Comparing the use of aggregate data and various methods of integrating individual patient data to network meta-analysis and its application to first-line ART

Background The 2018 World Health Organization HIV guidelines were based on the results of a network meta-analysis (NMA) of published trials. This study employed individual patient-level data (IPD) and aggregate data (AgD) and meta-regression methods to assess the evidence supporting the WHO recommendations and whether they needed any refinements. Methods Access to IPD from three trials was granted through ClinicalStudyDataRequest.com (CSDR). Seven modelling approaches were applied and compared: 1) Unadjusted AgD network meta-analysis (NMA) – the original analysis; 2) AgD-NMA with meta-regression; 3) Two-stage IPD-AgD NMA; 4) Unadjusted one-stage IPD-AgD NMA; 5) One-stage IPD-AgD NMA with meta-regression (one-stage approach); 6) Two-stage IPD-AgD NMA with empirical-priors (empirical-priors approach); 7) Hierarchical meta-regression IPD-AgD NMA (HMR approach). The first two were the models used previously. Models were compared with respect to effect estimates, changes in the effect estimates, coefficient estimates, DIC and model fit, rankings and between-study heterogeneity. Results IPD were available for 2160 patients, representing 6.5% of the evidence base and 3 of 24 edges. The aspect of the model affected by the choice of modeling appeared to differ across outcomes. HMR consistently generated larger intervals, often with credible intervals (CrI) containing the null value. Discontinuations due to adverse events and viral suppression at 96 weeks were the only two outcomes for which the unadjusted AgD NMA would not be selected. For the first, the selected model shifted the principal comparison of interest from an odds ratio of 0.28 (95% CrI: 10.17, 0.44) to 0.37 (95% CrI: 0.23, 0.58). Throughout all outcomes, the regression estimates differed substantially between AgD and IPD methods, with the latter being more often larger in magnitude and statistically significant. Conclusions Overall, the use of IPD often impacted the coefficient estimates, but not sufficiently as to necessitate altering the final recommendations of the 2018 WHO Guidelines. Future work should examine the features of a network where adjustments will have an impact, such as how much IPD is required in a given size of network.

Hierarchical Bayesian Models of Reinforcement Learning: Introduction and comparison to alternative methods

Reinforcement learning models have been used extensively and with great success to capture learning and decision-making processes in humans and other organisms. One essential goal of these computational models is generalization to new sets of observations. Extracting parameters that can reliably predict out-of-sample data can be difficult, however: reinforcement learning models often face problems of non-identifiability, which can lead to poor predictive accuracy. The use of prior distributions to regularize parameter estimates can be an effective way to remedy this issue. While previous research has suggested that empirical priors estimated from a separate dataset improve identifiability and predictive accuracy, this paper outlines an alternate method for the derivation of empirical priors: hierarchical Bayesian modeling. We provide a detailed introduction to this method, and show that using hierarchical models to simultaneously extract and impose empirical priors leads to better out-of-sample prediction while being more data efficient.

A Bayesian framework for inter-cellular information sharing improves dscRNA-seq quantification

Motivation Droplet-based single-cell RNA-seq (dscRNA-seq) data are being generated at an unprecedented pace, and the accurate estimation of gene-level abundances for each cell is a crucial first step in most dscRNA-seq analyses. When pre-processing the raw dscRNA-seq data to generate a count matrix, care must be taken to account for the potentially large number of multi-mapping locations per read. The sparsity of dscRNA-seq data, and the strong 3’ sampling bias, makes it difficult to disambiguate cases where there is no uniquely mapping read to any of the candidate target genes. Results We introduce a Bayesian framework for information sharing across cells within a sample, or across multiple modalities of data using the same sample, to improve gene quantification estimates for dscRNA-seq data. We use an anchor-based approach to connect cells with similar gene-expression patterns, and learn informative, empirical priors which we provide to alevin’s gene multi-mapping resolution algorithm. This improves the quantification estimates for genes with no uniquely mapping reads (i.e. when there is no unique intra-cellular information). We show our new model improves the per cell gene-level estimates and provides a principled framework for information sharing across multiple modalities. We test our method on a combination of simulated and real datasets under various setups. Availability and implementation The information sharing model is included in alevin and is implemented in C++14. It is available as open-source software, under GPL v3, at https://github.com/COMBINE-lab/salmon as of version 1.1.0.

A Bayesian framework for inter-cellular information sharing improves dscRNA-seq quantification

MotivationDroplet based single cell RNA-seq (dscRNA-seq) data is being generated at an unprecedented pace, and the accurate estimation of gene level abundances for each cell is a crucial first step in most dscRNA-seq analyses. When preprocessing the raw dscRNA-seq data to generate a count matrix, care must be taken to account for the potentially large number of multi-mapping locations per read. The sparsity of dscRNA-seq data, and the strong 3’ sampling bias, makes it difficult to disambiguate cases where there is no uniquely mapping read to any of the candidate target genes.ResultsWe introduce a Bayesian framework for information sharing across cells within a sample, or across multiple modalities of data using the same sample, to improve gene quantification estimates for dscRNA-seq data. We use an anchor-based approach to connect cells with similar gene expression patterns, and learn informative, empirical priors which we provide to alevin’s gene multi-mapping resolution algorithm. This improves the quantification estimates for genes with no uniquely mapping reads (i.e. when there is no unique intra-cellular information). We show our new model improves the per cell gene level estimates and provides a principled framework for information sharing across multiple modalities. We test our method on a combination of simulated and real datasets under various setups.AvailabilityThe information sharing model is included in alevin and is implemented in C++14. It is available as open-source software, under GPL v3, at https://github.com/COMBINE-lab/salmon as of version [email protected], [email protected]

The radically embodied conscious cybernetic Bayesian brain: Towards explaining the emergence of agency

This manuscript attempts to characterize a broad range of intentional phenomena in terms of embodied self-models (ESMs), understood as body maps with agentic properties, functioning as predictive-memory systems and cybernetic controllers. ESMs may constitute a dominant organizing principle for neural architectures due to their initial and ongoing significance for the processes by which inference problems are solved in cognitive (and affective) development. Specifically, embodied experiences may provide a source of foundational lessons in learning curriculums in which agents explore increasingly challenging inference spaces along zones of proximal development, so helping to solve an unresolved problem in Bayesian cognitive science: what are biologically plausible mechanisms for equipping learners with sufficiently constraining/empowering inductive biases? Drawing on models from neurophysiology, psychology, and developmental robotics, I suggest a potentially surprising answer to how this problem might be solved: body maps are the primary source of (empirical) priors, or very reliably learnable posterior expectations. If ESMs play this kind of foundational role in bootstrapping cognitive development, then we ought to expect bidirectional linkages between all sensory modalities and frontal-parietal control hierarchies, so infusing all senses with somatic-motoric properties, thereby structuring all perception by relevant affordances, so solving frame problems for embodied learners/agents.

Hypotheses about the relationship of cognition with psychopathology should be tested by embedding them into empirical priors

Mechanistic hypotheses about psychiatric disorders are increasingly formalized as computational models of information-processing in the brain. Model parameters, characterizing for example decision-making biases, are hypothesized to correlate with clinical constructs. This is promising, but here we draw attention to some techniques used to minimize noise in parameter estimation which are in common use but may be unhelpful. Namely, the use of empirical priors that do not incorporate relationships between psychopathology and modelled processes will suppress the very relationships of interest. This is because the variability associated with psychopathology will be indistinguishable from that due to noise from the point of view of the hierarchical, or random-effects, fit that used the empirical priors in question. We advocate incorporating cross-domain, e.g. psychopathology-cognition relationships into the parameter inference itself.