Deep learning models have fundamentally changed structural biology. By treating molecular docking—the prediction of how a small-molecule ligand binds to a target protein—as a geometric optimization or generative modeling task, architectures can predict binding poses thousands of times faster than traditional physics-based software.
Yet, a glaring issue has emerged: these models frequently predict poses that are physically impossible. They produce "geometric hallucinations" characterized by steric clashes—configurations where atoms from the ligand and the protein overlap in the exact same physical space, or where bond lengths and angles are severely distorted.
To make deep learning models truly useful for drug discovery, researchers are shifting away from pure geometric optimization and toward methods that explicitly enforce physical validity.
The Illusion of Geometric Accuracy
The root cause of physical invalidity lies in how deep learning models are trained and evaluated. Most models optimize for a geometric metric like Root-Mean-Square Deviation (RMSD) relative to a ground-truth crystal structure.
As Buttenschoen et al. point out in their work on the PoseBusters benchmark, data-driven models often treat molecules as flexible point clouds rather than constrained chemical entities. When a loss function only minimizes the average distance between predicted and true atomic coordinates, it lacks any mathematical awareness of electron cloud repulsion or the Pauli exclusion principle.
Coordinate-Only Loss (RMSD):
Predicted Atom ---> [ Overlap Zone ] <--- Protein Atom
(Low coordinate error, but physically impossible)
Physics-Constrained Loss:
Predicted Atom ---> | Energy Barrier | <--- Protein Atom
(Penalizes overlap, forcing a valid conformation)
Consider a scenario where a ligand atom sits in a tight binding pocket. If the model predicts its position with a slight error that pushes it inside a protein atom, a standard RMSD loss penalizes this error linearly or quadratically based on distance. However, in the real world, the energy penalty for such an overlap approaches infinity.
As a result, Yan et al. demonstrate that standard geometric metrics are fundamentally insufficient; a pose can achieve a deceptively low, "accurate" RMSD while remaining chemically absurd. This failure mode is widespread: Morehead et al. show via the PoseBench stress test that even state-of-the-art co-folding models fail basic chemical sanity checks, while Nittinger et al. observe similar structural violations when models attempt to predict complex allosteric and orthosteric poses. The issue even extends beyond small molecules; Abanades et al. note that high-speed structural prediction models for immune proteins regularly output unphysical atomic overlaps and distorted bond lengths.
Why Pure Data-Driven Models Hallucinate
When deep learning models prioritize spatial coordinates over chemical principles, they tend to memorize specific pocket shapes rather than learning the underlying physics of non-covalent interactions. Lai et al. argue that without explicit modeling of the "molecular glue"—such as hydrogen bonds and hydrophobic effects—models fail to generalize to novel molecular structures.
In a purely data-driven framework, the network might determine that placing an atom inside the protein boundary yields a lower average coordinate error across the training set than risking a larger coordinate penalty by placing it further away in an open, unoptimized region. The model lacks an internal understanding of steric hindrance, leading it to propose binding configurations that are energetically impossible.
Strategies for Enforcing Physical Validity
To resolve this tension between computational speed and physical reality, the field has converged on three primary paradigms to prevent steric clashes and ensure structural validity.
1. Hybrid Pipelines and Post-Processing Relaxation
The most straightforward approach uses deep learning as a fast "global search" mechanism to identify a rough binding pose, followed by a classical physics-based refinement step to clean up the details.
- Li et al. developed a hybrid pipeline that uses AI to quickly guess a starting conformation, then applies physics-based constraints to resolve overlapping atoms and distorted angles.
- Jiang et al. show that while AI models can defeat traditional physics-based methods on difficult cross-docking benchmarks, they still require a final touch of physics-based energy minimization to fix atomic clashes and guarantee chemical validity.
2. Physics-Informed Loss Functions
Rather than fixing clashes after inference, researchers are embedding physical laws directly into the neural network's loss function during training.
By adding an explicit penalty term that scales exponentially when two non-bonded atoms get too close, the network learns to treat protein boundaries as hard barriers. In FeatureDock, the authors demonstrate that incorporating these fundamental physical laws directly into the training loss eliminates the need for separate post-processing filtering steps. Similarly, Zhang et al. leverage explicit chemical and biological constraints within a multi-task architecture (FABFlex) to eliminate steric clashing while maintaining high inference speeds.
3. Force-Guided Generative Dynamics
The most sophisticated approach integrates physical forces directly into the generative trajectories of diffusion models and flow-matching frameworks. Instead of letting the model generate unconstrained point clouds and penalizing them at the end, physical forces actively steer the generation process at every step.
Generative Trajectory (e.g., Flow Matching / Diffusion):
[Unbound State] ---> Step 1 ---> Step 2 (Force Guidance Applied) ---> [Valid Bound Pose]
|
v
(Vector field deflected away from protein atoms)
- Newton-Euler Guidance: Huang et al. introduce a diffusion bridge model (Re-Dock) guided by Newton-Euler physics to simultaneously update ligand poses and protein sidechains without generating clashing structures.
- Unbalanced Flow Matching: Corso et al. develop FLEXDOCK, which uses Unbalanced Flow Matching to model the physical transition of a protein from its unbound to bound state, more than doubling the proportion of energetically favorable poses.
- Force-Guided Flow Matching: Guo et al. introduce ForceFM, which integrates explicit "force guidance" into a flow-matching framework, steering the ligand away from high-energy steric clashes and toward realistic, low-energy pockets.
- Equivariant Constraints: Zhang et al. combine SE(3)-equivariant transformers with diffusion models and explicit physical constraints (DynaPhArM) to ensure that the adaptive "molecular dance" between a drug and a target protein never violates basic physical laws.
Measuring True Physical Validity
The shift toward physics-constrained deep learning has fundamentally altered how molecular docking models are benchmarked. It is no longer enough to report a high percentage of predictions under a 2.0 Å RMSD threshold. Models must now pass rigorous physical validation suites.
The PoseBusters suite has become a standard gatekeeper for this transition. Newer architectures are designed from the ground up to satisfy its criteria. For example, Gao et al. demonstrate that by prioritizing physical and chemical plausibility alongside geometric accuracy, Uni-Mol Docking V2 ensures that over 95% of its predictions are entirely free from steric clashes or chirality inversions on the PoseBusters benchmark.
By treating physical laws as non-negotiable architectural constraints rather than optional post-processing steps, modern deep learning docking frameworks are successfully combining the statistical power of data-driven models with the immutable rigor of classical chemistry.
Key Foundational Work
Interformer: an interaction-aware model for protein-ligand docking and affinity prediction (2024) — contributed 3 concept elements to this direction.
DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (2024) — contributed 2 concept elements to this direction.
Deep Learning for Protein-Ligand Docking: Are We There Yet? (2024) — contributed 1 concept elements to this direction.