Network-based restoration strategies maximize ecosystem recovery

The study addresses how to design effective, generalizable restoration strategies for degraded mutualistic ecosystems (e.g., plant–pollinator networks) experiencing species loss. Mutualistic networks are highly interdependent and prone to cascading secondary extinctions following primary losses, so prioritizing which species to reintroduce is critical. Prior efforts often focus on single species, habitat/site prioritization, or low-dimensional models and lack consensus on quantifying species’ restoration importance under varying perturbations. The authors hypothesize that network topology, combined with dynamical models, can identify a universal, near-optimal multi-species reintroduction strategy that maximizes recovery (abundance), ensures persistence, and promotes rapid stabilization across diverse ecosystems and perturbation magnitudes.

Recent work shows complex ecosystems often follow universal collapse patterns, suggesting potential universal recovery patterns. Traditional restoration has emphasized single-species interventions, habitat identification, spatial rescue effects, or low-dimensional models. More recent approaches advocate network-based restoration, sometimes aiming to reverse extinction sequences or use phylogenetic information, and leverage centrality metrics to identify keystone species. However, centralities are correlated yet yield divergent prioritizations, leading to uncertainty in optimal reintroduction strategies. The literature links network properties such as connectance and nestedness with stability in mutualistic networks, motivating their consideration alongside recovery criteria (abundance, persistence, settling time).

- Data and networks: Analyzed 30 real-world plant–pollinator mutualistic networks from the Web of Life (unweighted, undirected bipartite species interaction networks), plus 27 synthetic networks spanning a range of sizes and structures. Measured network attributes: size (number of species), asymmetry (ratio of pollinators to plants), connectance (realized links proportion), and nestedness (NODF). - Dynamical models: Simulated abundances using three models: (i) n-dimensional mutualistic dynamical model explicitly representing all species with intrinsic growth, intra-/interspecific competition, immigration, and saturating mutualistic benefits; (ii) a 2-D dimension-reduced model aggregating plants and pollinators (effective abundances P_eff and A_eff) with saturating mutualistic terms; and (iii) a 1-D dimension-reduced model based on the projected adjacency, incorporating logistic growth with carrying capacity, migration, Allee effects, and mutualism via nearest-neighbor weighted degree. Parameters and derivations follow Jiang et al. and Gao et al. (details and values in Supplementary Table 4). - Perturbations: Implemented species loss by node removal with obligate mutualism (species go extinct if they lose all partners). Considered three primary extinction scenarios: random, generalist-preferred (probability proportional to degree), and specialist-preferred (probability inversely proportional to degree), at removal levels of 30%, 60%, and 90% (nine combinations), with 10 ensembles each. - Restoration strategies: After perturbation, species were reintroduced sequentially according to network-based priority metrics recalculated at each step: degree centrality, closeness centrality, betweenness centrality, and nearest-neighbor weighted degree (β_eff, selecting species to maximize incremental β_eff). A null model restored species in random order. Upon each reintroduction, removed mutualistic links were restored and systems were simulated to steady state. - Criteria: Assessed after each reintroduction step: (1) mean abundance X, (2) settling time ST (time to steady state within tolerance 1e-6 for at least five time steps, up to 300 steps ≈100 years with 4-month time steps), and (3) persistence P (proportion of surviving species). Initial abundances were set to 1e-6. - Analysis: For each network, perturbation scenario, and ensemble, computed average criterion values over restoration steps, enabling comparison across networks of varying sizes. Identified “winning” strategies by best-vs-rest policy: higher X and P, lower ST (ties broken by abundance then settling time). Examined correlations between criteria and network attributes using Spearman rank correlation and multivariate linear regression (normalized variables).

- Centrality-guided restoration consistently outperformed random reintroduction in recovering mean abundance and persistence across 30 real-world and 27 synthetic networks, perturbation magnitudes (30%, 60%, 90%), and removal scenarios (random, generalist-, specialist-preferred). - First-order metric sufficiency: Prioritizing reintroduction by degree (total number of mutualistic interactions) achieved near-optimal recovery. Higher-order strategies (betweenness, closeness, compartmentalization) produced only marginal gains over degree; across many cases, performance differences between degree- and betweenness-guided strategies were within ~1% for abundance, settling time, and persistence. - Example ecosystem (Dominica, 74 species): Degree, closeness, and betweenness strategies yielded near-optimal mean abundance across all 720 possible reintroduction sequences for a 20% removal; nearest-neighbor average degree performed well only under generalist-preferred perturbations. - Trade-offs: Strategies yielding higher mean abundance tended to have lower settling times but also lower persistence; random restoration often stabilized fastest (winner for lowest settling time in 58.52% of simulations) but at the cost of fewer persisting species and lower abundance. - Winners across simulations: For persistence, betweenness-guided restoration was the winner in 90.37% of simulations; other centralities accounted for 9.63%. In the 1-D model, betweenness guided 37.04% of wins and closeness 34.81% (context as reported in the study). - Correlation with network structure: In the 2-D model, mean abundance was negatively related to network size and asymmetry and positively related to connectance and nestedness, consistent with ecological theory. The 1-D model reproduced connectance/nestedness relationships but reversed the signs for size and asymmetry, highlighting limitations of more aggressive dimension reduction. - Nestedness dynamics: Early restoration steps showed the strongest gains in nestedness when generalist species were restored first, aligning with generalism’s positive influence on nestedness and explaining why degree-based strategies are near-optimal. - Strong correlations among centrality metrics across real networks explain small performance differences among network-based strategies.

The results show that simple, topology-driven strategies leveraging species’ degree provide robust, nearly optimal restoration outcomes across diverse mutualistic systems and perturbation regimes. Because degree is strongly correlated with higher-order centralities, prioritizing generalist species implicitly captures benefits attributed to network compartmentalization and nestedness, especially during early restoration when improvements in network organization are most impactful. However, the study also reveals a trade-off: maximizing abundance and minimizing settling time can reduce persistence, indicating that single-criterion optimization can produce undesirable ecosystem states. Correlative analyses further validate established relationships (connectance and nestedness promote stability), while discrepancies in the 1-D model’s predictions for size and asymmetry emphasize the need for caution when using reduced models that may not fully capture bipartite mutualisms. Overall, network-based strategies are broadly applicable in data-poor contexts because they require only interaction topology, not interaction strengths or trait/phylogenetic data.

Across 30 real-world and 27 synthetic mutualistic networks, restoring extirpated species based on simple connectivity (degree) yields near-optimal recovery of abundance and strong persistence, with higher-order centrality metrics offering only marginal performance gains. Differences among the best network-based strategies are generally within ~1% across criteria, indicating that the precise choice of centrality is not critical. The framework provides a practical, system-agnostic approach to guide restoration when habitat amelioration is infeasible and interaction strength or trait data are lacking. Given observed trade-offs among abundance, persistence, and settling time, restoration planning should balance multiple objectives. Future work should incorporate temporal environmental variability and develop adaptive, real-time strategies to mitigate ongoing global change impacts on mutualistic networks.

- Network representation is unweighted and undirected; interaction strengths and temporal variation are not explicitly modeled. - Assumes obligate mutualism and uses fixed model parameters; real ecosystems may include facultative interactions and context-dependent dynamics. - Dimension-reduced models (especially 1-D) can misrepresent relationships with structural attributes (e.g., network size, asymmetry). - Focuses on species reintroduction sequences without concurrent habitat restoration or landscape context; outcomes may differ when habitat constraints are binding. - Results rely on known interaction topologies; data gaps or mis-specified networks in data-poor regions could affect prioritizations. - Trade-offs among criteria imply that optimizing for one may degrade others; multi-objective decision-making is required.