Decoding the Grammar of Evolution: Foundation Models and deep learning Population Genomic Inference

Population genomics is being reshaped by machine learning. Foundation models now read DNA sequences directly; neural simulation-based inference engines estimate parameters that were once intractable; generative models produce synthetic genomes with realistic linkage structure; and graph-based architectures are beginning to capture variation that linear references cannot. At the same time, the field faces hard questions. How do we trust what these models learn? How well do they generalize across species, populations, and genomic contexts? How do we connect their predictions back to evolutionary theory and biological mechanisms? When do they fail, and how would we know?This Research Topic convenes methodological, theoretical, applied, and critical contributions that together map the role of AI in population and evolutionary genomics today. We welcome work across the full spectrum — from foundation models trained on petabases of sequence to lightweight architectures tailored to specific inference problems, from human biobanks to non-model organisms, microbes, plants, livestock, wildlife, and ancient samples. Our aim is not to crown a single approach but to build a community-wide conversation about what works, what does not, and what comes next.We invite contributions on (but not limited to):1. Foundation models and large-scale representation learning for genomes — variant effect prediction, non-coding constraint, regulatory element discovery, cross-species transfer, and mechanistic interpretation of learned representations.2. Generative models and simulation-based inference — realistic and LD-aware data synthesis, neural likelihood-free inference, normalizing flows, diffusion models, and validation against classical estimators and theory.3. Demographic history, selection, and adaptation — robust inference under model misspecification, detection of soft and polygenic sweeps, time-stratified inference, integration of ancient DNA, and disentangling demography from selection.4. Pan-genome graphs and structural variation — evolutionary dynamics on graph genomes, complex variant discovery and genotyping, and trait/disease association in graph references.5. Complex traits, disease, and polygenic prediction — evolutionary perspectives on heritability and architecture, PRS portability across ancestries, rare variant inference, and gene-environment interaction.6. Microbial, viral, and pathogen evolution — fitness landscapes, antimicrobial resistance, real-time epidemiological inference, within-host evolution, and deep mutational scanning.7. Conservation, agricultural, and ecological genomics — population assessment, adaptive potential under climate change, genomic prediction for breeding, and biodiversity monitoring.8. Interpretability, robustness, and benchmarking — saliency and attribution methods, mechanistic interpretability, calibration and uncertainty quantification, distribution shift, adversarial robustness, and shared benchmarks and datasets.9. Theoretical foundations and critical perspectives — when and how do neural methods recapitulate (or contradict) coalescent and diffusion theory? Where do they fail? What are the limits of inference from sequence alone, and what role remains for classical estimators? 10. Somatic evolution, cancer genomics, and clinical translation. Cancer is an evolutionary process in somatic cell populations, and the same machinery that infers demography and selection from germline data can — with care — be repurposed for tumors. We welcome work on deep-learning-based subclonal reconstruction, simulation-based inference of tumor evolutionary parameters, foundation models for somatic variant effect and mutational signatures, cfDNA and liquid biopsy as non-invasive population sampling, germline-somatic interactions, cancer PRS portability across ancestries, therapy induced selective pressures or bottlenecks and theoretical reconciliation of classical population genetics with the asexual, spatially structured, clinically sampled reality of tumor populations.In scope:• Methods papers introducing new algorithms, architectures, or training strategies• Empirical and applied studies in any organism — humans, non-human primates, model organisms, livestock, crops, wildlife, microbes, and viruses• Reviews, perspectives, and opinion pieces, including critical or skeptical ones• Benchmarks, replication studies, negative results, and head-to-head comparisons• Tutorials, datasets, simulators, and infrastructure that lower the barrier to entry• Theory papers connecting deep learning to population genetic models and inference Topic editors: Badri K Padhukasahasram, Keyan Zhao Source: https://www.frontiersin.org/research-topics/77067/decoding-the-grammar-of-evolution-foundation-models-and-deep-learning-population-genomic-inference