Hepatic clearance from in-vitro data, with an honest error bar.

Enter CLint, fu,mic and fu,p, or just a SMILES, and get hepatic and plasma clearance by well-stirred or parallel-tube IVIVE, together with how far off a prediction like this has been for real drugs.

Open the calculator Use it from your LLM

No account needed.

Worked example

Midazolam Verapamil Diltiazem Warfarin

CLint 177.8 µL/min/mg · fu,mic 0.88 · fu,p 0.05 · Rb 0.53

Plasma CL: 4.73 mL/min/kg
EH: 0.43
Clearance: medium

Plasma CL, mL/min/kg, log scale

Predicted plasma CL 4.73
Observed CL fell here for half of 301 reference drugs predicted like this: 4.32–27.7
Observed for midazolam: 5.3

CLint 135.6 µL/min/mg · fu,mic 0.43 · fu,p 0.10 · Rb 0.77

Plasma CL: 9.87 mL/min/kg
EH: 0.62
Clearance: medium

Plasma CL, mL/min/kg, log scale

Predicted plasma CL 9.87
Observed CL fell here for half of 301 reference drugs predicted like this: 9.02–57.9
Observed for verapamil: 18.0

CLint 16.7 µL/min/mg · fu,mic 0.76 · fu,p 0.22 · Rb 1.0

Plasma CL: 3.34 mL/min/kg
EH: 0.16
Clearance: low

Plasma CL, mL/min/kg, log scale

Predicted plasma CL 3.34
Observed CL fell here for half of 301 reference drugs predicted like this: 3.05–19.5
No observed plasma CL for diltiazem in the app's reference set

CLint 0.58 µL/min/mg · fu,mic 0.47 · fu,p 0.01 · Rb 0.55

Plasma CL: 0.0101 mL/min/kg
EH: 0.0009
Clearance: low

Plasma CL, mL/min/kg, log scale

No band here: the prediction falls outside the range the band was calibrated on, and the calculator says why. “Renal filtration (fu,p × GFR) rivals predicted hepatic CL. Total CL may be much higher.”

Predicted plasma CL 0.0101
Observed for warfarin: 0.06

Computed by the calculator: human liver microsomes, well-stirred. Inputs from Obach 1999 (Drug Metab Dispos 27:1350); observed plasma CL from the Lombardo et al. 2018 compilation, via PKSmart.

What you get

IVIVE with every intermediate: Scaled and unbound CLint, extraction ratio, plasma and blood clearance, and how far each input moves the answer. Nothing hidden behind one number.
The observed-clearance band: Beside each prediction, the range in which observed plasma clearance fell for half of 301 reference drugs predicted like this one.
A second opinion from structure: Human clearance and Vss from the SMILES alone, with fu,p and fu,inc, and an applicability verdict from similarity to the training set.
Your lab's correction: Fit a correction on compounds your own lab has measured. It narrows the band for your assay rather than someone else's.
Species scaling: Rat, dog or monkey clearance to human by six methods, each listed with the error it showed on 369 drugs.
Target CLint: Work backwards: the CLint an assay must read for a compound to reach a target clearance.

Validation, plainly

IVIVE from human liver microsomes and hepatocytes under-predicts in-vivo clearance by about 3-fold on average. ivive.app does not hide that: every result carries the range observed for comparable drugs, calibrated on 301 reference drugs with published in-vitro and in-vivo data.

On an independent Pfizer dataset (Tess et al. 2023) the 50% band held the observed value 58–68% of the time, never below its nominal coverage.
The band is wide on purpose. It is the method's measured error, not decoration.
Compounds with fu,p below 0.01 are flagged instead of given a range, because no range held for them.
Multi-species allometry scored worst of the scaling methods measured, and the page says so.

Sources: Obach 1999 (Drug Metab Dispos); Tess et al. 2023; Mahmood & Balian 1996; Hallifax & Houston 2019 (Drug Metab Dispos). Each method in the app links its equation and source; training data and licences are on Data sources.

How often the 50% band held, on drugs it never saw

Coloured: 58–68% across subsets of the Tess et al. 2023 Pfizer data. Dark line: the 50% the band promises.

Use it from your LLM

ivive.app is also an MCP server, so Claude, ChatGPT or any MCP client can call the calculations instead of doing the arithmetic itself. No key needed.

predict_inputs
calc_clhep
target_clint
input_sensitivity
similar_drugs
scale_to_human

A structure you send through an LLM also reaches that LLM's vendor. For proprietary compounds, check that the vendor is cleared for them, or ask about a private deployment.

Claude Code:

claude mcp add --transport http ivive https://api.ivive.app/mcp/

Claude Desktop, in claude_desktop_config.json:

"ivive": {
  "command": "npx",
  "args": ["mcp-remote", "https://api.ivive.app/mcp/"]
}

Any other client: the streamable HTTP endpoint is https://api.ivive.app/mcp/, with the trailing slash.

Use the API

Every calculation is a plain HTTP endpoint, the same one the calculator calls. Single compounds need no key; batches of up to 25 need a free account. For bigger batches or more volume, ask for a key at hello@ivive.app.

The full reference, with every endpoint, field and unit: api.ivive.app/docs (try requests in the browser), ReDoc, or the OpenAPI document to generate a client.

Python:

import requests

r = requests.post("https://api.ivive.app/api/ivive/clhep", json={
    "clint": 177.8, "fu_mic": 0.88, "fu_p": 0.05,
    "rb": 0.53, "species": "human",
})
print(r.json()["clp"], r.json()["band"])

Midazolam, with curl:

curl -X POST https://api.ivive.app/api/ivive/clhep \
  -H 'Content-Type: application/json' \
  -d '{"clint": 177.8, "fu_mic": 0.88, "fu_p": 0.05,
       "rb": 0.53, "species": "human"}'

The answer, shortened: plasma CL, extraction ratio, and the band with its basis.

{
  "clp": 4.7278,
  "eh": 0.430938,
  "classification": "medium",
  "band": { "low": 4.3218, "high": 27.7116,
            "coverage": 0.5, "basis": "reference", "n": 301 },
  "upper_bound": false,
  "flags": []
}

Free, and more

Feature	Free	On request
Single-compound calculations	Yes, no account
Batch, numbers or SDF	up to 25 compounds, with a free account	up to 500
Saved lab corrections	1, with a free account	unlimited
MCP for LLMs	30 requests a minute, 300 a day	API key, higher volume
Private deployment		on your own servers; structures never leave your network

Nothing is sold yet. If you need more, write to hello@ivive.app and say what for.

Your data

Structures and inputs you enter are processed in memory and not stored. A lab correction or IVIVC dataset you save keeps its data, visible only to you. An account holds your email address and a password hash. The API runs in Nuremberg (netcup), the account database in Frankfurt (Neon); the privacy policy has the details.