Local interaction networks reconstructed from global biodiversity data improve pollinator restoration decision making
Baiotto, T., Cosma, C., Cheung, Y. Y. J., Cheung, Y. Y. J., Narango, D., Woodard, J., McCarville, P., Echeverri, A., et al.
ecology
Quick verdict
NECTAR infers statewide plant-pollinator networks from global biodiversity data for restoration; strong validation of visitation links, host links untested.
Summary
The authors present NECTAR, a modular framework that converts global biodiversity databases into spatially explicit plant–pollinator interaction networks to guide restoration plant selection. Applied across California's 35 ecoregions, it integrates 6.6 million cleaned occurrence records, 309,506 interaction records, ensemble species distribution models, circular-statistics phenology, genus-level phylogenetic constraints, and graph embedding with transfer learning, predicting 2,473,729 interactions among 3,792 pollinators and 4,363 native plants. Cross-validation recovered 53.5% of withheld interactions versus 25.1% under a phylogenetic null and 0.1% under a random null, with pseudo-precision of 2.33–4.68; recall exceeded 90% on three of four independent field datasets. NECTAR's plant mixes are predicted to support up to 2.4× more pollinator species than existing resources (Xerces, Pollinator Partnership) or random draws, and scenario optimizations under realistic constraints improved support 2–11×, including specialist bees and declining butterflies. Main limitations: Lepidopteran host-plant networks, which underpin the 11× butterfly result, were not validated; the headline benchmark scores all selection methods on NECTAR's own predicted networks; the graph-embedding step (~65% of predicted pairs) was excluded from the primary cross-validation; and recall is lower for data-poor pollinators such as moths. Code and data are publicly available.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- NECTAR integrates occurrence, phenological, phylogenetic and network-embedding data to predict one of the largest spatially explicit plant–insect metawebs to date: 2,473,729 interactions across California's 35 ecoregions.
- Predicted networks recover 53.5% of withheld interactions — more than double both null models — and achieve >90% recall on three of four independent field datasets.
- NECTAR-derived plant mixes are predicted to support up to 2.4× more pollinator species than existing resources or random selection of native plants.
- Under realistic planting constraints, optimized mixes outperform random selection by 2–11×, including an 11× gain in host-plus-nectar coverage for declining butterflies.
- Predicted networks show strong spatial turnover — the top six plants for Redding and Los Angeles share no species — so restoration guidance must be region-specific.
Strengths
Scale and multi-modal data integration
Aggregates 6.65M cleaned occurrence records and 309,506 interaction records into ensemble SDMs (median CBI 0.644–0.813), circular von Mises phenology, genus-level phylogenetic filters and graph embedding, expanding species-level plant–pollinator pairs from 30,514 to 632,628 across 35 ecoregions (Table 1).
Layered, null-controlled validation
Withheld-interaction recall of 53.5% vs 25.1% (phylogenetic null) and 0.1% (random) across 25 withholding iterations × 100 null draws; pseudo-precision 2.33–4.68; four independent field datasets with >90% recall on three (Supp. Figs. S14–18).
Extensive sensitivity analysis
End-to-end deployment with realistic constraints
Weaknesses
Host-plant networks unvalidated yet load-bearing
Benchmark circularity in the 2.4× comparison
Primary CV omits the largest extrapolation component
Reliability drops exactly where coverage is celebrated
Minor numeric inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a bioRxiv preprint (posted 2026-09-17, DOI 10.64898/2026.03.30.715389, not yet peer reviewed) from a Cornell-led team with the Conservation Biology Institute and the California Native Plant Society. It presents NECTAR, a modular ecoinformatics framework that reconstructs spatially explicit plant–pollinator interaction networks from aggregated biodiversity data and converts them into planting recommendations for pollinator habitat restoration. The California demonstration spans 35 Jepson eFlora districts and yields 2,473,729 predicted spatially explicit interactions (2,256,597 flower visitation plus 217,132 Lepidopteran host-plant links) among 3,792 pollinators and 4,363 native plants (Table 1). Central claims: (i) a ~20× expansion of species-level plant–pollinator pairs (30,514 raw to 632,628 predicted); (ii) 53.5% recall of withheld interactions versus 25.1% under a phylogenetically constrained null and 0.1% under a random null; (iii) plant mixes supporting up to 2.4× more pollinator species than existing resources (Xerces, Pollinator Partnership) or random draws; and (iv) 2–11× gains under realistic planting-constraint scenarios. The framework ships as an open pipeline and a public tool (Calscape Pollinator Companion, Box 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is NECTAR?
NECTAR (Network-Enhanced Conservation Tool for Analysis and Recommendation) is an open-source framework that combines species occurrence data, species distribution models, flowering/flight phenology, phylogenetic constraints and graph embedding to predict which plants interact with which pollinators in which regions, then recommends native plant mixes for habitat restoration. It was demonstrated across California and is live in the Calscape Pollinator Companion tool.
How much better are NECTAR's plant recommendations than existing pollinator plant lists?
In the paper's predictions, 10-plant NECTAR mixes support up to 2.4 times more pollinator species than random draws of regionally native plants or random picks from existing expert lists (Xerces Society, Pollinator Partnership). Note this comparison is scored on NECTAR's own predicted networks, and scenario-level optimizations improved support 2–11× versus random selection.
How accurate are NECTAR's predicted plant–pollinator interactions?
In cross-validation, NECTAR recovered 53.5% of interactions withheld from its training data, versus 25.1% for a phylogeny-based null model and 0.1% for random guessing. On four independent field datasets it recovered over 90% of interactions known in other regions for three datasets, and up to 60% of completely novel interaction pairs. Roughly half of withheld interactions were still missed, and accuracy is lower for poorly surveyed groups like moths.
Can I use NECTAR outside California?
The framework itself is region-agnostic — the authors describe it as modular and transferable to other regions, spatial units, goals and interaction types — but the published, validated implementation and the Calscape tool cover California only. Applying it elsewhere requires re-running the full data pipeline with local occurrence and interaction data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated several passages (parts of Results around Fig. 2, portions of Methods and Discussion); judgments rely on the surviving text.
- Figures and supplementary materials were unavailable; supplementary-dependent numbers (per-dataset external recall, embedding-only recall) could not be verified.
- This is a non-peer-reviewed preprint (posted 2026-09-17); the assessment reflects that version.
- Zenodo/GitHub deposits were not inspected, so completeness of archived intermediates is assumed from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.