Gary Livingston

Predicting dire outcomes of patients with community acquired pneumonia

Community-acquired pneumonia (CAP) is an important clinical condition with regard to patient mortality, patient morbidity, and healthcare resource utilization. The assessment of the likely clinical course of a CAP patient can significantly influence decision making about whether to treat the patient as an inpatient or as an outpatient. That decision can in turn influence resource utilization, as well as patient well being. Predicting dire outcomes, such as mortality or severe clinical complications, is a particularly important component in assessing the clinical course of patients. We used a training set of 1601 CAP patient cases to construct 11 statistical and machine-learning models that predict dire outcomes. We evaluated the resulting models on 686 additional CAP-patient cases. The primary goal was not to compare these learning algorithms as a study end point; rather, it was to develop the best model possible to predict dire outcomes. A special version of an artificial neural network (NN) model predicted dire outcomes the best. Using the 686 test cases, we estimated the expected healthcare quality and cost impact of applying the NN model in practice. The particular, quantitative results of this analysis are based on a number of assumptions that we make explicit; they will require further study and validation. Nonetheless, the general implication of the analysis seems robust, namely, that even small improvements in predictive performance for prevalent and costly diseases, such as CAP, are likely to result in significant improvements in the quality and efficiency of healthcare delivery. Therefore, seeking models with the highest possible level of predictive performance is important. Consequently, seeking ever better machine-learning and statistical modeling methods is of great practical significance.

Transcriptional Networks Inferred from Molecular Signatures of Breast Cancer

Global genomic approaches in cancer research have provided new and innovative strategies for the identification of signatures that differentiate various types of human cancers. Computational analysis of the promoter composition of the genes within these signatures may provide a powerful method for deducing the regulatory transcriptional networks that mediate their collective function. In this study we have systematically analyzed the promoter composition of gene classes derived from previously established genetic signatures that recently have been shown to reliably and reproducibly distinguish five molecular subtypes of breast cancer associated with distinct clinical outcomes. Inferences made from the trends of transcription factor binding site enrichment in the promoters of these gene groups led to the identification of regulatory pathways that implicate discrete transcriptional networks associated with specific molecular subtypes of breast cancer. One of these inferred pathways predicted a role for nuclear factor-kappaB in a novel feed-forward, self-amplifying, autoregulatory module regulated by the ERBB family of growth factor receptors. The existence of this pathway was verified in vivo by chromatin immunoprecipitation and shown to be deregulated in breast cancer cells overexpressing ERBB2. This analysis indicates that approaches of this type can provide unique insights into the differential regulatory molecular programs associated with breast cancer and will aid in identifying specific transcriptional networks and pathways as potential targets for tumor subtype-specific therapeutic intervention.

Closing the loop: an agenda- and justification-based framework for selecting the next discovery task to perform

We propose and evaluate an agenda- and justification-based architecture for discovery systems that selects the next tasks to perform. This framework has many desirable properties: (1) it facilitates the encoding of general discovery strategies using a variety of background knowledge, (2) it reasons about the appropriateness of the tasks being considered, and (3) it tailors its behavior toward a users interests. A prototype discovery program called HAMB demonstrates that both reasons and estimates of interestingness contribute to performance in the domains of protein crystallization and patient rehabilitation.

Closing the loop: heuristics for autonomous discovery

Autonomous discovery systems will be able to peruse very large databases more thoroughly than people can. In a companion paper by G.R. Livingston et al. (see ibid., p.385-92, 2001), we describe a general framework for autonomous systems. We present and evaluate heuristics for use in this framework. Although these heuristics were designed for a prototype system, we believe they provide good initial solutions to problems encountered when implementing fully autonomous discovery systems. As such, these heuristics may be used as the starting point for future research into fully autonomous discovery systems.

Machine‐learning techniques for macromolecular crystallization data

Systematizing belief systems regarding macromolecular crystallization has two major advantages: automation and clarification. In this paper, methodologies are presented for systematizing and representing knowledge about the chemical and physical properties of additives used in crystallization experiments. A novel autonomous discovery program is introduced as a method to prune rule-based models produced from crystallization data augmented with such knowledge. Computational experiments indicate that such a system can retain and present informative rules pertaining to protein crystallization that warrant further confirmation via experimental techniques.

What's new? Using prior models as a measure of novelty in knowledge discovery

One of the challenges of knowledge discovery is identifying patterns that are interesting, with novelty an important component of interestingness. Another important aspect of knowledge discovery is making efficient use of background knowledge. This paper develops a definition of novelty relative to a prior model of the domain. The definition of novelty is tested using pneumonia outcome data and a prior model of pneumonia severity.

Toward a Computational Model of Hypothesis Formation and Model Building in Science

We present our two-part approach to computational scientific discovery. The first part is a heuristic search through the space of hypotheses to find good hytheses. The second stage is a heuristic search through the space of sets of hypotheses to find coherent models. Both of these heuristic searches strive for four criteria: empirical support, simplicity, novelty and utility. Four types of domain knowledge inform both searches: transforms, taxonomies, sequences and default models. We mention several experimental results that suggest the viability of our approach.

Towards predicting protein-protein interactions in novel organisms.

Machine learning methods are often used to predict Protein-Protein Interactions (PPI). It is common to develop methods using known PPI from well-characterised reference organisms, drawing from that organism data for inferring a predictive model and evaluating the model. We present evidence that this practice does not give a meaningful indication of the models performance on genetically distinct organisms. We conclude that this practice cannot be applied to proteins inferred from the genetic sequence of a novel organism for which no PPI data is available, and that there is need for evaluating such methods on organisms distinct from their training organisms.