Estimate total-tissue and interstitial exposure across 16 healthy organs, then connect predicted accessible concentration to affinity-based receptor occupancy. Bispecific mode adds separate arm estimates.
Awaiting profile
Start with the HER2 example or enter a target and dosing profile.
Model notes
Shah and Betts showed that, for many antibodies, each organ receives a fairly repeatable fraction of the concentration found in plasma. Their coefficients often estimated tissue concentration within roughly twofold of the observed value. That gives this model a physiological starting point instead of asking it to learn antibody distribution from scratch. Read the 2013 paper.
A fixed organ ratio treats antibodies in the same class alike. This model keeps that baseline, then adjusts it using molecular weight, drug type, dose, plasma exposure, species, target, and tissue. It can explore immunoglobulin G (IgG) antibodies, antibody–drug conjugates (ADCs), bispecific antibodies, smaller antibody fragments, and nanobodies. The result can reflect target-related uptake as well as ordinary antibody accumulation.
A fixed Shah-style lookup predicts each organ from a passive-distribution average. The complete research model keeps that physiological starting point and can correct it using the molecule's format, dose, plasma exposure, species, target identity, and organ protein-expression data. Across the broader drug set, adding that context reduced typical error from 5.17-fold to 1.42-fold.
The advantage held when the test was made harder. With every example for one target withheld, the fixed ratios had 7.25-fold typical error and the model had 1.37-fold error. In the slice closest to Shah's original scope— unmodified immunoglobulin G antibodies in healthy human organs—the fixed ratios had 5.28–5.31-fold error and the model had 1.29–1.32-fold error.
“Typical error” here is the median absolute fold difference between a prediction and its measured value. A 1.42-fold error means that a value predicted as 100 would typically correspond to roughly 70–142 in the same units. Leave-one-out (LOO) testing removes every row for one drug or target, trains on the rest, and evaluates that unseen group; the headline then takes the median within each held-out group and the median across groups. These matched figures belong to the complete research model; the next panel reports validation of the streamlined model running on this page. The comparison covers newer targeted imaging data that extend beyond Shah's original passive-distribution setting.
Organ exposure starts with an estimate of the drug concentration in plasma. The tool combines dose, dosing interval, molecular weight, and typical pharmacokinetic values for the selected antibody format. It estimates clearance as distribution volume × 0.693 ÷ half-life, then calculates average steady-state plasma concentration as dose ÷ clearance ÷ dosing interval. Molecular weight converts that result to nanomolar. A repeated-dose intravenous model supplies the peak and trough. The predicted concentration in each tissue is the plasma concentration multiplied by that organ's Kp. These format-level assumptions provide a useful baseline when drug-specific plasma measurements are unavailable.
The streamlined model on this page had 1.45-fold typical error when asked to predict targets excluded entirely from training. For each test, all examples involving one target were set aside, the model learned from the remaining targets, and the excluded target was then predicted. Repeating this across targets measures performance on unfamiliar targets.
The same result is 0.161 dex. Dex is a base-10 scale used when values span a wide range: 0.3 dex means about twofold and 1 dex means tenfold. At 0.161 dex, a prediction of 100 nanomolar (nM) has a typical error range of roughly 69–145 nM. This is retrospective internal validation using previously collected data. Training used 12,257 target–drug–organ examples derived from a smaller number of underlying molecules and studies.
A tissue measurement can include drug still in blood vessels as well as drug bound to cells, taken inside cells, or tracer left behind after drug breakdown. Cell-surface receptors are more directly exposed to drug in the interstitial fluid (ISF) surrounding cells. A separate model estimates that accessible concentration before occupancy is calculated. Its supporting data are sparser, so the displayed uncertainty range matters.
Two molecules can reach the same tissue concentration but bind very different fractions of their targets. The occupancy calculation combines predicted ISF concentration with the dissociation constant (Kd), the supplied measure of binding strength, to estimate the fraction of target molecules bound by the drug. This matters because exposure alone cannot distinguish strong binding from weak binding. Occupancy follows the tissue distribution prediction as a separate calculation. It assumes one drug binding one target at equilibrium and leaves out multivalent binding, receptor turnover, and clearance caused by target binding.