Machine Learning a Probabilistic Network of Ecological Interactions Alireza Tamaddoni-Nezhad
Department of Computing, Imperial College London
[email protected]
David Bohan
INRA
[email protected]
Alan Raybould
Syngenta Ltd
[email protected]
Stephen Muggleton
Department of Computing, Imperial College London
[email protected]
Overview
• Problem and data • Learning trophic links (i.e. food-webs) from FSE data • Functional food-webs • Pitfall vs Vortis food-webs • Automatic corroboration of learned trophic links • Discussion and further work
Ecosystems as networks of biomass flow Carabid larva
PREDATORS
Araneae
HERBIVORS
Can we apply machine learning on FSE data to construct food-webs from scratch?
Trechus quadristriatus
PLANTS
• Networks of trophic links (food webs) are important for explaining ecosystem structure and dynamics • It is difficult to establish trophic relationships between the many hundreds of species in an ecosystem • Project goal:
Auchenorhyncha
Isotomidae
Aphidoidea
Farm Scale Evaluation (FSE) of Genetically Modified herbicide-tolerant (GMHT) crops in the UK Samples from 257 sites, 4 different crops (spring beet, spring maize, spring oilseed rape and winter oilseed rape) across the UK between 2000 and 2004. Sites information: SITE_CODE SEASON CROP REGION B10 2000 B EASTERN B11 2000 B EASTERN B12 2000 B MIDLANDS&WESTER B13 2000 B MIDLANDS&WESTER B14 2000 B MIDLANDS&WESTER ... ... ... ...
SIZE 4.4 14.0 7.3 9.9 3.0 ...
INTENSITY 5 5 4 4 5 ...
BIODIVERSITY 1 1 0 1 0 ...
SOIL Medium Medium Medium Light Heavy ...
Species information: ARTHROPOD weight length 6453 4703 0 3.5 6455 59501 0 7 6453 3503 0.09 7 6453 3513 6.17E-02 8 ... ... ...
size class activity 1 1 3 1 3 1 4 0 ... ...
sap 0 0 0 0 ...
leaves 0 0 0 0 ...
seeds detritus spec pred gen pred 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 ... ... ... ...
Abundance of species in conventional and GMHT farms: ARTHROPOD 6400 111 6400 111 6400 111 6400 111 6400 111 ...
SITE_CODE CONVEN. B14 0 B15 0 B17 9 B18 1 B20 0 ... ...
GMHT 0 0 11 9 1 ...
ARTHROPOD 6400 111 6400 111 6400 111 6400 111 6400 111 ...
SITE_CODE abundance B18 up B21 down B34 down B66 down B67 up ... ...
Figure 1. Distribution of the 66 spring-sown beet (), 59 spring maize (), 67 spring oilseed rape () and 65 winter oilseed rape () fields sampled as part of the Farm Scale Evaluations (FSE) of Genetically Modified, herbicide-tolerant (GMHT) crops.
Abduction of trophic links Background knowledge
Observables
abundance(X, S, up):predator(X), abundance(Y, S, up), co_occurs(S, X, Y), bigger_than(X, Y), eats(X, Y).
X
Abduced hypotheses
eats(a1, a3). eats(a5, a6). eats(a5, a7). . .
Y
Z
Direction of effect
abundance(a1, s1, up). abundance(a2, s1, up). abundance(a3, s1, up). abundance(a1, s2, up). abundance(a2, s2, down). abundance(a3, s2, up). . . .
A probabilistic trophic network from Vortis data
D. Bohan, G. Caron-Lormier, S. Muggleton, A. Raybould and A. Tamaddoni-Nezhad, Automated Discovery of Food Webs from Ecological Data Using Logic-Based Machine Learning. PLoS ONE 6(12): e29028, 2011 A. Tamaddoni-Nezhad, D. Bohan, A. Raybould and S. Muggleton, Machine learning a probabilistic network of ecological interactions, In Proceedings of the 21st International Conference on Inductive Logic Programming, LNAI 7207, pages 332-346, 2012.
Corroboration of hypotheses in the literature and potential novel hypotheses (in red)
* Spearman's correlation between frequencies and number of references: 0.77 with p-value 0.009
Literature which support abduced trophic links 1. K.N.A. Alexander. The invertebrates of living and decaying timber in britain and ireland–a provisional annotated checklist. English Nature Research Reports, 467:1–142, 2002. 2. T. Bauer. Prey-capture in a ground-beetle larva. Animal Behaviour, 30(1):203–208, 1982. 3. J.R. Bell, R. Andrew King, D.A. Bohan, and W.O.C. Symondson. Spatial co-occurrence networks predict the feeding histories of polyphagous arthropod predators at field scales. Ecography, 33(1):64–72, 2010. 4. K. Berg. The role of detrital subsidies for biological control by generalist predators evaluated by molecular gut content Analysis. PhD thesis, Universit¨ats-und Landesbibliothek Darmstadt, 2007. 5. K. Desender and M. Pollet. Ecological data on clivina fossor(coleoptera, carabidae) from a pasture ecosystem. ii. reproduction, biometry, biomass, wing polymorphism and feeding ecology. Rev. Ecol. Biol. Sol., 22(2):233– 246, 1985. 6. A. Dinter. Intraguild predation between erigonid spiders, lacewing larvae and carabids. Journal of Applied Entomology, 122(1-5):163–167, 1998. 7. J.D. Lattin. Bionomics of the nabidae. Annual review of entomology, 34(1):383–400, 1989. 8. J. Lindsey. Ecology of Commanster, http://www.commanster.eu/commanster/insects/bugs/spbugs/ saldula.saltatoria.html. 9. X. Pons, B. Lumbierres, and R. Albajes. Heteropterans as aphid predators in inter-mountain alfalfa. European Journal of Entomology, 106(3):369–378, 2009. 10. C.W. Schaefer and A.R. Panizzi. Heteroptera of economic importance. CRC, 2000. 11. KD Sunderland. The diet of some predatory arthropods in cereal crops. Journal of Applied Ecology, 12(2):507–515, 1975. 12. K.D. Sunderland, G.L. Lovei, and J. Fenlon. Diets and reproductive phenologies of the introduced ground beetles harpalus-affinis and clivina-australasiae (coleoptera, carabidae) in new-zealand. Australian Journal of Zoology, 43(1):39–50, 1995. 13. S. Toft. The quality of aphids as food for generalist predators: implications for natural control of aphids. European Journal of Entomology, 102(3):371, 2005. 14. B.D. Turner. Predation pressure on the arboreal epiphytic herbivores of larch trees in southern england. Ecological Entomology, 9(1):91–100, 1984. 15. D.J. Warner, LJ Allen-Williams, S. Warrington, AW Ferguson, and IH Williams. Mapping, characterisation, and comparison of the spatio-temporal distributions of cabbage stem flea beetle (psylliodes chrysocephala), carabids, and collembola in a crop of winter oilseed rape (brassica napus). Entomologia experimentalis et applicata, 109(3):225–234, 2003. 16. D.C. Weber and J.G. Lundgren. Assessing the trophic ecology of the coccinellidae: their roles as predators and as prey. Biological Control, 51(2):199–214, 2009.
Functional Food-Webs: Can the FSE data be used to test the theory that ecosystems at the field-scale behave as networks that emerge from interactions between trophic functional types ?
A probabilistic trophic network for functional groups
A trophic network constructed by learning trophic interactions between functional groups.
Functional group
Name
Size
11 Large Predator
4
12 Aphid predator
3
10 Medium-large predator
3
23 Predatory heteroptera
3
13 L. pilicornis (1 species group)
3
9 Medium predator
2
22 Medium predator
2
14 Aphid predator (adult)
4
3 Leaf chewer 18 Sucking predator 1 Sap sucker
3 2 1
17 Sucking predator
1
25 Parasitoids
1
7 Detritivores
1
8 Small predators
1
2 Sap sucker
2
A probabilistic trophic network for functional groups
Each functional group is represented by a species which can be viewed as an archetype for the functional group.
Overall learning curves for functional vs individual food-webs (Vortis)
Pitfall vs Vortis sampling
Vortis suction sampler
ARTHROPOD 6400 111 6400 111 6400 111 6400 111 6400 111 ...
SITE_CODE CONVEN. B14 0 B15 0 B17 9 B18 1 B20 0 ... ...
GMHT 0 0 11 9 1 ...
Abundance of species
A pitfall trap
A (new) probabilistic trophic network from pitfall data
Overall learning curves for functional vs individual food-webs (pitfall)
Automatic corroboration of new food-webs (work-in-progress)
Manual corroboration of trophic links in the literature
Automatic corroboration of trophic links (text-mining)
?
Goal of automatic corroboration: Can we adapt existing text-mining techniques and tools (e.g. Pubgene, SciMiner, Lit.Search, ..) to generate Literature Network for the species involved in the learned food-webs ?
Examples of literature networks for genes
A literature network generated by LitSearch
A literature network generated by PubGene
Automatic corroboration of food-webs (using text-mining) Publication Databases
Interaction Lexicon Pair of Species
Food-web
Query
Text Mining
Literature Network
Publication Search Engines
WWW
Literature Network for Merged Vortis & pitfall data
Literature Network for merged Vortis & pitfall data Null hypothesis: The frequencies of merged trophic links (from HFE ) and the number of hits for these links from text-mining are not correlated. 12
Hypothesis Frequency (HFE)
10
8
6
4
2
0 0
500
1000
1500
2000
2500
3000
Total number of hits from text-mining
Results: Correlation between the hypothesis frequencies (HFE) for merged trophic links and the number of hits for these links. Spearman's correlation ρ value is 0.821 with p-value 0.01
Summary • Using FSE data to generate testable food-webs from scratch • Corroboration of many links in the literature • Potential novel hypotheses which could be tested • Functional food-webs are at least as accurate as individual species food-webs (importance of functional ecology) • Initial results on text-mining of trophic links are consistent with frequencies from HFE
Future work • Extending the model (e.g. plant data, temporal data, ..) • Cross-validation of text-mining and improving the accuracy • Learning functional groups as well as trophic links using predicate invention Acknowledgment: G. Afroozi Milani (Imperial College), G. CaronLormier (Rothamsted Research), S. Dunbar (Syngenta)