# ADMET prediction Send a small-molecule shortlist to ADMET-AI, Admetica, eToxPred, and structural alert screening, then compare predicted absorption, metabolism, and toxicity side by side. [Run ADMET-AI](https://proteiniq.io/app/admet-ai)[Run ADMET prediction](https://proteiniq.io/app/workflows/templates?template=admet-toxicity-triage)[Contact sales](https://proteiniq.io/contact) ## More small-molecule discovery workflows ## Tools for ADMET prediction Use these methods to prepare inputs, run the core analysis, inspect outputs, and validate the evidence described in this workflow. [![ADMET-AI](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fadmet-ai.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### ADMET-AI Predicts 41 ADMET endpoints with DrugBank percentiles, physicochemical properties, and alert counts.](https://proteiniq.io/app/admet-ai)[![Admetica](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fadmetica.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Admetica Adds 22 Chemprop endpoints for metabolism, clearance, permeability, and hERG as a second model.](https://proteiniq.io/app/admetica)[![Boltz ADME](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fboltz-adme.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Boltz ADME Predicts lipophilicity, permeability, and solubility with a hosted model.](https://proteiniq.io/app/boltz-adme)[![eToxPred](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fetoxpred.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### eToxPred Scores general toxicity and synthetic accessibility from molecular fingerprints.](https://proteiniq.io/app/etoxpred)[![Structural alert screening](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Ftoxicity-prediction.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Structural alert screening Screens compounds against PAINS, Brenk, and NIH structural alert catalogs.](https://proteiniq.io/app/toxicity-prediction)[![Molecular descriptors](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fmolecular-descriptors.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Molecular descriptors Calculates RDKit physicochemical properties, QED, and rule violations for context.](https://proteiniq.io/app/molecular-descriptors)[![Lipinski's rule of 5](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Flipinskis-rule-of-5.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Lipinski's rule of 5 Checks molecular weight, logP, hydrogen bond donors, and acceptors against the oral-drug limits.](https://proteiniq.io/app/lipinskis-rule-of-5)[![Veber's rule](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fvebers-rule.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Veber's rule Checks rotatable bonds and polar surface area as a flexibility and polarity screen.](https://proteiniq.io/app/vebers-rule)[![Lead-likeness filter](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Flead-likeness-filter.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Lead-likeness filter Applies stricter size and lipophilicity limits to compounds meant for optimization.](https://proteiniq.io/app/lead-likeness-filter)[![PAINS filter](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fpains-filter.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### PAINS filter Flags scaffolds that give false positives across biochemical assays.](https://proteiniq.io/app/pains-filter)[![Brenk filter](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fbrenk-filter.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### Brenk filter Reports toxic, reactive, and pharmacokinetically poor fragments.](https://proteiniq.io/app/brenk-filter)[![SDF to SMILES](https://proteiniq.io/_next/image?url=%2Fimages%2Ftools%2Fsdf-to-smiles.webp&w=96&q=75&dpl=dpl_GWnEYP2z6eAGLas7af1mz73qKh8L) ### SDF to SMILES Converts docked or aligned SDF records to SMILES for the prediction tools.](https://proteiniq.io/app/sdf-to-smiles) ## What is ADMET prediction? ADMET prediction is a computational technique for estimating how the body will absorb, distribute, metabolize and excrete a small molecule, and how toxic it may be, from its chemical structure alone. Models trained on measured assay data turn a SMILES string into numbers such as the probability of intestinal absorption, the percentage bound to plasma protein, whether a CYP enzyme is inhibited, or the risk of blocking the hERG channel. Run aspirin through [ADMET-AI](https://proteiniq.io/app/admet-ai) and it returns a 0.96 probability of human intestinal absorption, which sounds excellent until the DrugBank percentile column places it at only the 32nd percentile among approved drugs. A raw prediction says little until it is ranked against compounds that already became medicines, and the better tools report both. Prediction sits between cheap property rules and expensive assays. Rule-based filters count atoms and bonds, predictive models estimate a specific outcome, and laboratory assays measure it. Each step costs more per compound and answers more reliably, so ADMET prediction is mainly used to decide which few compounds deserve an assay. ### When to use ADMET prediction Run a prediction step when a list of candidates is still too long to test: - A docking or similarity screen has produced a shortlist of tens to a few thousand compounds, and you must cut it before ordering or synthesizing anything. - You are comparing analogs in one series and want to see whether a change moves solubility, CYP inhibition or hERG risk. - A hit carries an uncomfortable feature, such as a reactive group or high lipophilicity, and you want evidence before investing in chemistry. - You need a quick profile of a compound set to choose which in vitro assays to run first. ### Strengths and limitations One ADMET-AI job accepts up to 1,000 molecules and returns 41 model predictions, eight physicochemical properties and three alert counts for each, so a whole shortlist can be profiled before a single compound is ordered. Because the models are independent programs, the same structure can be sent through several of them, and disagreement between them is itself information. The models learn from public assay datasets that differ in size, species and noise. A structure far from that chemistry still receives a confident-looking number. Predictions rank compounds and flag liabilities; they do not replace measurement, and a clean profile does not establish safety. ## Choosing an ADMET prediction tool The hosted tools answer different questions, so the choice depends on what the next decision needs. ADMET-AI is the usual starting point; the others add a second model, a narrower endpoint set, or a different kind of evidence. | Question | Tool | What it returns | | --- | --- | --- | | Broad ADMET profile with context for a shortlist | ADMET-AI | 41 endpoints trained on Therapeutics Data Commons datasets, eight RDKit properties, three alert counts, and a DrugBank percentile for every one of those 52 values | | A second model for metabolism, clearance and hERG | [Admetica](https://proteiniq.io/app/admetica) | 22 Chemprop models covering Caco-2, solubility, P-gp, plasma protein binding, volume of distribution, CYP inhibitors and substrates, clearance, half-life, hERG and LD50, with a selectable subset | | Only lipophilicity, permeability and solubility | [Boltz ADME](https://proteiniq.io/app/boltz-adme) | LogD, permeability and a solubility flag from Boltz's hosted model, up to 1,000 molecules per job | | General toxicity and ease of synthesis | [eToxPred](https://proteiniq.io/app/etoxpred) | A Tox-score and a synthetic accessibility score, both on a 0 to 1 scale | | Substructure red flags | [Structural alert screening](https://proteiniq.io/app/toxicity-prediction) | PAINS, Brenk and NIH alert counts, with the matched patterns and atom mapping | Start with ADMET-AI, because its percentiles give every number a reference point and it covers the most endpoints. Add Admetica when the decision turns on a CYP call, clearance or hERG, and read the two side by side. Both are Chemprop graph neural networks, so they can share blind spots, and agreement is weaker evidence than it looks. Boltz ADME is priced per molecule, at 20 credits each, where ADMET-AI quotes 50 credits for 1,000 molecules. Use it when a hosted LogD and permeability model is specifically what you need, not as the default for a large shortlist. eToxPred was trained on 4,550 compounds with an optimal cutoff of 0.58 in its validation, so treat it as a coarse toxicity flag that also tells you whether a compound will be hard to make. ## Rule-based filters and ADMET models Rule-based filters are not ADMET predictions. They apply fixed cutoffs to calculated properties, which makes them fast, transparent and a reasonable first screen. [Lipinski's rule of five](https://proteiniq.io/app/lipinskis-rule-of-5) flags poor oral absorption when more than one of four limits is exceeded: molecular weight above 500 Da, logP above 5, more than 5 hydrogen bond donors, or more than 10 acceptors. Atorvastatin breaks two of them, at 558.65 Da and a logP of 6.31, and is still sold as a tablet. [Veber's rule](https://proteiniq.io/app/vebers-rule) adds flexibility and polarity, limiting rotatable bonds to 10 and polar surface area to 140 Ų. The [lead-likeness filter](https://proteiniq.io/app/lead-likeness-filter) is stricter on purpose, with molecular weight at or below 350 Da and logP at or below 3, because compounds grow heavier and greasier during optimization and a hit that already sits at the limit has no room left. [Molecular descriptors](https://proteiniq.io/app/molecular-descriptors) calculates the RDKit properties behind all three, along with QED and a synthetic accessibility score. Two filters look for structure instead of size. The [PAINS filter](https://proteiniq.io/app/pains-filter) catches scaffolds that give false positives across many biochemical assays, a pattern first catalogued by [Baell and Holloway](https://doi.org/10.1021/jm901137j). The [Brenk filter](https://proteiniq.io/app/brenk-filter) matches 105 patterns for toxic, reactive and pharmacokinetically poor fragments. An alert is a prompt to look, not a verdict: aspirin triggers one Brenk alert and is an approved drug. Property rules help, but they do not carry the whole load. An [analysis of attrition](https://doi.org/10.1038/nrd4609) at AstraZeneca, Eli Lilly, GlaxoSmithKline and Pfizer linked physicochemical properties to clinical failure for safety reasons, and also concluded that tighter control of properties alone is unlikely to cut attrition much further. Predictive models for specific endpoints are the next layer, not a replacement for the first. ## Where ADMET prediction fits in a screening workflow In [virtual screening](https://proteiniq.io/use-cases/virtual-screening), ADMET prediction normally comes last among the computational steps. A campaign prepares the library, applies the cheap property and alert filters, docks what remains, checks the poses, and only then profiles the shortlist. Docking scores and ADMET outputs answer different questions, so they stay in separate columns instead of being averaged into one number. The order is a matter of economics. In a [high-throughput virtual screening](https://proteiniq.io/use-cases/high-throughput-virtual-screening) run over a very large library, rule-based filters and alert checks cut the set before docking because they cost almost nothing per compound. The slower predictive models then run on the few hundred to thousand compounds that survive docking, which also matches the 1,000-molecule size of a single ADMET-AI job. After [structure-based virtual screening](https://proteiniq.io/use-cases/structure-based-virtual-screening) with a docking engine such as GNINA, the poses arrive as SDF files. ADMET-AI and eToxPred read SMILES, so convert the selected records first with [SDF to SMILES](https://proteiniq.io/app/sdf-to-smiles), keeping the compound identifiers that tie each prediction back to its pose. ADMET prediction also applies earlier and later: to purchasable libraries before screening, and to analog series during lead optimization. ## How to run ADMET prediction online The featured workflow sends one shortlist to four predictors in parallel and keeps every result tied to the compound that produced it. 1. **Standardize the shortlist.** Collect one SMILES per compound with a stable identifier, remove salts and counter-ions so the parent structure is predicted, and convert docked SDF poses to SMILES. Duplicates and invalid structures are easier to remove here than to explain later. 2. **Run ADMET-AI for the broad profile.** Keep DrugBank percentiles on. If the compounds are meant to join a therapeutic class, set the ATC filter to it so the percentiles compare against relevant drugs instead of all approved ones. Check how many molecules were valid. 3. **Run Admetica for the metabolism and clearance endpoints.** Deselect the models the decision does not need, so the output stays short, and keep the native units for each column. 4. **Add eToxPred and the structural alert screen.** Read the Tox-score against its 0.58 cutoff, and open the matched patterns to see what triggered each alert instead of relying on the counts. 5. **Merge the results by identifier and look for disagreement.** Never average values from different models. Mark compounds where the models disagree, or where one strong liability sits next to otherwise good numbers. 6. **Pick compounds for assays.** Choose for chemical diversity as well as for good predictions, and order the assays that test each flagged endpoint, such as microsomal stability, Caco-2 permeability or hERG. ## How to read ADMET prediction results Each column is one of two kinds. A classification endpoint gives the probability that a molecule has a property, such as crossing the blood-brain barrier or blocking hERG. A regression endpoint predicts a value in the unit shown, such as log mol/L for solubility or L/kg for volume of distribution. Regression values are returned as the model produces them and can fall outside a physical range. The public aspirin example in ADMET-AI predicts a half-life of -19.7 hours, which is a sign that the model has little to offer for that molecule, not a measurement. Percentiles compare predictions with predictions, not with measured drug properties, and the direction that counts depends on the property. A high percentile is good for absorption and solubility and bad for hERG or DILI. Aspirin's DILI probability of 0.67 sits at the 64th percentile, so roughly a third of the reference drugs score higher on the same model, and the raw 0.67 is not a liver-injury verdict. Treat disagreement as the useful result. If ADMET-AI and Admetica both flag CYP3A4 inhibition for a compound, plan an interaction assay early. If they split, the compound needs the measurement more than either number does. If both agree that a compound is unremarkable, that is a reason to move on, not a guarantee. ## Limits of in-silico ADMET prediction Data size varies enormously by endpoint. The ADMET-AI CYP inhibition models learned from 12,000 to 13,000 compounds each, while DILI, carcinogenicity and hERG came from only 278 to 648. Several endpoints are also species-specific, such as rat LD50, rodent carcinogenicity and mouse skin reaction, so a number from them does not translate directly into human risk. The [ADMET-AI paper](https://doi.org/10.1093/bioinformatics/btae416) reports the highest average rank on the TDC ADMET Leaderboard at publication, but that benchmark measures held-out public data, not your chemical series. Domain coverage is the second limit. Macrocycles, peptides, [PROTACs](https://proteiniq.io/guides/protein-degraders-clinical-trials) and organometallic compounds lie far from the drug-like chemistry the models saw, and the tools still return confident-looking numbers with no applicability-domain warning. Predictions also describe the parent structure alone. Metabolites, dose, formulation and tissue exposure decide whether a liability matters, and no structure-based model sees them. Use predictions to rank and flag within a series, and confirm the endpoints that decide a project with assays. A good in silico profile narrows a list; it does not show that a compound is absorbed, stable or safe. ## How ADMET prediction works The featured workflow sends one standardized shortlist to four independent predictors. Their outputs keep their native scales, so a flag from one model can be compared with the others without being blended into one score. 1. **Standardize the shortlist.** Collect one SMILES per compound with a stable identifier, remove salts, and convert docked SDF records so every prediction can be traced back to its pose. 2. **Profile the compounds with ADMET-AI.** Predict the broad ADMET panel and compare each value with DrugBank percentiles, optionally narrowed to a therapeutic class. 3. **Cross-check with Admetica.** Run the 22 Chemprop endpoints that matter for metabolism, clearance, and hERG, and read them next to the ADMET-AI profile. 4. **Screen toxicity with eToxPred and structural alerts.** Score general toxicity and synthetic accessibility with eToxPred, then open the matched PAINS, Brenk, and NIH patterns in the structural alert screen. 5. **Compare results and choose assays.** Merge the outputs by compound identifier, mark disagreements and strong liabilities, and select compounds and confirmatory assays. ## Inputs and outputs Check formats before running, then inspect and download the result from every workflow step. ### Inputs #### Compound shortlist `SMILES``CSV``SDF` One structure per compound with a stable identifier. Convert docked SDF poses to SMILES first, and keep the parent structure rather than a salt form. #### Endpoint and reference choices Decide which Admetica endpoints you need and whether the ADMET-AI percentiles should compare against a therapeutic class instead of all approved drugs. #### Decision thresholds Write down the property cutoffs and alert rules that exclude a compound before looking at the results, so the review stays consistent across the shortlist. ### Outputs #### ADMET profile table `CSV` One row per molecule with probabilities and regression values in their stated units, plus DrugBank percentiles from ADMET-AI. #### Second-model endpoints `CSV``JSON` Admetica predictions for the selected endpoints, with the run log and version records that document how they were produced. #### Toxicity and synthesis scores `CSV` The eToxPred Tox-score and synthetic accessibility score for each compound, which are ranking aids rather than verdicts. #### Structural alert matches `CSV` PAINS, Brenk, and NIH alert counts with the matched pattern names and atom mapping for each compound. ## Frequently asked questions AI answers from our docs. No sign-in needed. ## Run your first analysis today Start with 200 free credits and no installs. Upgrade when your project needs more compute. [Start for free](https://proteiniq.io/sign-up)[View pricing](https://proteiniq.io/pricing)