# Pharma & R&D for ML Engineers What the key numbers in drug research mean, from toxicity doses and binding strength to potency, drug levels in the body and viral inhibition, with notes on modelling them\. By sunny\-moon\-25a0495f Topic: Human Body & Medicine [Open the interactive Jig](https://ask.jig.com/j/uuLIiyx19V7j) ## Transcript ### 1. A lab bench for model builders Drug research runs on measurements, and every number has a precise meaning\. Around this bench sit six stations, one for each kind of number you will meet in pharma data\. Drag to turn the bench, and use the guide to visit each station\. The molecule above the centre is caffeine, a small drug most of us take every day\. ### 2. Toxicology: dose and response Toxicologists give groups of animals fixed doses and record harmful effects\. The highest tested dose with no harmful effect is the NOAEL, the no\-observed\-adverse\-effect level\. The lowest dose where harm appears is the LOAEL\. Plotting harm against dose gives the dose\-response curve\. The NOAEL can only be one of the doses tested, much like a threshold picked from a coarse grid search\. These column heights are illustrative\. ### 3. Binding: what Kd means A drug works by sticking to a target, often a receptor protein on a cell\. Kd, the dissociation constant, is the concentration at which half of the targets are occupied\. A lower Kd means the drug holds on more tightly\. Slide the concentration and watch the receptors fill\. Occupancy against log concentration is an S\-shaped curve, much like a sigmoid\. ### 4. Potency: IC50 and Ki IC50 is the drug concentration that halves an activity in one particular test\. Add more of the natural substance the drug competes with, and the IC50 moves to higher doses\. Ki is the drug's own binding strength with the test conditions removed, so it stays put\. IC50 is like a score measured on one test set; Ki is closer to the model parameter itself\. Compare IC50 values from different labs with care\. ### 5. ADME: the body's journey After a dose, the body absorbs the drug, distributes it, metabolises it and excretes it: ADME for short\. Blood levels over time form one curve\. Its peak is Cmax at time Tmax, the area under it \(AUC\) is total exposure, and the half\-life is how long the level takes to halve\. Most drugs decay exponentially: after about five half\-lives, roughly 97% is gone\. ### 6. A knowledge graph of findings Papers report facts such as "this gene encodes this receptor" or "this drug binds it"\. A knowledge graph stores each fact as a typed edge between typed nodes, along with its source\. Show the predicted link: a model can suggest an edge, but it is a hypothesis to test, not a finding\. Public sources include DrugBank for drug–target pairs, ChEMBL for measured potencies and Open Targets for target–disease evidence\. ### 7. Viral inhibition In a dish of cells, a virus leaves clear spots called plaques where cells have died\. More antiviral drug means fewer plaques\. EC50 is the concentration that halves the virus; CC50 is the concentration that harms half of the healthy cells\. Add the drug to the plate and compare the columns\. Selectivity index is CC50 divided by EC50; bigger is safer\. Remdesivir in one cell line: EC50 under 1 micromolar, CC50 above 100\. ### 8. The rulebooks Regulators set how these studies are run, so results can be trusted and compared\. OECD Test Guideline 407 describes a 28\-day rat study with at least three dose groups that yields a NOAEL\. FDA guidance then turns a NOAEL into a cautious first human dose\. ICH M3\(R2\) sets which safety studies must come before each stage of human trials\. Think of these as the data spec and the test protocol\. ### 9. From bench to model Each station produces labels for your models: doses, potencies, curves, graph edges and plate counts\. Every number depends on its test, units and species, so keep that context as features\. Then your model learns biology instead of lab quirks\. Turn the bench to revisit any station\.