Finding a drug means narrowing millions of possible molecules down to one that binds its target, reaches it in the body, and does little else. Simulation and machine learning help at almost every step, mostly by deciding what is worth making next. Docking and drug carriers are explained on the drug delivery and docking page.
How tightly a drug binds is measured by the dissociation constant Kd, which is the concentration at which half of the targets are occupied. It is set by the binding free energy.
Every 1.4 kcal/mol is a factor of ten. A typical hit binds at a micromolar Kd (about −8 kcal/mol), and a drug at a nanomolar Kd (about −12). Therefore, the difference between a weak hit and a drug is only a few kcal/mol, which is smaller than the error of most scoring functions. This is why careful free energy methods, such as alchemical FEP, are used in lead optimization, where an accuracy of about 1 kcal/mol pays off.
Try: set the ligand to 10−7 M, and move ΔG from −8 to −12. The occupancy goes from almost nothing to almost full.
A library may hold millions of compounds, and only a handful of them bind. Docking or a machine learning model scores them all, and only the top of the list is tested. The score does not need to be accurate in absolute terms. It only needs to put more binders near the top than chance would, and the enrichment factor measures exactly this.
Two habits matter. First, test the method on known actives and decoys before trusting it on new compounds. Second, split the data by scaffold when training ML models, or the reported performance will be too optimistic, as I found in project 1.
Try: at r = 0, the curve follows the diagonal (random). At r = 0.5, the top 5% already holds a large share of the binders, even though the scatter looks poor.
A molecule that binds strongly but does not dissolve, crosses no membranes, or blocks the hERG channel is not a drug. Lead optimization balances several properties at once, such as potency, solubility, permeability, metabolic stability, and safety. Many of them pull against each other through lipophilicity, so there is rarely one best molecule, but rather a front of good compromises.
Simple rules, such as Lipinski's rule of five (molecular weight under 500, logP under 5, and few hydrogen bond donors and acceptors), are a first filter. QSAR and ML models predict the rest, and free energy calculations check the most promising ideas before synthesis.
Try: move the weight from potency to solubility, and watch the chosen candidate slide along the front. If you raise the solubility limit, the most potent candidates drop out.