Fast generative models for molecular structure have nearly removed iterative inference, the need to run a model many times in sequence to refine a structure. One example is a protein structure predictor distilled into a single forward pass. These models still produce physically impossible outputs: atoms overlapping, rings threaded through side chains, ligands stuck inside a pocket wall. The usual repair is to run a classical force-field minimizer afterwards, which pushes atoms downhill on a physics energy. That brings back the sequential latency the fast model was built to remove. It also often fails, because an entangled pose is a local minimum in 3D. No small motion separates two interpenetrated groups without first pushing them harder into each other.
We propose to repair such poses in a lifted space. Each atom gets one extra coordinate, and the minimizer can move atoms "around" each other through a fourth dimension. A penalty then pulls everything back into 3D. The key observation is that this lifting turns clash-induced traps into saddles with a guaranteed downhill direction. We then distill the lifted repair into a refiner that is applied once or twice, so the final pipeline keeps single-pass speed. In two sentences: clashes that trap 3D minimizers become escapable once atoms may briefly leave 3D. That escape procedure can be learned and amortized, so fast generators get physical validity without paying for long relaxation.
Core Claim
The idea of converting spurious minima into strict saddles comes from low-rank matrix sensing. A strict saddle is a point where the slope is zero but at least one direction curves downward. Over-parametrization via Lifting for Low-rank Matrix Sensing lifts a vector into a higher-order tensor. Spurious points stay stationary under the lift, but their curvature changes: the smallest Hessian eigenvalue at a spurious point moves from 2.67 to −0.17. Global recovery rises from 47.5% to 75%. That lift costs $n^l$ parameters. Our lift costs one scalar per atom, and the physics tells us exactly where the negative curvature comes from.
Give atom $i$ a 3D position $x_i$ and an auxiliary coordinate $w_i$. We minimize
$$ E_{\text{lift}}(x,w)=\sum_{i<j}\phi_{ij}\!\left(\sqrt{\|x_i-x_j\|^2+(w_i-w_j)^2}\right)+\mu\sum_i w_i^2 , $$
where $\phi_{ij}$ is the force-field pair term (Lennard-Jones plus electrostatics; bonded terms are kept on 3D distances), and $\mu$ is a penalty weight that we raise over time. Because the energy is symmetric under $w\to -w$, every 3D stationary point remains stationary at $w=0$, exactly as in the tensor-lifting result. The second derivative along a relative displacement $w_i-w_j$ at $w=0$ equals $\phi_{ij}'(r_{ij})/r_{ij}+2\mu$. For an overlapping pair the repulsive force makes $\phi'_{ij}<0$. So whenever clash strength exceeds the penalty, a trapped 3D pose becomes a strict saddle with an explicit escape direction.
This gives a falsifiable mechanistic prediction. Poses that remain invalid after standard 3D minimization should show significant negative curvature in the lifted Hessian. Poses that 3D minimization does repair should not.
Research Program
The work has three stages.
Diagnosis. We take poses from fast generators and run conventional 3D minimization on them. For the ones that remain invalid, we compute the smallest lifted-Hessian eigenvalue using Lanczos iterations, a standard method for extracting extreme eigenvalues cheaply. We test whether residual invalidity is predicted by lifted negative curvature.
Lifted relaxation as a solver. We minimize $E_{\text{lift}}$ with $\mu$ annealed from small to large, ending in an exact projection to $w=0$ and a short 3D polish.
Amortization. We train a small equivariant graph network to map a generator's raw pose to the endpoint of the lifted relaxation, in either one or two applications. Training data comes from lifted relaxations of generated poses on PDBBind training complexes.
Each stage builds on specific prior work. The idea of auxiliary dimensions untangling geometric folds is supplied by Dimension expansion and customized spring potentials for sensor localization, which lets folded 2D sensor layouts flip through a third dimension. Self-Calibration for the Time-of-Arrival Positioning reduced positioning failures from over 60% to near zero with auxiliary coordinates. Neither works with molecular force fields or a learned amortizer.
The closest molecular work is Neural Network Reparametrization for Accelerated Optimization in Molecular Simulations. It optimizes an over-parametrized latent space through a graph network built on Hessian slow modes, and it reaches 1.5–3× speedups on small proteins. Its authors note that each step remains expensive and nothing is amortized. We differ in two ways: our lift is explicit and analyzable rather than learned per instance, and we distill the trajectory away.
The amortization stage borrows its motivation from DCFold, which collapses AlphaFold3's recycling and diffusion into one pass at 78.6% versus 82.9% PoseBusters V2 success. DCFold stays in Euclidean coordinates and has no means of escaping a bad geometry. We treat it as a source of poses to repair, not as a competitor to replace.
The constraint motivating all of this comes from PoseBusters. It shows learned docking poses are often physically invalid, with DiffDock's valid-and-accurate rate at 12% on its 308-complex set. It also shows that force-field minimization helps, so the bar is classical relaxation, not raw output.
Baselines
The naive reference point is the raw output of a fast generator with no repair. We will use DiffDock, run both at its default 20 steps and at truncated step counts of 1–4, and DCFold where its structures are available. DiffDock's reported valid-and-accurate rate is 12% on the PoseBusters Benchmark Set and 47% on Astex, as reported by PoseBusters. Its PDBBind top-1 success is 38.2%.
The strongest existing competitor is the same generator followed by standard 3D force-field minimization, as evaluated in PoseBusters. It is the strongest because it already fixes many violations and is what practitioners actually do. We found no single reported valid-and-accurate figure for minimized DiffDock in the file ("no reported result found on Lacuna" for that combination). We will therefore rerun it ourselves.
To keep the comparison fair, it will use the same force field (AMBER through OpenMM), the same protein preparation, the same ligand-pocket atom set, and an identical energy-evaluation budget to our lifted relaxation. We will sweep the budget from 50 to 5,000 evaluations so the comparison is a curve rather than a point. Classical tools Gold and AutoDock Vina provide physically grounded reference points: 64% valid-and-accurate for Gold on Astex, and 58% and 55% physical validity respectively on the PoseBusters set.
The key ablations target what is actually new. The first runs lifted relaxation with the auxiliary coordinates frozen at zero. This is identical code and annealing schedule, so it isolates the fourth dimension itself. The second trains the amortized refiner on unlifted 3D minimization endpoints instead of lifted ones. This separates the value of the lift from the value of a learned refiner.
Beating the field means three things. Lifted relaxation must beat equal-budget 3D minimization on valid-and-accurate rate at every budget. The one-to-two-pass refiner must match or exceed 3D minimization at its full budget while costing no more than a few generator-equivalent forward passes in wall-clock time. And it should approach Gold's 64% on Astex.
Why not just…?
Why not just run longer 3D force-field minimization? This is the real threat. If residual invalidity comes from poor convergence rather than topological traps, more steps will close the gap and the lift is pointless. The diagnosis stage settles this. If poses that stay invalid after 5,000 3D evaluations show no lifted negative curvature, and the frozen-$w$ ablation matches the lifted version, we will report that lifting adds nothing for docking and confine the claim to amortization.
Why not just use more diffusion steps, or restrict generation to physical degrees of freedom? DiffDock already diffuses only over translation, rotation and torsions, and it uses 20 steps. PoseBusters still finds its validity collapsing on novel complexes. Restricting the degrees of freedom does not stop torsion-space trajectories from ending in entangled poses. We will include DiffDock at 20 and 40 steps to show the latency–validity tradeoff directly.
Why not just coarse-grain the system? Reducing atoms to beads smooths the landscape, but it requires mapping beads back to full atoms. Neural Network Reparametrization argues this loses atomic fidelity. Clash and ring-planarity checks are atomistic, so a coarse-grained repair would have to be followed by full-atom minimization anyway. That reintroduces the baseline above.
Why not just add a clash penalty when training the generator? This is plausible and cheap. We will train a DiffDock-sized single-pass variant with a steric loss as an extra baseline. Our expectation is that it lowers average overlap but cannot express the non-local "pass-through" moves needed to undo threading. The comparison on the subset diagnosed as lifted-negative-curvature will show whether that expectation holds.
Decisive Checks
We will run on the PoseBusters Benchmark Set, Astex Diverse, and the PDBBind 2019 test split. The metrics are valid-and-accurate rate (RMSD under 2 Å and passing all PoseBusters checks), RMSD success alone, and wall-clock time per complex on one GPU.
The proposal stands or falls on three results. First, on poses that 3D minimization cannot fix, the fraction showing a negative lifted eigenvalue should be high, and lifted relaxation should repair a clear majority of them. Second, the frozen-$w$ ablation should lose most of that gain. Third, the distilled refiner should keep the gain at one or two passes on the post-2021 PoseBusters complexes, which none of its training data overlaps.
If the first check fails, the curvature account is wrong for real force fields. If the third fails, the lift is a good solver but not an amortizable one. In that case the honest contribution is a better post-hoc minimizer, not a fix to iterative-refinement latency.
References
- Over-parametrization via Lifting for Low-rank Matrix Sensing: Conversion of Spurious Solutions to Strict Saddle Points
- Dimension expansion and customized spring potentials for sensor localization
- Self-Calibration for the Time-of-Arrival Positioning
- Neural Network Reparametrization for Accelerated Optimization in Molecular Simulations
- DCFold: Efficient Protein Structure Generation with Single Forward Pass
- PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences
- DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking