QDX (@qdxco) header
QDX (@qdxco) profile picture

Nature isn't classical, dammit, and if you want to make a simulation of nature, you'd better make it quantum mechanical. — Richard Feynman

Singapore
Joined November 2023
The pKa method with the lowest benchmark error was the worst at identifying reactive cysteines (21%). The method with 3× higher RMSD got it right 57% of the time. General accuracy ≠ biological relevance.
1
33
A Tc of 0.16 (Morgan) can be as statistically significant as 0.44 (RDKit). Raw Tanimoto scores are measurements without units — you need fingerprint, reference molecule, and database context to interpret them. Converting Tc to p-values fixes this.
1
1
20
We ran 16 different naive umbrella sampling simulations. Got 16 wildly different barrier estimates: 18.27 kcal/mol to 30.64 kcal/mol. For covalent drug design, this is the difference between hitting and missing your target.
1
20
We tested this on SARS-CoV-2 inhibitors: - Y180: predicted 25.45 kcal/mol, experimental 23.27 ✓ - GC376: predicted 22.67 kcal/mol, experimental 21.98 ✓ Small sample, but aligned predictions.
1
11
If you're designing covalent binders, this matters. Accurate barrier predictions = faster iteration = better molecules. See our blog for the full writeup: use.qdx.co/inactivation-rate
10
"The Tanimoto similarity is 0.7" Cool. Which fingerprint? Without that context, you've said almost nothing. And even WITH a fingerprint, 0.7 might mean very different things depending on your molecule and database. Here's how to actually interpret Tc 🧵
1
1
19
TL;DR: ✅ Always report your fingerprint with Tc ✅ Use conditional models for virtual screening, not universal thresholds ✅ Convert Tc to a significance score / p-value for interpretable ranking ✅ Use prospective ROC curves to pick the right fingerprint for your target
1
9
Stop saying "the Tanimoto score is." Start saying what it means. For references and the full writeup see our blog: use.qdx.co/tanimoto-score
7