Benchmarks & validation

How well does it work?

Precomputed runs on well-studied molecules, next to what experiments measured, and the validation gates that were written down before any of them were run. Failed gates are shown as failed.

Validation gates

Criteria were frozen before the runs (docs/pull/VALIDATION.md). An engine that fails a gate ships with an "experimental" label.

GateWhat is testedResultDetail

Proteins

ProteinDirectionPeak tension at 0.1 Å/τ (ε/Å)xβ (Å)ΔLc total (nm)Experiment (approx.)

Unfolding movies

Standard protocol, slowest speed (0.05 Å/τ), 10 runs. Runs are grouped by which block of native contacts breaks first, and the most typical run of each group is shown. Yellow: the first contacts to break. Blue → orange: residues losing their native contacts. Make your own from any result on the simulator.

Titin I27: one pathway (10/10 runs)

In every run the N-terminal residues 1–7 separate from residues 71–77 first, then from 78–84. The terminal strand pairs carry the load, and the core follows.

Ubiquitin N–C: two pathways

7 of 10 runs open at the β1–β5 pair (residues 1–6 ↔ 67–72) first. In 2 of 10, the β1/β2–α1 region (13–18 ↔ 25–30) gives way first; one run matched neither group.

GFP pulling directions (gate 2b)

Protocol P_std (0.5 Å/τ, 20 replicas). Simulated peak tension against the unfolding force Dietz et al. (2006) measured for each pair of attachment points. The model's ranking follows how much chain lies between the attachment points, the signature of drag.

Pulled residuesSpanModel (ε/Å)Experiment (pN)

DNA & RNA

ConstructConditionsΔG fold (kcal/mol)F½ (pN)Δx (nm)Unfold / refold at 100 nm/s (pN)

Hairpin class benchmark (gate 4)

The one fully reported hairpin of Woodside et al. (PNAS 2006), 20R55/4T, has F½ = 13.2 ± 0.7 pN and Δx = 18 ± 2 nm at 23 °C, 200 mM KCl. Its exact sequence could not be retrieved, so the gate uses 30 random 20-bp stems with the same GC content (11 G/C) and loop.