Predicting disease risk from DNA presents an unprecedented emerging challenge as biobanks approach population scale sizes ($N>10^6$ individuals) with ultra-high-dimensional features ($L>10^5$ genotypes). Current methods, often linear and reliant on summary statistics, fail to capture complex genetic interactions and discard valuable individual-level information. We introduce **PRSformer**, a scalable deep learning architecture designed for end-to-end, multitask disease prediction directly from million-scale individual genotypes. PRSformer employs neighborhood attention, achieving linear $O(L)$ complexity per layer, making Transformers tractable for genome-scale inputs. Crucially, PRSformer utilizes a stacking of these efficient attention layers, progressively increasing the effective receptive field to model local dependencies (e.g., within linkage disequilibrium blocks) before integrating information across wider genomic regions. This design, tailored for genomics, allows PRSformer to learn complex, potentially non-linear and long-range interactions directly from raw genotypes. We demonstrate PRSformer's effectiveness using a unique large private cohort ($N \approx 5$M) for predicting 18 autoimmune and inflammatory conditions using $L \approx 140$k variants. PRSformer significantly outperforms highly optimized linear models trained on the *same individual-level data* and state-of-the-art summary-statistic-based methods (LDPred2) derived from the *same cohort*, quantifying the benefits of non-linear modeling and multitask learning at scale. Furthermore, experiments reveal that the advantage of non-linearity emerges primarily at large sample sizes ($N > 1$M), and that a multi-ancestry trained model improves generalization, establishing PRSformer as a new framework for deep learning in population-scale genomics.

MOTIVATION: ASTRAL is the current leading method for species tree estimation from phylogenomic datasets (i.e. hundreds to thousands of genes) that addresses gene tree discord resulting from incomplete lineage sorting (ILS). ASTRAL is statistically consistent under the multi-locus coalescent model (MSC), runs in polynomial time, and is able to run on large datasets. Key to ASTRAL's algorithm is the use of dynamic programming to find an optimal solution to the MQSST (maximum quartet support supertree) within a constraint space that it computes from the input. Yet, ASTRAL can fail to complete within reasonable timeframes on large datasets with many genes and species, because in these cases the constraint space it computes is too large. RESULTS: Here, we introduce FASTRAL, a phylogenomic estimation method. FASTRAL is based on ASTRAL, but uses a different technique for constructing the constraint space. The technique we use to define the constraint space maintains statistical consistency and is polynomial time; thus we prove that FASTRAL is a polynomial time algorithm that is statistically consistent under the MSC. Our performance study on both biological and simulated datasets demonstrates that FASTRAL matches or improves on ASTRAL with respect to species tree topology accuracy (and under high ILS conditions it is statistically significantly more accurate), while being dramatically faster-especially on datasets with large numbers of genes and high ILS-due to using a significantly smaller constraint space. AVAILABILITYAND IMPLEMENTATION: FASTRAL is available in open-source form at https://github.com/PayamDiba/FASTRAL. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.