficulty having fun, social withdrawal, excessive fatigue, irritability, low self‐esteem, and depressed feelings) assessed with the Children’s Depression Rating Scale‐Revised at 4–6 weeks predicted treatment outcomes with fluoxetine at 10–12 weeks with an average accuracy of 73% in the training dataset. The same six symptoms predicted 10–12 week outcomes at 4–6 weeks in (a) duloxetine testing datasets with an average accuracy of 76% and (b) placebo‐treated patients with accuracies of 67%. In placebo‐treated patients, the accuracies of predicting response and remission were similar to antidepressants. Accuracies for predicting nonresponse to placebo treatment were significantly lower than antidepressants. Conclusions PGMs provided clinically meaningful predictions in samples of depressed children and adolescents treated with fluoxetine or duloxetine. Future work should augment PGMs with biological data for refined predictions to guide the selection of pharmacological and psychotherapeutic treatment in children and adolescents with depression.

Age at depressive onset (AAO) corresponds to unique symptomatology and clinical outcomes. Integration of genome-wide association study (GWAS) results with additional “omic” measures to evaluate AAO has not been reported and may reveal novel markers of susceptibility and/or resistance to major depressive disorder (MDD). To address this gap, we integrated genomics with metabolomics using data-driven network analysis to characterize and differentiate MDD based on AAO. This study first performed two GWAS for AAO as a continuous trait in (a) 486 adults from the Pharmacogenomic Research Network-Antidepressant Medication Pharmacogenomic Study (PGRN-AMPS), and (b) 295 adults from the Combining Medications to Enhance Depression Outcomes (CO-MED) study. Variants from top signals were integrated with 153 p180-assayed metabolites to establish multi-omics network characterizations of early (<age 18) and adult-onset depression. The most significant variant (p = 8.77 × 10−8) localized to an intron of SAMD3. In silico functional annotation of top signals (p < 1 × 10−5) demonstrated gene expression enrichment in the brain and during embryonic development. Network analysis identified differential associations between four variants (in/near INTU, FAT1, CNTN6, and TM9SF2) and plasma metabolites (phosphatidylcholines, carnitines, biogenic amines, and amino acids) in early- compared with adult-onset MDD. Multi-omics integration identified differential biosignatures of early- and adult-onset MDD. These biosignatures call for future studies to follow participants from childhood through adulthood and collect repeated -omics and neuroimaging measures to validate and deeply characterize the biomarkers of susceptibility and/or resistance to MDD development.

Objective To test the ability of machine learning (ML) approaches with clinical and genomic biomarkers to predict methotrexate treatment response in patients with early rheumatoid arthritis (RA). Methods Demographic, clinical, and genomic data from 643 patients of European ancestry with early RA (mean age 54 years; 70% female) subdivided into a training (n = 336) and validation cohort (n = 307) were used. The genomic data comprised 160 single‐nucleotide polymorphisms (SNPs) previously associated with RA or methotrexate metabolism. Response to methotrexate monotherapy was defined as good or moderate by the European Alliance of Associations for Rheumatology (EULAR) response criteria at the 3‐month follow‐up. Supervised ML methods were trained with 5 repeats and 10‐fold cross‐validation using the training cohort. Prediction performance was validated in the independent validation cohort. Results Supervised ML methods combining age, sex, smoking, rheumatoid factor, baseline Disease Activity Score in 28 joints (DAS28) scores and 160 SNPs predicted EULAR response at 3 months with the area under the receiver operating curve of 0.84 ( P = 0.05) in the training cohort and achieved a prediction accuracy of 76% ( P = 0.05) in the validation cohort (sensitivity 72%, specificity 77%). Intergenic SNPs rs12446816, rs13385025, rs113798271, and ATIC (rs2372536) had variable importance above 60.0 and along with baseline DAS28 scores were among the top predictors of methotrexate response. Conclusion Pharmacogenomic biomarkers combined with baseline DAS28 scores can be useful in predicting response to methotrexate in patients with early RA. Applying ML to predict treatment response holds promise for guiding effective RA treatment choices, including timely escalation of RA therapies.

We set out to determine whether machine learning–based algorithms that included functionally validated pharmacogenomic biomarkers joined with clinical measures could predict selective serotonin reuptake inhibitor (SSRI) remission/response in patients with major depressive disorder (MDD). We studied 1,030 white outpatients with MDD treated with citalopram/escitalopram in the Mayo Clinic Pharmacogenomics Research Network Antidepressant Medication Pharmacogenomic Study (PGRN‐AMPS; n = 398), Sequenced Treatment Alternatives to Relieve Depression (STAR*D; n = 467), and International SSRI Pharmacogenomics Consortium (ISPC; n = 165) trials. A genomewide association study for PGRN‐AMPS plasma metabolites associated with SSRI response (serotonin) and baseline MDD severity (kynurenine) identified single nucleotide polymorphisms (SNPs) in DEFB1, ERICH3, AHR, and TSPAN5 that we tested as predictors. Supervised machine‐learning methods trained using SNPs and total baseline depression scores predicted remission and response at 8 weeks with area under the receiver operating curve (AUC) > 0.7 (P < 0.04) in PGRN‐AMPS patients, with comparable prediction accuracies > 69% (P ≤ 0.07) in STAR*D and ISPC. These results demonstrate that machine learning can achieve accurate and, importantly, replicable prediction of SSRI therapy response using total baseline depression severity combined with pharmacogenomic biomarkers.