MetaDiffusion: A process-aware conditional diffusion model for multi-species spatial range reconstruction from sparse observations

Poster presented at the 25th European Conference on Computational Biology (ECCB 2026), Geneva, under the theme "Biodiversity, AI & Health".

For most species we know where they have been seen, not where they live. A typical species is known from a handful of occurrence records, and IUCN Red List area metrics are derived from those same sparse records. Correlative species distribution models fit presence to environment and return a single map, which hides how uncertain that map is and cannot distinguish a true absence from a survey gap.

We present MetaDiffusion, a conditional denoising diffusion model that reconstructs multi-species ranges from sparse records and returns an ensemble of plausible maps rather than one. It is trained on Lotka-Volterra individual-based metacommunity simulations, so every reconstruction can be scored against a range that is known exactly. The model is conditioned on a per-cell environmental field, a species-interaction graph, and a short occupancy history, fused into one per-species vector that steers a species-parallel U-Net. Observed cells are pinned during sampling, so every ensemble member honours the records and differs only in what it infers.

On held-out communities at environment seeds never used in training, reconstructed ranges match the truth in geometry (Kolmogorov-Smirnov distances of 0.13 for range size, 0.23 for fragmentation and 0.095 for spatial spread). Recall on hidden occupied cells is 20-44% within two grid cells of a record and at chance beyond; a tuned Gaussian kernel smoother given the same records also recovers nothing beyond that radius, so the limit lies in the information rather than in the method. Ensemble coverage rises with record count, from 23% (fragmentation) and 43% (spread) at one record to 75% and 80% at five to nine, below the 88% a calibrated ensemble would attain. A leave-one-out ablation finds the occurrence records to be the only conditioning input carrying measurable signal: what the model knows about ecology it learned during training, and no interaction network needs to be supplied at prediction time.

Preliminary and unpublished results, shared for discussion. Full manuscript in preparation.