PhiFold: AI Designs Protein Motion, Not Just Structure, with 10,000x Speedup

AI protein design has progressed from structure prediction (AlphaFold) to de novo structure generation (RFdiffusion), but proteins are dynamic machines whose function depends on conformational fluctuations, collective motions, and transient pockets. Current generative models target static 3D structures and cannot guarantee desired dynamic behavior.

A joint team from Peking University, Fudan University, and Tsinghua University proposes PhiFold (Towards Dynamic Protein Design with Physics-Structured Covariance Modeling), a unified framework that jointly generates protein backbones and predicts near-equilibrium dynamics statistics, enabling both forward dynamics prediction and inverse dynamics-guided design.

Core Representation: Residue-Displacement Covariance

Instead of learning full molecular dynamics (MD) trajectories — which are stochastic, path-dependent, and computationally expensive — PhiFold learns the residue-displacement covariance matrix , a second-order statistic capturing per-residue fluctuation amplitudes, directional preferences, and inter-residue correlations. This compact representation retains essential near-equilibrium statistical features without modeling temporal details.

Physics-Structured Covariance Modeling: Three Interpretable "Motion Knobs"

Directly predicting the 3N×3N covariance matrix is high-dimensional and entangles distinct physical factors. PhiFold decomposes the covariance into three physically meaningful components:

Local flexibility — per-residue 3D covariance describing anisotropic fluctuation magnitudes.

Collective motion — low-dimensional modes capturing coordinated multi-residue movements.

Participation ratio — per-residue coefficients weighting the contribution of collective vs. local motion.

The model assembles these components with normalization and variance calibration, guaranteeing a positive-definite overall covariance while preserving each residue's predicted local covariance. This yields a compact, interpretable, and mathematically consistent dynamics representation.

Differentiable Dynamics-Guided Generation

PhiFold attaches a dynamics branch to the pre-trained FoldFlow backbone generator. During training the backbone is frozen and the dynamics branch remains differentiably connected. At sampling time, gradients from user-specified regional flexibility targets (e.g., increase or decrease RMSF) steer the evolving backbone. Thus dynamics become a design condition rather than a post-hoc evaluation metric.

Experimental Validation

Controllable flexibility: On generated proteins of 100–300 residues, targeting increased flexibility raised regional RMSF by 0.104–0.221 Å; targeting increased rigidity lowered RMSF by 0.025–0.101 Å.

Independent MD verification: GROMACS all-atom simulations on designed samples confirmed the intended flexibility direction in the majority of cases despite run-to-run variability.

Benchmark on ATLAS dataset: PhiFold was compared against MDGen, ANM, BioEmu, and Str2Str across 12 metrics covering local flexibility, inter-residue cooperativity, conformational distributions, and dynamic contact statistics. PhiFold achieved the best score on 8 of 12 metrics, including 5 of 6 conformational-ensemble metrics. Notably, RMS displacement correlation reached 0.739 vs. 0.516 (MDGen) and 0.663 (ANM).

Sampling Efficiency: One Prediction, Batch Sampling

Because PhiFold learns the equilibrium covariance, a single forward pass yields the full statistical model. Conformational ensembles are then generated by cheap probabilistic sampling from the learned Gaussian, avoiding repeated neural-network evaluations. For 100–300 residue proteins:

End-to-end amortized cost: 1.89–1.97 ms per conformation .

Speedups: 30–50× vs. MDGen, 280–1,100× vs. BioEmu, 7,000–11,000× vs. Str2Str (up to 10,000×).

Incremental sampling after covariance prediction: 0.311–0.619 µs per conformation (sub-microsecond).

Context: From ProAR to PhiFold

The same team previously introduced ProAR (Probabilistic Autoregressive Modeling for Molecular Dynamics, accepted at AAAI 2026), which models MD trajectories as probabilistic autoregressive processes with anti-drift sampling for variable-length trajectory generation. ProAR addresses temporal evolution ; PhiFold addresses statistical dynamics-aware design . Together they form complementary pillars toward a "Molecular Dynamics World Model" that understands, predicts, and actively designs molecular dynamic behavior.