Research
PathFold combines the representational power of large protein language models with constraints from energy-landscape theory to act as a scalable surrogate for molecular dynamics, enabling kinetic inference and misfolding divergence detection across large and diverse protein families.
Training signal
Direct folding trajectories exist only for a handful of small proteins. PathFold therefore integrates three complementary data tiers, which also makes the central data-availability problem of trajectory-based modelling explicit and tractable.
Long all-atom MD simulations of ~12 fast-folding proteins under 80 residues and Folding@home subsets, discretised with Markov state models into backbone dihedral distributions, contact probability maps and secondary-structure assignments.
For larger proteins without complete trajectories: NMR conformational ensembles, HDX-MS protection factors and Φ-value data serve as partial ordering constraints.
Gō-type energy functions, contact-order folding-rate estimates and statistical contact potentials enforce thermodynamic plausibility across the whole dataset, even without direct experimental supervision.
Each trajectory step is represented as a triple of dihedral-angle distribution, contact probability map and secondary-structure vector. For proteins under 100 residues, typically 10 to 30 discrete states capture the main events of structure formation. The modelling target is a coarse-grained, kinetically ordered trajectory, explicitly not atomistic MD resolution.
Foundation
Pre-trained protein language models such as ESM-2 and ProtT5-XL capture rich co-evolutionary dependencies, but their latent spaces are not organised by free energy and do not encode the directed, downhill nature of folding trajectories.
PathFold adapts these embeddings with an energy-informed contrastive learning scheme grounded in energy-landscape theory. Every conformation is described by residue-resolved geometric descriptors (contact maps, torsion angles, solvent exposure, secondary-structure states). A free-energy surrogate, the fraction of native contacts Q combined with a statistical contact potential, assigns a scalar energy to each conformation. Conformations with similar Q form positive pairs, distant ones negative pairs, and an InfoNCE loss in the residue-resolved embedding space pulls energetically adjacent conformations together.
The result is a thermodynamically ordered latent manifold in which neighbouring points represent plausible sequential steps along a folding pathway, the geometric foundation for PathFold-P. A lightweight two-layer MLP projects PLM embeddings (ESM-2 650M: 1280 dims; ProtT5-XL: 1024 dims) into this energy-aligned space.
Trajectory generation model
A transformer-based generative model that predicts coarse-grained folding trajectories from amino acid sequences while preserving kinetic order, intermediate-state plausibility and thermodynamic consistency along the whole pathway.
Misfolding & aggregation model
A probabilistic model that treats misfolding as a divergence from normal folding trajectories and predicts residue-resolved misfolding and aggregation risk with calibrated uncertainty.
Open platform & benchmark
A modular open-source software suite that integrates both models, plus DynBench, the first standardised benchmarking framework with a public leaderboard for sequence-to-trajectory models.
FAIR datasets
A curated, FAIR-compliant sequence–trajectory dataset that harmonises nine open molecular resources into a single versioned, metadata-rich release, a scientific contribution in its own right.
Folded and structurally well-characterised proteins, built from PDB ensembles, NMR conformers and Folding@home trajectory subsets, providing the folding supervision signal.
Intrinsically disordered and amyloidogenic proteins with residue-resolved misfolding and aggregation labels from DisProt, AmyLoad, AmyPro and ClinVar.
Translational relevance
As a concrete health-relevance deliverable, PathFold produces variant-resolved dynamic misfolding maps for amyloid-β (Aβ40, Aβ42), tau and α-synuclein in collaboration with DZNE, the three clinically most important proteinopathies.
Variant-resolved dynamic misfolding maps for amyloid-β (Aβ40/Aβ42), tau and α-synuclein, developed together with DZNE as a proof of concept for dynamic biomarker discovery and variant prioritisation.
Residue-level misfolding and aggregation risk scores for disease-associated variants and post-translational modifications, as molecular descriptors for early diagnosis and risk stratification.
Trajectory-level stability indicators to support variant interpretation, stability services and precision-medicine oriented protein design.
Outputs
Everything PathFold produces is released under open licenses.