Local interaction networks reconstructed from global biodiversity data improve pollinator restoration decision making

The paper introduces NECTAR, a novel framework that integrates global biodiversity data to reconstruct local plant-pollinator interaction networks, enabling the design of habitat restoration plant mixes that support significantly more pollinator species than existing methods.

The Gist

Imagine you are trying to throw the perfect party for a very diverse group of guests: bees, butterflies, moths, and flies. You want to make sure everyone is happy, fed, and able to do their job (pollinating your garden).

The problem? You don't have a guest list. You don't know exactly which snacks (flowers) each specific guest prefers. Some guests are picky eaters who only like one specific type of flower, while others are adventurous. If you just guess and plant whatever looks pretty, you might end up with a party where half the guests go hungry and leave early.

This is the challenge scientists face when trying to restore habitats for pollinators. They know plants are good, but they often lack the detailed "menu" of who eats what, where, and when.

Enter NECTAR: The Ultimate Party Planner

The paper introduces a new tool called NECTAR (Network-Enhanced Conservation Tool for Analysis and Recommendation). Think of NECTAR not as a simple list of plants, but as a super-smart, data-driven matchmaking algorithm for nature.

Here is how it works, broken down into simple concepts:

1. The "Big Data" Detective Work

Traditionally, scientists only knew about plant-pollinator relationships that had been directly observed (like seeing a bee on a flower). But nature is huge, and we haven't seen everything. It's like trying to map a city by only walking the main streets and ignoring the alleyways.

NECTAR acts like a detective that fills in the missing map. It takes three huge piles of information and mixes them together:

• Where they live: Where do the plants grow, and where do the bugs hang out?
• When they are awake: Do the flowers bloom at the same time the bugs are flying? (If a flower blooms in winter and the bee is hibernating, they will never meet).
• Family trees: If a bee is known to like a specific type of flower, its "cousin" bees probably like similar flowers too.

By combining these clues, NECTAR can predict interactions that no human has ever seen yet. It's like guessing that a new guest at the party will love the cheese platter because their whole family loves cheese.

2. The "Local Menu" vs. The "Generic Menu"

Imagine you are in Los Angeles. A generic gardening guide might say, "Plant sunflowers!" But in a specific neighborhood in LA, the local bees might actually prefer a different native plant that isn't on the generic list.

NECTAR creates a hyper-local menu. It doesn't just say "Plant native flowers." It says, "In this specific neighborhood, if you plant these 6 specific plants, you will support 3 times more local moth species than if you just picked plants at random."

3. The "Smart Shopping List" (Optimization)

The most powerful part of NECTAR is its ability to solve puzzles. Let's say you have a small garden, a tight budget, and you want to help a specific endangered butterfly.

• The Old Way: You pick 10 random native plants. You hope for the best.
• The NECTAR Way: You tell the computer, "I have space for 10 plants. I need them to be drought-tolerant. I want to save the Monarch butterfly."

NECTAR runs thousands of simulations in seconds to find the perfect combination. It might say, "Don't plant 10 different flowers. Plant these 6 specific ones. They cover the most ground, feed the most different bugs, and fit your budget."

The Results: Why This Matters

The researchers tested this in California (a place with huge biodiversity). Here is what they found:

• Better than Random: If you just picked plants at random, you'd support about 12% of the local pollinator community.
• Better than Expert Lists: Even existing "expert" plant lists (which are good!) only supported about 12-18% because they rely on old data.
• NECTAR Wins: Using NECTAR's optimized list, you could support up to 34% of the pollinator community with the same number of plants. That is nearly tripling the impact of your garden!

The "Calscape" App

To make this useful for regular people, the researchers built a free tool called the Pollinator Companion (part of the Calscape app). You can type in your zip code, tell the app what your soil is like, and how much sun you get. The app then gives you a custom shopping list of plants that will bring the most life to your specific patch of earth.

The Big Picture

This paper is a game-changer because it moves conservation from "guessing" to "precision engineering."

Think of it like upgrading from a paper map (old plant lists) to Google Maps with real-time traffic (NECTAR). The paper map gets you in the right direction, but the smart app gets you to your destination faster, avoids the traffic jams (data gaps), and finds the best route for your specific car (your specific garden goals).

By using big data and smart math, NECTAR helps us turn small actions—like planting a few flowers in a backyard—into a massive, coordinated effort to save our planet's pollinators.

Technical Summary

1. Problem Statement

Global insect declines, particularly among pollinators, threaten ecosystem stability and food security. A primary driver of this decline is habitat loss, making habitat restoration a critical conservation strategy. However, effective restoration is hindered by the "Eltonian Shortfall": a lack of comprehensive data on biotic interactions (specifically which plants support which pollinators).

Current restoration tools rely on expert-curated plant lists that are often:

• Generalized: Designed for broad audiences rather than specific local contexts.
• Data-Poor: Rely on known interactions, missing vast numbers of unobserved or rare interactions (especially for moths and specialized bees).
• Biased: Heavily skewed toward charismatic, easily identifiable species (e.g., butterflies, bumblebees) and accessible locations.
• Static: Fail to account for spatial heterogeneity, phenological overlap, or specific restoration constraints (e.g., drought tolerance, nursery availability).

Consequently, practitioners lack evidence-based, spatially explicit guidance to optimize plant mixes for specific conservation goals.

2. Methodology: The NECTAR Framework

The authors present NECTAR (Network-Enhanced Conservation Tool for Analysis and Recommendation), a modular, data-driven framework designed to generate spatially explicit plant-pollinator interaction networks and optimize planting recommendations.

A. Data Aggregation and Preprocessing

• Scope: Applied to California (35 Jepson eFlora districts/ecoregions).
• Data Sources: Integrated millions of occurrence records from Calflora, GBIF, Ecdysis, and Moth Photographers Group, alongside interaction data from GloBI, HOSTS, and specialized literature.
• Cleaning: Used taxonomic harmonization (GNparser/GNverifier) and spatial cleaning to filter for native species and georeferenced records.
• Scale: Included 5,178 native plant species and 5,131 pollinator species (bees, hoverflies, butterflies, moths).

B. Species Distribution and Phenology Modeling

• Species Distribution Models (SDMs): Used an ensemble modeling approach (GLM, GAM, Random Forest, MAXNET) with 50 environmental predictors (climate, soil, land cover) at 5 arcminute resolution.
  • Tiered Approach: Applied different modeling strategies based on data density (Alpha-hull for ultra-rare, ESM for rare, full ensemble for common) to include species with sparse data.
• Phenometrics: Estimated flowering periods (plants) and flight periods (insects) using circular statistics (sine-skewed von Mises distributions) to handle year-end continuity. A tiered approach was used based on record availability.

C. Interaction Network Prediction

NECTAR predicts interactions by filtering potential pairs through three layers of constraints:

1. Phylogenetic Constraints: If a pollinator interacts with a plant genus, it is assumed to potentially interact with other species in that genus.
2. Graph Embedding & Transfer Learning:
   • Used graph embedding (t-SVD) to convert the interaction network into a vector space.
   • Calculated dot products between pollinator and plant vectors to quantify "interaction potential."
   • Applied transfer learning (ancestral character estimation) to predict interactions for pollinators with no existing data based on their phylogenetic relatives.
3. Spatial and Temporal Overlap:
   • Calculated the probability of interaction in a specific ecoregion ($k$) as a function of spatial co-occurrence ($p_{co-occ}$) and phenological overlap ($p_{co-phen}$).
   • Flower Visitation: $P_{ijk} = p_{co-occ} \times p_{co-phen}$.
   • Host Plant (Lepidoptera): $P_{ijk} = p_{co-occ}$ (independent of phenology).
   • Interactions were classified as "highly plausible" if their probability score exceeded the 25th percentile of confirmed spatially explicit interactions.

D. Optimization and Decision Support

• Genetic Algorithms: Used to select optimal sets of 10 complementary plants that maximize pollinator richness (or specific subgroups) while satisfying constraints (e.g., drought tolerance, nursery availability).
• Comparison: Results were benchmarked against:
  1. Random selection of native plants.
  2. Existing expert-based lists (Pollinator Partnership, Xerces Society).
  3. Random draws from NECTAR's top-ranked plants.

3. Key Contributions

• Scalable Framework: NECTAR is the first working example of scaling data-driven, network-based restoration planning from local to regional scales using heterogeneous global biodiversity data.
• Data Imputation: Successfully bridged the "Eltonian Shortfall" by using machine learning (graph embedding) and phylogenetic inference to predict interactions for data-poor taxa, increasing species-specific plant-pollinator pairs by 12x and spatially explicit interactions by 200x.
• Public Tool Integration: The framework was operationalized into the Calscape Pollinator Companion, an interactive tool allowing users to input constraints and receive optimized plant lists.
• Modularity: The pipeline is designed to be transferable to other regions, interaction types (e.g., natural enemies), and objectives.

4. Key Results

• Network Recovery: The predicted networks recovered 53.8% of withheld local interactions in cross-validation, significantly outperforming random null models (0.1%) and phylogenetic null models (25.2%).
• Taxonomic Bias Correction: While observed data was biased toward bees and butterflies, the predicted networks significantly expanded representation for moths (the most speciose group), increasing flower-visitation interactions for moths from ~20 observed to >360,000 predicted.
• Restoration Efficacy:
  • Random Selection: Supported ~12.8% of the local pollinator community.
  • Expert Lists: Supported ~12.3% (comparable to random in this specific metric, suggesting lists are good but not optimized for local gaps).
  • NECTAR (Optimized): Supported 34.3% of the pollinator community (up to 2.8x more than random selection and existing lists).
• Spatial Heterogeneity: Optimal plant mixes varied drastically by location (e.g., Redding vs. Los Angeles shared zero species in common for the top 6 plants), proving that a "one-size-fits-all" approach is ineffective.
• Constraint Scenarios:
  • Chaparral Restoration: Optimized mixes supported 123–176% more pollinators than random.
  • Specialist Bees: Targeted optimization nearly doubled support for oligolectic bees compared to general optimization.
  • Declining Butterflies: Optimized mixes provided a 10-fold increase in species supported with both host and nectar resources compared to random selection.

5. Significance and Implications

• Evidence-Based Conservation: NECTAR moves pollinator restoration from generalized assumptions to quantitative, evidence-based decision-making. It demonstrates that integrating multiple data modalities (distribution, phenology, phylogeny) significantly improves restoration outcomes.
• Addressing Data Gaps: The framework proves that machine learning and transfer learning can effectively mitigate spatial and taxonomic sampling biases, making restoration viable even in data-poor regions or for under-sampled taxa like moths.
• Multi-Objective Optimization: The tool allows practitioners to balance competing goals (e.g., supporting endangered species vs. maximizing general diversity) and practical constraints (drought, cost), which is crucial for small-scale urban gardening and large-scale restoration alike.
• Global Applicability: By providing open-source code and a modular workflow, NECTAR offers a blueprint for translating global biodiversity data into local conservation actions worldwide, potentially revolutionizing how habitat restoration is planned and executed.