Morroniside was the compound that looked promising in my preliminary Keap1 calculations. It did not keep that position when I compared all five candidates under consistent conditions. Catalpol and geniposide were the two I retained for experimental testing.
Keap1 helps regulate the cellular response to oxidative stress. Another protein, Nrf2, interacts with a groove on its surface. I modelled small molecules near that groove to ask which ones remained there consistently enough to justify a direct binding experiment.
A favourable starting position was only the beginning of that comparison. The protein moves. Water moves around it. A molecule that initially fits the groove can turn away or leave it as the simulation proceeds.
Rebuilding the comparison
The preliminary work used a nonhuman Keap1 structure and inconsistent representations of ligand charge. I revisited the full set of five compounds with the same human receptor structure and a common method for assigning charges. Each compound then had three simulations, with each run following 200 nanoseconds of modelled motion.
Those nanoseconds describe time inside the molecular model. They are not the time the computer spent running. Five compounds, three runs each, gave 15 trajectories to compare.
Several things changed together: the receptor, charge treatment, duration, and replication. I cannot assign the change in the shortlist to repetition alone. The narrower finding is that the early preference for morroniside did not survive this revised comparison.
The repeats also shared an important limitation. For each compound, they began from the same top-ranked docking pose, with different stochastic seeds. They let me examine variation in subsequent motion; they did not explore every possible starting arrangement.
A contact can survive while the pose changes
I looked at movement relative to the starting pose alongside contact with the pocket. Either measure by itself could give an incomplete impression.
In one of morroniside's three runs, pocket-anchor contact remained frequent even though the molecule changed its orientation substantially. The other two runs showed larger displacement. Describing all three as the same outcome would lose that variation.
Coptisine offered a different example. It ranked third among the five compounds by its starting docking score, then showed large displacement in all three simulations. That made the initial score insufficient for deciding which compounds to test first in this study.
Catalpol and geniposide stayed near the pocket more consistently. To summarize changes in pose, I aligned the protein and measured ligand RMSD, which describes deviation from the reference arrangement. I first took a median within each run, then compared the three run summaries. Thousands of saved frames from one continuous simulation are not thousands of independent repeats.
Keeping two candidates
I then used saved structures from catalpol and geniposide for an endpoint energy calculation. This estimates an energy difference from the modelled protein–ligand complex and its separated components. It is a further computational check, with assumptions of its own.
The mean was slightly more favourable for geniposide, but the ranges of the three run-specific means overlapped. I kept both candidates. That small average difference did not give me a sound reason to declare one the winner.
The energy calculation omitted entropy and covered only these two retained compounds. It therefore could not supply a complete binding free energy or a new energy ranking of the original five.
I published the study in In Silico Pharmacology on 18 August. Its practical result is an order for follow-up work: test whether catalpol and geniposide bind Keap1 directly, then examine competition with the relevant Nrf2 peptide. A cellular response alone would not settle that binding question. Those experiments were outside this computational study.
Read the paper · My detailed account, with the numerical comparisons