Framework for the Use of External Controls to Evaluate Treatment Outcomes in Precision Oncology Trials.

Advances in genomics have enabled anticancer therapies to be tailored to target specific genomic alterations. Single-arm trials (SATs), including those incorporated within umbrella, basket, and platform trials, are widely adopted when it is not feasible to conduct randomized controlled trials in rare biomarker-defined subpopulations. External controls (ECs), defined as control arm data derived outside the clinical trial, have gained renewed interest as a strategy to supplement evidence generated from SATs to allow comparative analysis. There are increasing examples demonstrating the application of EC in precision oncology trials. The prospective application of EC in conducting comparative studies is associated with distinct methodological challenges, the specific considerations for EC use in biomarker-defined subpopulations have not been adequately discussed, and a formal framework is yet to be established. In this review, we present a framework for conducting a prospective comparative analysis using EC. Key steps are (1) defining the purpose of using EC to address the study question, (2) determining if the external data are fit for purpose, (3) developing a transparent study protocol and a statistical analysis plan, and (iv) interpreting results and drawing conclusions on the basis of a prespecified hypothesis. We specify the considerations required for the biomarker-defined subpopulations, which include (1) specifying the comparator and biomarker status of the comparator group, (2) defining lines of treatment, (3) assessment of the biomarker testing panels used, and (4) assessment of cohort stratification in tumor-agnostic studies. We further discuss novel clinical trial designs and statistical techniques leveraging EC to propose future directions to advance evidence generation and facilitate drug development in precision oncology.

Development and validation of a deep learning model for improving detection of nonmelanoma skin cancers treated with Mohs micrographic surgery.

Background: Real-time review of frozen sections underpins the quality of Mohs surgery. There is an unmet need for low-cost techniques that can improve Mohs surgery by reliably corroborating cancerous regions of interest and surgical margin proximity. Objective: To test that deep learning models can identify nonmelanoma skin cancer regions in Mohs frozen section specimens. Methods: Deep learning models were developed on archival images of focused microscopic views (FMVs) containing regions of annotated, invasive nonmelanoma skin cancer between 2015 and 2018, then validated on prospectively collected images in a temporal cohort (2019-2021). Results: The tile-based classification models were derived using 1423 focused microscopic view images from 154 patients and tested on 374 images from 66 patients. The best models detected basal cell carcinomas with a median average precision of 0.966 and median area under the receiver operating curve of 0.889 at 100x magnification (0.943 and 0.922 at 40x magnification). For invasive squamous cell carcinomas, high median average precision of 0.904 was achieved at 100x magnification. Limitations: Single institution study with limited cases of squamous cell carcinoma and rare nonmelanoma skin cancer. Conclusion: Deep learning appears highly accurate for detecting skin cancers in Mohs frozen sections, supporting its potential for enhancing surgical margin control and increasing operational efficiency.

Synonymous alterations of cancer-associated Trp53 CpG mutational hotspots cause fatal developmental jaw malocclusions but no tumors in knock-in mice.

Intragenic CpG dinucleotides are tightly conserved in evolution yet are also vulnerable to methylation-dependent mutation, raising the question as to why these functionally critical sites have not been deselected by more stable coding sequences. We previously showed in cell lines that altered exonic CpG methylation can modify promoter start sites, and hence protein isoform expression, for the human TP53 tumor suppressor gene. Here we extend this work to the in vivo setting by testing whether synonymous germline modifications of exonic CpG sites affect murine development, fertility, longevity, or cancer incidence. We substituted the DNA-binding exons 5-8 of Trp53, the mouse ortholog of human TP53, with variant-CpG (either CpG-depleted or -enriched) sequences predicted to encode the normal p53 amino acid sequence; a control construct was also created in which all non-CpG sites were synonymously substituted. Homozygous Trp53-null mice were the only genotype to develop tumors. Mice with variant-CpG Trp53 sequences remained tumor-free, but were uniquely prone to dental anomalies causing jaw malocclusion (p < .0001). Since the latter phenotype also characterises murine Rett syndrome due to dysfunction of the trans-repressive MeCP2 methyl-CpG-binding protein, we hypothesise that CpG sites may exert non-coding phenotypic effects via pre-translational cis-interactions of 5-methylcytosine with methyl-binding proteins which regulate mRNA transcript initiation, expression or splicing, although direct effects on mRNA structure or translation are also possible.

Computational prediction of multidisciplinary team decision-making for adjuvant breast cancer drug therapies: a machine learning approach.

BACKGROUND: Multidisciplinary team (MDT) meetings are used to optimise expert decision-making about treatment options, but such expertise is not digitally transferable between centres. To help standardise medical decision-making, we developed a machine learning model designed to predict MDT decisions about adjuvant breast cancer treatments. METHODS: We analysed MDT decisions regarding adjuvant systemic therapy for 1065 breast cancer cases over eight years. Machine learning classifiers with and without bootstrap aggregation were correlated with MDT decisions (recommended, not recommended, or discussable) regarding adjuvant cytotoxic, endocrine and biologic/targeted therapies, then tested for predictability using stratified ten-fold cross-validations. The predictions so derived were duly compared with those based on published (ESMO and NCCN) cancer guidelines. RESULTS: Machine learning more accurately predicted adjuvant chemotherapy MDT decisions than did simple application of guidelines. No differences were found between MDT- vs. ESMO/NCCN- based decisions to prescribe either adjuvant endocrine (97%, p = 0.44/0.74) or biologic/targeted therapies (98%, p = 0.82/0.59). In contrast, significant discrepancies were evident between MDT- and guideline-based decisions to prescribe chemotherapy (87%, p < 0.01, representing 43% and 53% variations from ESMO/NCCN guidelines, respectively). Using ten-fold cross-validation, the best classifiers achieved areas under the receiver operating characteristic curve (AUC) of 0.940 for chemotherapy (95% C.I., 0.922-0.958), 0.899 for the endocrine therapy (95% C.I., 0.880-0.918), and 0.977 for trastuzumab therapy (95% C.I., 0.955-0.999) respectively. Overall, bootstrap aggregated classifiers performed better among all evaluated machine learning models. CONCLUSIONS: A machine learning approach based on clinicopathologic characteristics can predict MDT decisions about adjuvant breast cancer drug therapies. The discrepancy between MDT- and guideline-based decisions regarding adjuvant chemotherapy implies that certain non-clincopathologic criteria, such as patient preference and resource availability, are factored into clinical decision-making by local experts but not captured by guidelines.

Gene-disease association with literature based enrichment.

MOTIVATION: Gene set enrichment analysis (GSEA) annotates gene microarray data with functional information from the biomedical literature to improve gene-disease association prediction. We hypothesize that supplementing GSEA with comprehensive gene function catalogs built automatically using information extracted from the scientific literature will significantly enhance GSEA prediction quality. METHODS: Gold standard gene sets for breast cancer (BrCa) and colorectal cancer (CRC) were derived from the literature. Two gene function catalogs (CMeSH and CUMLS) were automatically generated. 1. By using Entrez Gene to associate all recorded human genes with PubMed article IDs. 2. Using the genes mentioned in each PubMed article and associating each with the article's MeSH terms (in CMeSH) and extracted UMLS concepts (in CUMLS). Microarray data from the Gene Expression Omnibus for BrCa and CRC was then annotated using CMeSH and CUMLS and for comparison, also with several pre-existing catalogs (C2, C4 and C5 from the Molecular Signatures Database). Ranking was done using, a standard GSEA implementation (GSEA-p). Gene function predictions for enriched array data were evaluated against the gold standard by measuring area under the receiver operating characteristic curve (AUC). RESULTS: Comparison of ranking using the literature enrichment catalogs, the pre-existing catalogs as well as five randomly generated catalogs show the literature derived enrichment catalogs are more effective. The AUC for BrCa using the unenriched gene expression dataset was 0.43, increasing to 0.89 after gene set enrichment with CUMLS. The AUC for CRC using the unenriched gene expression dataset was 0.54, increasing to 0.9 after enrichment with CMeSH. C2 increased AUC (BrCa 0.76, CRC 0.71) but C4 and C5 performed poorly (between 0.35 and 0.5). The randomly generated catalogs also performed poorly, equivalent to random guessing. DISCUSSION: Gene set enrichment significantly improved prediction of gene-disease association. Selection of enrichment catalog had a substantial effect on prediction accuracy. The literature based catalogs performed better than the MSigDB catalogs, possibly because they are more recent. Catalogs generated automatically from the literature can be kept up to date. CONCLUSION: Prediction of gene-disease association is a fundamental task in biomedical research. GSEA provides a promising method when using literature-based enrichment catalogs. AVAILABILITY: The literature based catalogs generated and used in this study are available from http://www2.chi.unsw.edu.au/literature-enrichment.

BICEPP: an example-based statistical text mining method for predicting the binary characteristics of drugs.

BACKGROUND: The identification of drug characteristics is a clinically important task, but it requires much expert knowledge and consumes substantial resources. We have developed a statistical text-mining approach (BInary Characteristics Extractor and biomedical Properties Predictor: BICEPP) to help experts screen drugs that may have important clinical characteristics of interest. RESULTS: BICEPP first retrieves MEDLINE abstracts containing drug names, then selects tokens that best predict the list of drugs which represents the characteristic of interest. Machine learning is then used to classify drugs using a document frequency-based measure. Evaluation experiments were performed to validate BICEPP's performance on 484 characteristics of 857 drugs, identified from the Australian Medicines Handbook (AMH) and the PharmacoKinetic Interaction Screening (PKIS) database. Stratified cross-validations revealed that BICEPP was able to classify drugs into all 20 major therapeutic classes (100%) and 157 (of 197) minor drug classes (80%) with areas under the receiver operating characteristic curve (AUC) > 0.80. Similarly, AUC > 0.80 could be obtained in the classification of 173 (of 238) adverse events (73%), up to 12 (of 15) groups of clinically significant cytochrome P450 enzyme (CYP) inducers or inhibitors (80%), and up to 11 (of 14) groups of narrow therapeutic index drugs (79%). Interestingly, it was observed that the keywords used to describe a drug characteristic were not necessarily the most predictive ones for the classification task. CONCLUSIONS: BICEPP has sufficient classification power to automatically distinguish a wide range of clinical properties of drugs. This may be used in pharmacovigilance applications to assist with rapid screening of large drug databases to identify important characteristics for further evaluation.

Perpetrators of pharmacokinetic drug-drug interactions arising from altered cytochrome P450 activity: a criteria-based assessment.

AIMS: To catalogue the perpetrators of CYP-mediated pharmacokinetic drug-drug interactions (PK-DDIs) using clinically relevant criteria, and to compare this with an analogous catalogue. METHODS: Candidate inhibitors and inducers of CYP1A2, CYP2C9, CYP2C19, CYP2D6 and CYP3A ('perpetrators') were evaluated using published clinical pharmacokinetic interaction studies. Studies were selected on the basis of ≥six human subjects, use of a validated in vivo probe substrate for the CYP enzyme, and clinically relevant dosing. Inhibitors were described according to the FDA classifications of strong, moderate or weak, whereas inducers were classified as major (≥twofold decrease in AUC) or weak (

In silico prioritisation of candidate genes for prokaryotic gene function discovery: an application of phylogenetic profiles.

BACKGROUND: In silico candidate gene prioritisation (CGP) aids the discovery of gene functions by ranking genes according to an objective relevance score. While several CGP methods have been described for identifying human disease genes, corresponding methods for prokaryotic gene function discovery are lacking. Here we present two prokaryotic CGP methods, based on phylogenetic profiles, to assist with this task. RESULTS: Using gene occurrence patterns in sample genomes, we developed two CGP methods (statistical and inductive CGP) to assist with the discovery of bacterial gene functions. Statistical CGP exploits the differences in gene frequency against phenotypic groups, while inductive CGP applies supervised machine learning to identify gene occurrence pattern across genomes. Three rediscovery experiments were designed to evaluate the CGP frameworks. The first experiment attempted to rediscover peptidoglycan genes with 417 published genome sequences. Both CGP methods achieved best areas under receiver operating characteristic curve (AUC) of 0.911 in Escherichia coli K-12 (EC-K12) and 0.978 Streptococcus agalactiae 2603 (SA-2603) genomes, with an average improvement in precision of >3.2-fold and a maximum of >27-fold using statistical CGP. A median AUC of >0.95 could still be achieved with as few as 10 genome examples in each group of genome examples in the rediscovery of the peptidoglycan metabolism genes. In the second experiment, a maximum of 109-fold improvement in precision was achieved in the rediscovery of anaerobic fermentation genes in EC-K12. The last experiment attempted to rediscover genes from 31 metabolic pathways in SA-2603, where 14 pathways achieved AUC >0.9 and 28 pathways achieved AUC >0.8 with the best inductive CGP algorithms. CONCLUSION: Our results demonstrate that the two CGP methods can assist with the study of functionally uncategorised genomic regions and discovery of bacterial gene-function relationships. Our rediscovery experiments also provide a set of standard tasks against which future methods may be compared.