This paper introduces an open-source implementation of a non-monotonic learning method called XHAIL and shows how it can be applied to a whole-organism model of yeast metabolism. Unlike the previous XHAIL prototype, which could only handle networks of a few dozen reactions, our new system can accommodate networks with over a thousand. This scale–up was achieved though several implementation improvements which increase the method’s efficiency and support an enhanced language bias that further increases its usability. We test the system in a case study using real data collected by a Robot Scientist.
The 3rd Intl. Workshop on NL-based Software Engineering
ILP 2014 - Nonmonotonic Learning in Large Biological Network
1. Nonmonotonic Learning
in Large Biological Networks
Stefano Bragaglia, Oliver Ray
stefano.bragaglia@bristol.ac.uk, oray@cs.bris.ac.uk
Department of Computer Science
University of Bristol
14-17/09/2014 ILP '14, Nancy 1
2. • New open source XHAIL implementation
• Study of scalability on Biological Networks
• Found mistake in genome-scale network
14-17/09/2014 ILP '14, Nancy 2
3. XHAIL
• Nonmonotonic ILP
– normal (extended) logic programs
– abductive, deductive, inductive
• Prototype (2006)
– ASP-based (lparse/smodels)
– Prolog wrapper (SWI-Prolog)
– Not defeasible
– Not
• Metabolic Network Revision (ILP 09)
– AAA model (~30 reactions)
– Pathway-specific model
• Non
– normal
– Abductive
• Current application
– ASP-based (gringo/clasp)
– Java wrapper (Java 8)
– Defeasible (language bias)
– Open source
• Metabolic Network Revision
– ABER model (1100+ reactions)
– Whole-organism model
14-17/09/2014 ILP '14, Nancy 3
4. XHAIL
• Nonmonotonic ILP
– normal (extended) logic programs
– abductive, deductive, inductive
• Prototype (2006)
– ASP-based (lparse/smodels)
– Prolog wrapper (SWI-Prolog)
– Not defeasible
– Not
• Metabolic Network Revision (ILP 09)
– AAA model (~30 reactions)
– Pathway-specific model
• Non
– normal
– Abductive
• Current application
– ASP-based (gringo/clasp)
– Java wrapper (Java 8)
– Defeasible (language bias)
– Open source
http://github.com/cathexis-bris-ac-uk/XHAIL
• Metabolic Network Revision
– ABER model (1100+ reactions)
– Whole-organism model
14-17/09/2014 ILP '14, Nancy 4
12. Scalability Analysis
on Validation Experiments
standard
expressions
14-17/09/2014 ILP '14, Nancy 12
biased
expressions
# reactions hypotheses time (s) - means "out of memory"
14. NONMONOTONIC LEARNING
IN LARGE BIOLOGICAL NETWORKS
Stefano Bragaglia, Oliver Ray
stefano.bragaglia@bristol.ac.uk, oray@cs.bris.ac.uk
• Thanks for your attention
• Any questions?
14-17/09/2014 ILP '14, Nancy 14
Editor's Notes
Requirements Engineering (FLTL, ILP’06)
Deals with negation
Completion
Ross King
Fluent Linear Temporal Logic
Requirements Engineering (FLTL, ILP’06)
Deals with negation
Completion
Ross King
Fluent Linear Temporal Logic
Help the user to explore the part of the search space they want to focus on
We prepared a first set of tests to validate and revise the portion of ABER relating to AAA. These tests are divided into tasks.
Task A consists of two experiments that allow XHAIL to infer that YER090W is required in all enzyme complexes catalysing reaction 4.1.3.27.
Experiment 1 simply knocks out YER090W. ABER observes growth as it thinks YKL211C can catalyse 4.1.3.27 alone.
We prepared a first set of tests to validate and revise the portion of ABER relating to AAA. These tests are divided into tasks.
Task A consists of two experiments that allow XHAIL to infer that YER090W is required in all enzyme complexes catalysing reaction 4.1.3.27.
Experiment 1 simply knocks out YER090W. ABER observes growth as it thinks YKL211C can catalyse 4.1.3.27 alone.
We prepared a first set of tests to validate and revise the portion of ABER relating to AAA. These tests are divided into tasks.
Task A consists of two experiments that allow XHAIL to infer that YER090W is required in all enzyme complexes catalysing reaction 4.1.3.27.
Experiment 1 simply knocks out YER090W. ABER observes growth as it thinks YKL211C can catalyse 4.1.3.27 alone.
We prepared a first set of tests to validate and revise the portion of ABER relating to AAA. These tests are divided into tasks.
Task A consists of two experiments that allow XHAIL to infer that YER090W is required in all enzyme complexes catalysing reaction 4.1.3.27.
Experiment 1 simply knocks out YER090W. ABER observes growth as it thinks YKL211C can catalyse 4.1.3.27 alone.
We prepared a first set of tests to validate and revise the portion of ABER relating to AAA. These tests are divided into tasks.
Task A consists of two experiments that allow XHAIL to infer that YER090W is required in all enzyme complexes catalysing reaction 4.1.3.27.
Experiment 1 simply knocks out YER090W. ABER observes growth as it thinks YKL211C can catalyse 4.1.3.27 alone.
We have also prepared a second batch of 40 experiments using real data provided by a robot scientist.
The conventions are the same as before but we only compare the results in the case of the hardest problems.
We observe a gain in performance of about 30%.