Deep Learning for ADMET Prediction: How to Evaluate Models and Results

What does deep learning add to ADMET prediction?

Instead of applying only fixed rules, a trained model learns relationships between molecular representations and labeled properties. Graph-based models represent atoms and bonds and learn features useful for a particular prediction task.

The published ADMET-AI study, for example, describes Chemprop-RDKit graph neural networks trained on 41 Therapeutics Data Commons datasets. Those tasks include both classification and regression. They do not form a single universal measure of whether a compound is suitable for development.

For a researcher, the useful question is therefore specific: can this model help prioritize compounds for the endpoint we need to investigate? Predicting aqueous solubility and predicting a toxicity-assay label require different interpretations and validation evidence.

Start with the decision, then select the endpoint

Before running a library, identify what you need to decide. “Choose the best compound” is too broad. “Prioritize compounds for a solubility experiment” gives the prediction a defined role.

Use this sequence to structure the review:

• Define the decision. Specify the experiment, screening step, or research question the prediction will inform.
• Match the endpoint. Check the target label, assay context, units, and any transformation of the reported value.
• Check the structure. Retain the submitted SMILES and document how salts, charge states, and stereochemistry were handled.
• Inspect model evidence. Examine validation results relevant to that endpoint and chemistry.
• Choose the next check. Record the measurement or source needed before acting on the prediction.

This makes a screening table useful to the colleague who receives it: each result has an interpretation and a next step, rather than an unexplained red or green flag.

How should you evaluate an ADMET model?

Start with endpoint-level evidence. Model evaluation asks how well predictions match known outcomes across a relevant test set. Reviewing one response asks whether its fields have been interpreted correctly. An overall leaderboard position can help identify models to investigate, but it cannot tell you how a particular endpoint will perform on your compound series.

The TDC ADMET benchmark uses scaffold splits and different metrics for different tasks. Scaffold splitting separates compounds by structural scaffolds, making the evaluation more relevant to structural generalization than a random division alone. It still does not establish performance on every new chemical series.

Metrics answer different questions. AUROC assesses classification ranking across thresholds. AUPRC describes precision–recall performance and should be read with the positive-class prevalence in mind. MAE measures average absolute numerical error on the evaluated scale; Spearman correlation measures rank agreement. None is a confidence score for one submitted molecule. TDC specifies which metric it uses for each benchmark endpoint.

For a project evaluation, reserve compounds with relevant measured endpoints, keep them separate from model development, and compare predictions with those measurements under compatible definitions and units. Document the set and its limitations. If training-set overlap is unknown, avoid presenting the exercise as fully independent validation.

Tested example: reading an aspirin prediction response

On September 15, 2026, we submitted the public aspirin structure CC(=O)Oc1ccccc1C(=O)O to the ADMET prediction tool in Chemical Molecular MCP. The call returned molecular descriptors and endpoint predictions. The values below are rounded from that response.

TDC identifies Bioavailability_Ma as a binary classification task. That distinction matters: even a neatly formatted percentage would be misleading if it turned the class score into a measured biological quantity.

This test demonstrates retrieval and interpretation of an actual response. It does not validate predictive accuracy. We did not compare the output with an independent assay dataset or establish whether aspirin appeared in the model’s training data.

The response also did not include a deployed model version, uncertainty estimate, or applicability-domain flag. We retained those as missing information. A research handoff should preserve that gap rather than silently supplying confidence labels.

The next step would depend on the decision: for a bioavailability-related investigation, first establish the dataset’s label definition and the relevant experimental evidence before using the score to prioritize work.

Why model version and input preparation matter

Two services can name the same model family and still return different values. The ADMET-AI repository explicitly distinguishes its v2 implementation from the v1 implementation used for the publication and live web server. It notes that retraining changes predictions.

Record the model version when available; a connector’s version is not a substitute. In the example above, the response did not identify the deployed ADMET-AI version, so we do not assign one.

Keep the original structure as well as any standardized structure passed to the model. If a result changes, the reviewer needs to distinguish an input change from a model update. For repeated screens, retain endpoint names and raw values before rounding them for presentation.

Using Patsnap Open in an AI-assisted ADMET review

A useful AI workflow connects a prediction to the question your team is trying to answer. Supply the compounds, the endpoints of interest, and your review criteria; then ask the assistant to organize the returned values without turning predictions into experimental findings.

Chemical Molecular MCP documents ADMET-AI prediction alongside structure search, chemical record retrieval, and retrieval of structures associated with patent documents. That combination is useful when a compound review also needs chemical identity or patent context. ADMET-AI supplies the predictive models; Patsnap Open supplies the connected tool access and related chemical-data capabilities.

Model Context Protocol, or MCP, connects these tools to a compatible AI application. In a supported Claude Desktop or Codex configuration, the assistant can call the prediction tool and organize its output against criteria you provide. The language model is not itself the ADMET predictor.

An example review request after connection is:

Run ADMET prediction for these SMILES. Preserve the input structures and raw endpoint names. Separate descriptors, classification scores, and regression outputs. Identify missing units, model-version information, or uncertainty information. Organize the results against my stated screening criteria and list the experimental questions that remain unresolved.

Use the official Chemical Molecular page to obtain the connection instructions for your application. The setup prompt in the sidebar can guide that configuration; the research request above is for use once the tools are available.

What should the screening handoff contain?

A concise handoff should let another researcher inspect the basis for the next step:

• Compound and input: identifier, submitted structure, and preparation choices.
• Prediction: endpoint, raw value, interpreted scale, tool, date, and model version if available.
• Decision context: the criterion being considered and why that endpoint matters.
• Follow-up evidence: measurements to inspect or obtain, and unresolved applicability questions.

Keep measured findings in separate columns from predictions. A compound with an apparently favorable output may still need testing; a concerning output may justify a targeted check rather than automatic exclusion. The research team should define those actions for its assay strategy and project context.

Frequently asked questions

Is deep learning always better than other ADMET prediction methods?

No single model class should be assumed best for every endpoint or dataset. Compare methods on compatible data, splits, and metrics, then evaluate whether the evidence is relevant to your compounds and intended use.

Does an ADMET score of 0.9 mean a compound is 90% safe?

No. Establish which endpoint and class the score refers to. A classification score is not an overall safety assessment, and its numerical value does not establish a calibrated probability for your compound.

Can Claude or Codex predict ADMET properties through MCP?

In a supported MCP setup, they can invoke a connected prediction tool and help review its response. Chemical Molecular MCP documents an ADMET-AI prediction tool. The prediction comes from that tool’s models, while the assistant organizes the workflow and interpretation.

Can predicted ADMET results replace experiments?

Predictions can inform screening priorities and experimental questions. They do not establish measured ADMET behavior or clinical safety. The evidence needed depends on the endpoint and decision.