Pearl millet genomic vulnerability to climate change in West Africa highlights the need for regional collaboration

The study addresses how climate change threatens rainfed agriculture in sub-Saharan Africa by reducing yields and increasing climate extremes. Traditional approaches (ecological niche models) have assessed future suitability of crop areas but typically ignore intraspecific genetic diversity and local adaptation. The authors propose leveraging existing varietal (landrace) diversity as a short-term adaptation strategy, akin to assisted migration, to maintain productivity under future climates. Using pearl millet, a staple adapted to a broad rainfall range and cultivated largely as landraces, they aim to quantify genomic vulnerability to projected climates and evaluate whether targeted movement of currently cultivated, climate-adapted landraces could mitigate risk. The work integrates genomic data with agronomically relevant climate metrics aligned to the monsoon-driven growing season to identify vulnerable regions and inform regional adaptation and seed exchange strategies.

Prior work shows climate change is already depressing crop yields and will increase heat waves and heavy rainfall, with major implications for West African rainfed systems. Ecological niche modeling has been used to predict future suitability for crops but does not encompass within-species diversity and local genomic adaptation. Recent advances in landscape and ecological genomics (e.g., gradient forest approaches) combine genomic variation with environmental gradients to predict adaptive potential and genomic offsets under future climates. Assisted gene flow/migration concepts from wild species are proposed as analogs for crop variety movement. The study builds on CMIP5 multi-model climate projections and agronomically relevant metrics (e.g., monsoon onset) to improve prediction fidelity for cropping systems, addressing known uncertainties in Sahel precipitation projections.

- Plant material: 173 geo-referenced pearl millet landraces from 10 West/Central African countries collected by IRD (1974–1989). For each landrace, 100 seedlings were grown, and equal leaf tissue pooled for DNA extraction (total 17,300 seedlings). - Sequencing and SNP calling: Target capture of genic regions using 152,619 80 bp baits (first 1000 bp of annotated genes), library prep with HiSeq2500 sequencing. Reads trimmed (Cutadapt), quality filtered, mapped to the reference genome (BWA-MEM), low-quality/unmapped reads removed (SAMtools), coverage assessed (QualiMap), realignment (GATK IndelRealigner), variant calling (GATK UnifiedGenotyper). SNP filters: bi-allelic; per-accession depth >10 and <250; <3 SNPs in 5 bp window; alternate allele frequency AF > 0.003 (≥5 reads across dataset); per accession total read count ≥20 (else NA); only SNPs with complete data retained. Result: 138,948 polymorphic SNPs; allele frequencies estimated per landrace. Replication checks showed high correlation of allele frequency estimates (r = 0.91–0.96). - Population structure: PCA on allele frequencies (R prcomp) to assess structure and geographic clustering. - Climate data: Observed climate from EWEMBI (0.5° daily, 1979–2013). Extracted mean over 1979–1989 (matching collection period) for pixels containing landrace locations. Future projections from 17 CMIP5 models (bias-corrected via CDF-t), for RCP2.6 and RCP8.5 at 2050 (2049–2059) and 2100 (2089–2099), plus historical (1979–1989) runs. Cultivation area defined as convex hull of landrace coordinates (~3.1 million km²; 1041 pixels). - Climate metrics: 157 agronomically relevant metrics per pixel/model: monsoon onset; plus 156 statistics (mean/min/max counts) for precipitation, mean/max/min temperature, and shortwave radiation computed over windows of 30, 60, 90, 120, 150, 180 days post-onset. - Gradient Forest (GF) modeling: For SNPs with minor allele frequency >10% (16,632 SNPs), fitted GF models (500 trees/SNP) linking allele frequencies to climate metrics at the 173 sites. Identified important climate predictors and SNPs with predictive power (R² > 0). Built separate GF models for each of the 17 climate models using historical climate corresponding to collection dates. - Genomic vulnerability (genomic offset): Predicted genomic composition across the cultivation area under historical and future climates using GF turnover functions. Computed genomic vulnerability as Euclidean distance between current and future genomic compositions. Assessed for each climate model and RCP scenario; also tested using all metrics vs an uncorrelated subset (|r| < 0.7). - Flowering time phenotyping and GWAS: Flowering time measured in six trials at ICRISAT Sadoré, Niger (2016–2017), 10 plants per landrace under rainfed with supplemental irrigation. GWAS performed on 27,409 SNPs (MAF >5%) using LFMM (lmm R package) with K=5 latent factors (FDR 5%) after model selection via PCA screeplot and QQ-plots; candidate genes identified by intersecting SNPs with gene annotations. Compared GF R² for flowering-associated SNPs vs all SNPs (Wilcoxon test). - Yield validation of vulnerability: Common garden in Sadoré; measured 100-seed weight, mean seed weight on main spike, and total seed weight per plant (mean main spike seed weight × mean number of productive tillers). Computed genomic vulnerability for each landrace as Euclidean distance between predicted genomic composition under origin climate vs Sadoré climate using GF based on observed climate; correlated vulnerability with yield traits (Pearson r). - Assisted migration scenarios: For each climate model and RCP8.5 at 2050, identified vulnerable areas by clustering the top 10% most vulnerable pixels using DBSCAN (clusters ≥4 pixels, inter-pixel separation <1200 km). For each vulnerable cluster, selected the pixel with highest vulnerability, then computed genomic vulnerability between that pixel’s future climate and all current pixels to identify the minimum (EDmin) and its location (optimal source). Calculated migration distance (geodesic between source and target) and migration load (EDmin). Also evaluated near-optimal (closest among the 1% lowest ED) and sub-optimal (closest among the 5% lowest ED) scenarios, recording migration distance and load. Multi-model results combined across 17 models.

- Genomic structure: PCA of SNP allele frequencies showed landraces cluster by country/geography, indicating spatial genetic structure. - Climate predictors: In GF models, key predictors of genomic variation included minimum shortwave radiation intensity at the start of the growing season, monsoon onset, and early-season precipitation intensity. - Model fit: On average, 88% of SNPs (≈14,544–14,757 of 16,632; R² > 0) were predictable by climate variables across models. - Spatial vulnerability pattern: Genomic vulnerability (2050) exhibited a latitudinal pattern with higher values near ~10°N and ~15°N, with the northern edge of cultivation for both early and late flowering types being most vulnerable. Some regions (e.g., Niger under RCP8.5) showed high inter-model variability (CV > 50%) due to precipitation projection uncertainties. - Flowering time link: Flowering time across landraces ranged 41–124 days with bimodal distribution (~60 days early vs ~110 days late), spatially structured (long-cycle in south up to ~11°N; short-cycle in north ~13–16°N). Vulnerability was highest at the northern cultivation limits for both phenology groups. GWAS identified 103 SNPs (75 genes) associated with flowering time (mainly on chromosomes 1, 2, 5). These SNPs had higher GF explanatory power (mean R² = 0.53) than all SNPs (mean R² = 0.28; Wilcoxon p < 2×10^-16), indicating flowering-time loci drive vulnerability predictions. - Yield validation: Genomic vulnerability negatively correlated with yield-related traits in the Sadoré common garden: r = -0.412 (100-seed weight, p < 0.0001), r = -0.368 (mean seed weight on main spike, p < 0.0001), r = -0.310 (total seed weight per plant, p < 0.0001). Using uncorrelated climate metrics slightly reduced correlations (e.g., r = -0.374 for 100-seed weight). - Assisted migration feasibility (RCP8.5, 2050): Across 17 models, 80 vulnerable areas were identified. Optimal migration distances spanned 77–3665 km (mean 1059 km; SD 801 km), with 88.3% requiring transboundary movement. Migration load (EDmin) ranged 0.0010–0.0215 (mean 0.0045). High loads (>0.01) in some cases indicate future climates lacking close current analogs among cultivated areas. Near-optimal scenario reduced mean distance to 537 km (63.75% transboundary) with mean load 0.0052; sub-optimal reduced to 257 km (37.5% transboundary) with mean load 0.0062, implying 10–30% higher load than optimal.

Integrating genomic data with agronomically tuned climate metrics allowed the authors to quantify where and how much genetic change is needed for pearl millet to track climate change. The approach addresses limitations of niche models by capturing local adaptation signals embedded in allele frequency–environment relationships. The strongest vulnerabilities occur at the northern margins of cultivation for both early- and late-flowering types, consistent with phenology’s central role in synchronizing reproduction with favorable conditions. The tight association between vulnerability and yield loss in common gardens validates genomic vulnerability as a biologically meaningful predictor of performance under climate shifts. Assisted migration using existing landraces can partially mitigate projected impacts but often requires long-distance, transboundary seed movement, highlighting the need for coordinated regional strategies and policy frameworks in West Africa. Flowering time emerges as a key breeding and selection target to enhance resilience under warming and altered rainfall regimes. The multi-model climate ensemble underscores areas of robust signal versus high uncertainty (notably precipitation over parts of the Sahel), guiding where recommendations are more reliable. Social dimensions, including farmer seed networks and potential resistance to adopting non-local varieties, will influence the feasibility of varietal replacement, suggesting participatory breeding and regional collaboration as critical enablers.

This study develops and validates a genomic–climate framework to map pearl millet’s vulnerability to future climates in West Africa and to propose assisted migration strategies leveraging existing landrace diversity. Key contributions include: identifying the northern cultivation margins as hotspots of genomic vulnerability; demonstrating flowering time as a pivotal adaptive trait underpinning vulnerability; empirically linking genomic vulnerability to yield loss; and quantifying that effective mitigation will often require long-distance, transboundary seed transfers. The findings call for regional collaboration across West Africa to coordinate seed exchange, germplasm conservation, and deployment, and for integrating farmer participation to ensure adoption. Future research should refine trait-based genomic predictions (e.g., multi-trait adaptation), incorporate socio-economic constraints and seed system dynamics into migration planning, extend analyses to additional crops and regions, and update vulnerability maps with higher-resolution, next-generation climate projections and on-farm performance data.

- Climate projection uncertainty: High inter-model variability in precipitation projections for parts of the Sahel translates into uncertain vulnerability estimates locally (e.g., Niger under RCP8.5). - Analog gaps: High migration loads in some cases indicate future climates without close current analogs in cultivated areas, limiting adaptation via present-day landraces alone. - Genomic scope: Pool-seq of genic regions may miss regulatory or structural variants contributing to adaptation; only SNPs with MAF >10% were modeled in GF. - Environmental scope: Climate metrics focused on monsoon-aligned windows; non-climatic stresses (soils, pests, management) and extreme event dynamics beyond metrics used may affect adaptation. - Adoption and seed systems: Social, cultural, and logistical barriers to transboundary seed movement and farmer adoption are not modeled but are critical to implementation. - External validity: Yield validation used a single common garden site; broader multi-site trials would strengthen transferability of the vulnerability–yield relationship.