Inferring species interactions from co-occurrence data with Markov networks

Area of Science:

• Community ecology
• Ecological network analysis
• Statistical physics applications in ecology

Background:

• Inferring species interactions from co-occurrence data is challenging due to indirect ecological effects.
• A single interaction can cascade through a network, altering observed species correlations.

Purpose of the Study:

• To apply Markov networks (statistical physics models) for predicting direct and indirect species interactions.
• To evaluate the performance of Markov networks against existing methods in ecological network analysis.

Main Methods:

• Utilized Markov networks (Markov random fields) to model species interactions.
• Estimated interactions from co-occurrence rates using maximum likelihood, controlling for indirect effects.
• Compared Markov networks with six other approaches using simulated ecological landscapes.

Main Results:

• Markov networks consistently outperformed existing methods in identifying direct species interactions.
• Computationally efficient approximations using partial correlations or generalized linear models also showed strong performance.
• Null models failed to control for indirect effects and produced inaccurate inferences.

Conclusions:

• Markov networks provide a robust framework for accurately inferring direct species interactions in ecological networks.
• The method effectively isolates direct effects even when indirect interactions or abiotic factors are strong.
• This approach advances the reliability of ecological network analysis.