Protein Folding Experiments

Overview

What is this? A set of experiments at Erdős AI Lab sitting between deep learning and structural biology - characterizing how modern transformer-based folding stacks behave on small proteins, where they succeed, where they fail, and whether their internal confidence signals actually track real structural fidelity.

And the second thread? A head-to-head against classical molecular-dynamics baselines on the same targets, looking for systematic biases on either side - the kind that only show up when learned and physics-based approaches are scored against each other in earnest.

Status: Active and ongoing - February 2026 to present. Now live: head-to-head between ESMFold and the native RCSB structure, plus folding-trajectory dynamics (Q, Cα RMSD, R_g, Q-RMSD landscape) on a 1000 ns trajectory. In active development: expanded target sets (multi-chain complexes, intrinsically disordered regions), pLDDT calibration sweeps across model sizes, and a head-to-head against classical MD baselines on the same trajectories. New write-ups, target lists, and evaluation protocols will be added here as each track matures.

Threads

pLDDT-style confidence calibration. Do learned confidence scores actually track RMSD/lDDT on held-out structures? When do they break, and along which axes - loop length, disorder, novelty?
Folding trajectory dynamics. Instead of treating folding stacks as black boxes, inspecting intermediate representations to understand how structure forms layer by layer.
Transformer vs. classical MD. Head-to-head on small targets where classical MD is still tractable, isolating where each approach is reliable.

Selected Results

What are we looking at? A representative head-to-head and a folding-trajectory snapshot from one of the active tracks. The structural comparison shows where single-chain folding stacks land versus the native complex; the trajectory plots are the observables we use to score folding quality over time.

Side-by-side ribbon diagrams: ESMFold predicted protein structure (rainbow N to C) on the left, native RCSB structure (green, with bound nucleic-acid double helix) on the right. — ESMFold prediction (left, rainbow N→C) vs native RCSB structure (right, green with bound nucleic-acid double helix).

Backbone topology Helical bundle is recovered - the major secondary-structure elements line up with the native fold.

What's missing Bound nucleic-acid double helix - not modeled by single-chain folding stacks like ESMFold.

Takeaway Confidence on the protein chain alone is reasonable; complex-level fidelity needs a multi-entity model.

Four-panel folding trajectory: native contact fraction Q over time, Cα RMSD vs native over time, radius of gyration Rg over time, and a Q vs RMSD 2D folding landscape colored by simulation time across 1000 ns. — Folding observables across a 1000 ns trajectory - native contact fraction Q, Cα RMSD vs native, radius of gyration R_g, and the Q-RMSD 2D landscape coloured by time.

Q (native contacts) Hovers near 1 most of the run with brief unfolding excursions - structure is largely native-like.

Cα RMSD Drops from ~5 Å to ~1-2 Å after ~600 ns, indicating settled native-like geometry.

2D landscape Trajectory funnels toward high-Q / low-RMSD over time - consistent with productive folding.

Why It Matters

Hasn't structure prediction been solved? At the surface level, yes - the benchmark numbers are impressive. But reliable use downstream requires knowing when a model can be trusted, and the answer is rarely "always". Calibrated confidence and a mechanistic read on folding stacks are what turn benchmark scores into usable scientific tools.

Nilesh Sarkar / Research

Overview

Threads

Selected Results

Why It Matters

Related