We are working on an experiment planning system for synthetic biology. As part of this effort, we wish to identify the most informative experiments to perform, based on our current state of knowledge about some design or designs. To do so, we introduce a hierarchical Bayesian model of genetic circuits, assess its predictive adequacy, and explain how it is used to score candidate experiments. We conclude with a discussion of future research directions, including model extensions, improvements to the inference methods, and more ambitious uses of the predictive model.

Organisations conducting research programs often focus the work of their scientists and technologists on challenge problems (CPs). These challenges are designed to ensure that progress is measurable and relevant to the goals of the program sponsor. Generating and selecting pertinent CPs is difficult, as is assessing their value. We describe a method of generating and selecting CPs and its application in a highly collaborative, multi-organisation research program. Thirty-eight biologists, chemists, mathematicians and computer scientists across academic, commercial and government organisations generated and ranked their top choices from among 12 richly described candidate challenge problems. A ranked-choice voting formula was applied. Five CPs were highly scored; the remaining seven were distributed across a lower range of scores. The program sponsor subsequently directed researchers to address six CPs, including the elected five. Analysis of the rationales that participants offered for their CP rankings revealed four domain-independent dimensions of value: capability, speed, impact and synergy. These dimensions of value can help managers of interdisciplinary research programs systematically select a portfolio of CPs that will efficiently apply utilise resources towards program goals and facilitate measurement of scientific progress.

Experimental protocols are typically represented in either a natural language that is hard to replicate or compare, or in procedural languages that are difficult to automatically synthesize, detach from a specific experimental design for reuse, or analyze. We introduce a new approach based on techniques from automated planning. We describe how to represent transformation operators that manipulate samples in terms of applying conditions to samples. We define the semantics of this representation. We also present a simplified version of the notation that removes much of the modeling burden required of scientists. The resulting representation supports automated planning, provides sample provenance and metadata tracking at no cost by virtue of a plan’s causal structure, and separates protocol specification from experimental design.

This paper describes R3 ( Reading, Reasoning, and Reporting ), our system for deep language understanding and extension of mechanistic models. The overall purpose of R3 is to read about biochemical signaling pathways from PubMed Central journal articles and integrate information into its model. Its initial background model of these biochemical pathways is derived from an imported Reactome model of biological pathways, events, complexes, and proteins. We describe some significant issues for semantic parsing in this domain and how R3 uses pre-and post-analysis reasoning to bridge the differences between the semantic information that can be derived from a text and the codified mechanistic information in the curated biomedical database. We also present extensions to relational structure mapping to detect corroboration between the semantic parse and the model and extend the model via analogical inference from the parse. We close with a description of empirical results with R3, including semantic parsing, model extension, grounding entity and event references, and modeling entity behavior using knowledge learned by reading.

We describe preliminary work on XPlan, a system for experiment planning in synthetic biology. In synthetic biology, as in other emerging fields, scientific exploration and engineering design must be interleaved, because of uncertainty about the underlying mechanisms. Through its experiment planning, XPlan provides a coordinating linchpin in DARPA’s Synergistic Discovery and Design (SD2) platform to automate scientific discovery, closing the loop between multiple machine learning analysis and biological design tools and wet labs to guide the discovery and design process. To accomplish this, XPlan combines design of experiments techniques with hierarchical planning, based on the Shop2 planner, to develop experimental plans that are directly executable in highly automated wet labs and to project experimental costs. In particular, XPlan formulates experimental designs and translates them into goals representing biological samples, then uses Shop2 to plan construction and measurement of samples using available laboratory resources. In ongoing work, we are developing probability models that will support value of information computations to optimize experimental plans.

In the field of synthetic biology, as with other emerging fields of engineering, scientific exploration and engineering design are intimately entwined. Unlike established fields of engineering, synthetic biology has only highly uncertain and incomplete mechanistic models. As a result, engineering synthetic biological systems is an incremental process in which the production of designs is closely interleaved with execution of experiments to assess the success of those designs and data analysis to identify factors and mechanisms responsible for design successes and failures. In this demonstration, we show how a Hierarchical Task Network (HTN) planning system automates scientific discovery, closing the loop betweenmultiple machine learning analysis and biological design tools and wet labs to guide the discovery and design process. 1 Research Problem and Motivations Organization and planning of synthetic biology experiments is currently done almost entirely by hand. Several ongoing developments, however, are rapidly increasing the need for automation assistance in experiment planning. More and more laboratory automation is becoming available, increasing the scale and complexity of experiments that can be performed. Automation and information technology are supporting new business models with laboratory work done by technicians or outsourced to a “lab for hire.” Finally, new “multiplexing” protocols allow many tests to be conducted on a single experimental sample, and multiple experimental samples to be processed in parallel. In all of these cases, the growth in scale and complexity are rapidly outstripping the abilities of humans to create detailed experimental plans and to hand-curate the relationships between those plans and the large collections of data they generate. Furthermore, experiments are still costly both in money and time, and the space a researcher wishes to explore is often much larger than the ∗This work was supported by the Air Force Research Laboratory (AFRL) and DARPA under contract FA875017CO184. This document does not contain technology or technical data controlled under either U.S. International Traffic in Arms Regulation or U.S. Export Administration Regulations. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the AFRL and DARPA. Decision-Theoretic Hierarchical Planner