DiffeoMorph: Learning Distributed 3D Morphogenesis with Rotation-Invariant Spectral Shape Matching
Organizations: Department of Systems Biology, Harvard Medical School · Department of Data Science, Dana-Farber Cancer Institute · Departments of Biomedical Informatics, Harvard Medical School · Broad Institute of MIT and Harvard
Abstract
Biological systems can form complex three-dimensional structures through the collective behavior of agents that share a common update rule and operate without central control. How such distributed control gives rise to precise global patterns remains a central question not only in developmental biology but also in distributed robotics, programmable matter, and multi-agent learning. Here, we introduce DiffeoMorph, an end-to-end differentiable framework for learning a morphogenesis protocol that guides a population of agents to morph into a target 3D shape. Each agent updates its position and internal state using an SE(3)-equivariant graph neural network, based on its own internal state and signals received from other agents. To train this system, we introduce a new shape-matching loss based on 3D Zernike polynomials, which compares the predicted and target shapes as continuous spatial distributions, not as discrete point clouds, and is invariant to agent ordering, number of agents, and global orientation. To achieve rotation invariance while preserving reflection sensitivity, we include an alignment step that optimally rotates the predicted Zernike spectrum to match the target before computing the loss. We perform benchmarking to establish the advantages of our shape-matching loss over other standard distance metrics for shape comparison tasks. We then demonstrate that DiffeoMorph can form a range of complex shapes from minimally patterned initial conditions. DiffeoMorph provides a general framework for learning distributed control strategies for morphogenesis, swarm robotics, and programmable self-assembly.
Figures & tables
| Ellipsoid | Crescent | Starfish | Helical Ellipsoid | Bunny | |||||||
| Original | Rotated | Original | Rotated | Original | Rotated | Original | Rotated | Original | Rotated | ||
| GNCA | No noise | 16.21 | 17.34 0.48 | 17.32 | 19.74 1.26 | 8.98 | 10.43 0.70 | 19.69 | 21.43 1.39 | 12.15 | 12.66 0.27 |
| NPA | No noise | 0.19 | 1.21 0.39 | 0.18 | 2.44 0.97 | 0.24 | 5.22 2.03 | 0.15 | 3.11 0.60 | 0.17 | 5.93 3.33 |
| +0.05 | 0.25 0.03 | 1.33 0.48 | 0.20 0.02 | 2.57 0.98 | 0.61 0.07 | 5.85 2.87 | 0.41 0.06 | 3.23 0.58 | 0.38 0.02 | 6.39 3.34 | |
| +0.1 | 0.53 0.07 | 1.76 0.55 | 0.53 0.09 | 3.11 1.12 | 1.67 0.28 | 6.77 1.78 | 1.00 0.07 | 3.82 0.63 | 0.94 0.11 | 6.46 3.05 | |
| +0.2 | 3.55 3.14 | 3.39 0.97 | 2.05 0.21 | 4.93 1.29 | 2.90 0.57 | 8.02 1.39 | 3.31 0.52 | 5.47 0.75 | 2.45 0.47 | 6.87 2.44 | |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| NPA | DiffeoMorph | |
| Ellipsoid | 0.931 | 0.685 |
| Crescent | 0.898 | 0.792 |
| Starfish | 0.985 | 0.813 |
| Helical Ellipsoid | 0.925 | 0.875 |
| Bunny | 0.961 | 0.912 |
| Training loss | Pose-normalized test | 95% CI | Test [email protected] | Pred./target radius ratio |
| Chamfer without alignment | 0.03549 | 0.03388–0.03718 | 0.0598 | 0.548 |
| Chamfer after spectral pose alignment | 0.02325 | 0.02237–0.02418 | 0.1623 | 0.870 |
| DiffeoMorph spectral loss | 0.01745 | 0.01649–0.01845 | 0.1986 | 0.995 |
| Fitting objective | Dense radial RMSE | 95% CI | Success below 1% | Success below 2% |
| Raw Chamfer without alignment | 0.14194 | 0.12969–0.15390 | 0.0% | 0.0% |
| Joint shape-and-rotation Chamfer | 0.01372 | 0.01199–0.01562 | 36.8% | 81.2% |
| Rotation-aligned Chamfer | 0.01455 | 0.01276–0.01654 | 32.6% | 76.4% |
| DiffeoMorph, angular-only ablation | 0.00774 | 0.00551–0.01092 | 81.2% | 95.8% |
| DiffeoMorph, angular and radial moments | 0.00563 | 0.00466–0.00663 | 86.8% | 99.3% |
3D Shape Generation
3D shape generation creates and edits 3D objects from learned geometric priors and text, images, sketches, or structural constraints.
See all 112 papers