Fig. 1: Inferring aneuploid fitness landscapes with ALFA-K.

Conceptual overview. a Schematic representation of an evolving cell population passaged longitudinally over multiple timepoints (panels). Individual cells are colored based on the time point at which their specific karyotype first emerged in the population. b Karyotypes (rows) are determined from single-cell sequencing data (Supplementary Note 4.1). The left heatmap displays inferred chromosome copy numbers (fill color) for various detected karyotypes. In the right heatmap, the frequency (fill color) of distinct karyotypes (y-axis) across different timepoints (x-axis) is shown. c Conceptual visualization of karyotype evolution on a fitness landscape. Each point represents a unique karyotype, positioned according to a 2D projection of its high-dimensional state (x and y axes), with fitness indicated by height (z-axis). Points are colored by their time of first emergence, corresponding to panel (a). d Using the same representation as (c), but highlighting instead the region where fitness estimates are made by the pipeline. Fill color within the charted region indicates the stage of the inference process used to estimate fitness for that specific karyotype (e.g., direct frequency-based estimation vs. Gaussian process regression, see Supplementary Note 1). Validation of forecasting performance. e Two example Gaussian-random-field (GRF) fitness landscapes illustrate how increasing the wavelength (λ) alters topology. f Overview of ABM sampling strategy. Agent-based simulations incorporating MS-driven karyotype changes were run on the GRF landscapes in (e); the resulting karyotype counts served to train and validate ALFA-K. Black bars indicate the longest continuous fitness-increasing interval per simulation; colored ticks mark the initial timepoint for each training window; gray ticks indicate the final training timepoint, and red ticks mark the prediction horizon. g Fraction of ABM simulations whose forecasts outperform a Euclidean “no-evolution” baseline (see Supplementary Note 3.1), evaluated on unseen training data after excluding landscapes with poor internal consistency. Source data underlying these plots are available in the ALFA-K repository.