Model and methods

CaDRiN predicts drug response using compound structure, cellular context, transcription factor activity and response patterns learned from training cell lines.

Response patterns learned during training

Two-way main effects describe overall drug potency and cell sensitivity. A rank-8 response profile captures interaction patterns in the training response matrix.

Molecular and cellular features

Morgan fingerprints describe compound structure; ChemBERTa-MTR embeddings help estimate responses for unseen drugs. Context features describe the cellular setting, and VIPER estimates transcription factor activity. Each dataset uses its established Context and VIPER inputs.

Conditional drug-cell modeling

The conditional network combines drug, Context and VIPER representations, including cell-conditioned drug features, feature products and feature differences.

Predictions for known and unseen entities

The model selects a prediction route according to whether the drug and cell appeared in training. Known entities use learned main effects and response profiles; unseen entities obtain estimates from their molecular or cellular features.

Calibration and response correction

Validation data determine response-profile calibration and fusion weights. Multiresolution response correction applies to pairs whose drug and cell both appeared in training. The combined output is predicted ln(IC50).

The lower panel shows how the contextualized network produces Context features. For uploaded samples, the server applies the existing transforms and trained networks.

Evaluation & datasets

GDSC

Training and internal evaluation. Five holdout settings test random pairs, unseen cells, unseen drugs, unseen scaffolds, and unseen scaffolds together with unseen cells. Each run reports pooled Pearson correlation (PCC) and root mean squared error (RMSE); results are summarized across runs.

CCLE, CTRP1, CTRP2 and NCI60

External evaluation on independent drug-response datasets, using each dataset's established cellular features.

TCGA

Precomputed predictions for 8,248 tumor samples, 32 cancer types and 417 compounds. Clinical and molecular annotations support cohort comparisons and association analyses.

See Scope and data handling for interpretation across datasets and experimental endpoints.

Getting started

• Cell-line Prediction: choose catalog cell models and compounds, or add a new compound by name, PubChem CID or SMILES.
• TCGA Prediction: explore existing predictions through Drug response, Group comparison, Associations or Survival analysis. Expand a category, choose a tool and submit its form.
• External-sample Prediction: upload your sample data, check the inputs, choose compounds and run the prediction.

You do not need measured drug responses to submit a prediction. Each workflow provides templates or examples.

Cell-line Prediction

1. Choose a dataset

   Select a cell model catalog. Load example fills in a selection you can run immediately.

2. Select cells and compounds

   Every selected cell model is paired with every selected compound. Filter cells by Tumor type / tissue or search by name or ID. Filters keep your earlier selections; select the count above a catalog to review them.

3. Run the prediction

   Select Predict response. View the table, ranking or matrix, and open a pair to inspect its prediction components and reference distribution. Use Edit inputs to change the selection.

New compounds

Open Search compound and enter a compound name or PubChem CID. Check the returned name, structure and salt form, then select Add compound. Lookup requires a connection to PubChem.

Alternatively, open Add compounds by SMILES and enter one structure per line. You can add a name before the SMILES, separated by a tab. The server generates Morgan fingerprints and ChemBERTa-MTR embeddings before prediction.

External-sample Prediction

Upload gene expression, copy-number alterations (CNA) and somatic mutations for the same samples. The current preprocessing workflow supports human bulk tumor data.

Prepare your files

Use CSV or TSV matrices with genes in rows and samples in columns. Name the first column gene_id. Use the gene symbols in each template and identical sample IDs across all three matrices. Row and column order may differ.

Gene coverage. CaDRiN uses a 1,521-gene panel for CNA and mutations. The upload templates contain the 1,514 genes retained from that panel in the existing TCGA preprocessing workflow. Include every gene in each template. Expression requires its own, larger background gene list; do not restrict it to the CNA and mutation panel.

Required measurements must be present and numeric. Extra genes are reported and excluded. Raw counts, FASTQ/BAM, continuous copy-number ratios and variant files require upstream preparation. The public TCGA example uses its original processed expression scale.

Run a prediction

1. Upload the files and select Check inputs.
2. Review sample matching and gene coverage, then continue to Prepare & predict.
3. Choose compounds and select Prepare features & predict.

You can submit a single sample. CaDRiN applies the existing preprocessing models to each sample to generate Context and VIPER features, then predicts drug response. No PCA or model is fitted on the uploaded batch.

Each request supports up to 50 samples and 5,000 sample-compound pairs, with a combined request size of 16 MiB. Choose from 417 GDSC compounds.

You can browse other sections while processing continues. Select Cancel task to stop processing; cancellation may take a moment. Keep the browser tab open to retain access to the task.

View results

Use Table to inspect values, Compound ranking to compare compounds for one sample, or Response heatmap to compare all submitted samples. Sample names are hidden in the heatmap by default; enable Show sample names to display them. Predictions and the generated Context and VIPER matrices are available for download.

Interpret results

Predicted ln(IC50)

The natural logarithm of IC50 expressed in micromolar, on the GDSC2-aligned prediction scale. Lower values indicate greater predicted sensitivity. Negative values are valid and correspond to an IC50 below 1 micromolar.

Response percentile

A prediction's position within a reference population for the same compound. Lower percentiles indicate greater relative predicted sensitivity. Tied values receive their midpoint rank.

Interaction contribution

The sum of the effective profile, conditional and correction contributions. A negative value lowers the prediction relative to the fitted main-effect baseline.

Prediction components

The contribution of each model component after applying the prediction route and fusion weights. The six contributions sum to the final prediction.

These references stay fixed when you change the displayed selection or filters. External-sample rankings order the selected compounds within each sample.

Cell-line result controls

The metric selector updates ranking order, matrix colors and exported values; the table retains final predictions. Response-percentile ranking becomes available after loading references for all result compounds. Results open as a matrix when both multiple cells and multiple compounds are selected, and as a table otherwise.

Drug response

TCGA Prediction contains precomputed responses for 8,248 tumor samples across 32 cancer types and 417 compounds. Expand Drug response and choose a view:

• By cancer: compare cancer-level medians or select a cancer to view individual samples.
• By compound: examine one compound across tumors and clinical subgroups.
• By sample: search a full TCGA sample ID and inspect its response profile, prediction components and TF activities.

Set the cancer, sample type and any Clinical filters, then run the analysis. Use Edit filters to revise a completed query.

Read the heatmap

All selected samples appear in one heatmap. Hover over a cell to inspect its value. Sample names are hidden by default; enable Show sample names to display them. Long heatmaps scroll within the chart.

Predicted ln(IC50)

Raw predicted responses, or cancer medians when no cancer is selected.

Relative response

Subtracts the same compound's median across all 8,248 atlas samples. Blue indicates a value below that reference; orange indicates a value above it.

Interaction contribution

Contribution beyond the fitted main-effect baseline, shown on a symmetric scale around zero.

Color limits use the submitted cohort and compounds. Toggling sample names keeps the scale unchanged. Add up to two clinical annotation tracks; gray denotes unavailable values. The summary table retains raw predictions for the complete cohort.

Group comparison

Expand Group comparison and choose Mutation, Copy number, Clinical characteristics, Expression subtype or Custom groups.

1. Select the cancer and cohort filters.
2. Define two groups and review the included and excluded samples. Each group needs at least two distinct samples, with no overlap.
3. Run the comparison. Select a compound in the ranking, volcano plot or table to inspect its group distributions.

The effect is median(A) minus median(B) in predicted ln(IC50). Negative values indicate greater predicted sensitivity in group A. All 417 compounds are included. Search and highlight thresholds change the display while preserving the full comparison.

Associations

Analyze associations between predicted drug response and immune signature scores or transcription factor activity.

1. Expand Associations and choose Immune signatures or TF activity.
2. Select a cancer, compound and feature. Apply clinical filters and choose adjustment or scope settings if needed.
3. Run the analysis. Explore features for a compound or compounds for a feature, then open a relationship to inspect individual patients.

There are 28 immune ssGSEA signatures and 277 TF activities. A negative correlation means higher feature values accompany lower predicted ln(IC50). The scatterplot can show original values or residual ranks after adjustment.

Across cancers analyzes each cancer separately. Select a heatmap cell to open its cancer-specific result. The correlation scale stays fixed from -1 to +1.

Survival analysis

Open Survival analysis from the TCGA directory or a LUAD compound result. Choose a compound, cohort filters and overall survival (OS) or progression-free interval (PFI). The analysis uses LUAD primary tumors. Review patient, event and exclusion counts before running it.

The Cox model reports the hazard ratio per one-unit increase in predicted ln(IC50). A ratio above one indicates a higher event hazard as the predicted value increases. Kaplan-Meier curves compare samples with lower and higher predicted values using the cohort median as the cutoff.

Downloads and figure export

Data

• Cell-line Prediction: download full-precision predictions as CSV or JSON.
• External-sample Prediction: download prediction CSV/JSON and the generated Context and VIPER matrices.
• TCGA Prediction: Download data includes the submitted cohort, predictions, annotations, analysis settings and plotted values. Analysis-specific files contain group definitions, association inputs or survival endpoints and diagnostics. Download full matrix provides all atlas samples and compounds.

Figures

Select Export beside a chart for SVG, PDF or PNG. Choose the figure width and font size; for PNG, also choose the resolution. SVG and PDF retain editable Arial text. Rankings and heatmaps can include all rows and sample names. Long PDFs continue onto additional pages; use SVG or PDF when a PNG exceeds the image-size limit.

Share a TCGA analysis

Copy link preserves submitted filters, group definitions and display settings so the analysis can be reopened. Public TCGA sample IDs in custom groups are included in the link. Unsaved changes are stored separately as a browser-tab draft; restoring a draft opens its inputs without running an analysis.

Scope and data handling

Interpretation

CaDRiN predictions support research and experimental prioritization. Clinical treatment decisions require clinical evidence. Response percentiles describe relative model predictions, and group comparisons describe differences in predicted response. Survival associations describe prognosis; they do not estimate treatment benefit from a compound.

TF activity is a model input, and immune signatures share expression-derived information with the model. These associations do not establish independent treatment validation, feature importance or causality. Immune scores measure signature enrichment, without calibration as cell fractions or immunotherapy-response probabilities. Prediction components explain the numerical calculation of a response.

Dataset endpoints require separate interpretation: NCI60 GI50 measures growth inhibition, while the prediction scale is aligned with GDSC IC50.

Submitted data

The web server processes prediction inputs. Uploaded molecular matrices and sample information remain in memory during processing and are excluded from browser drafts and shared links. Uploaded samples are treated as unseen, regardless of their names. The original catalog and TCGA data remain unchanged. Compound lookup sends the entered name or CID to PubChem.

Result availability

• Prediction downloads: cell-line and external-sample CSV/JSON results share a server cache of up to 16 result sets. Downloads expire after one hour, or sooner when the cache fills.
• External-sample tasks: completed tasks and generated features remain available for up to one hour, with up to eight task records retained. Older completed tasks may be removed as new tasks arrive. Reloading or closing the browser tab loses its task link.
• TCGA analyses: results can be regenerated from the precomputed atlas and the settings in a shared link.

Server restarts clear temporary tasks and prediction downloads. Save results you want to keep. Catalog and TCGA input drafts belong to the current browser-tab session and can be cleared from the recovery notice.