We combine recent progress in seismic tomography and numerical modeling of thermochemical convection to infer robust features on mantle structure and dynamics. First, we separate the observed density anomalies into their thermal and compositional contributions. The tomographic maps of thermo-chemical variations were computed using a new approach that combines a careful equation of state modeling of the lower mantle, independent constraints on density from probabilistic tomography, and a full statistical treatment for uncertainties analyses. We then test models of thermo-chemical convection against these density components. We compute synthetic anomalies of thermal and compositional density from models of thermo-chemical convection calculated with the anelastic approximation. These synthetic distributions are filtered to make meaningful comparisons with the observed density anomalies. Our comparisons suggest that a stable layer (i.e., that no domes or piles are generated from it) of dense material with buoyancy ratio B ≥ 0.3 is unlikely to be present at the bottom of the mantle. Models of piles entrained upwards from a dense, but unstable layer with buoyancy ratio B ∼ 0.2, explain the observation significantly better, but discrepancies remain at the top of the lower mantle. These discrepancies could be linked to the deflection of slabs around 1000 km, or to the phase transformation at 670 km, not included yet in the thermo-chemical calculations.