Phenotype-based activity
Predict antiviral activity at the phenotype level from compound structures.
Predict antiviral activityAntiviral discovery · From phenotype to targets
Explore antiviral potential.
Prioritize candidate targets.
DeepAVC brings phenotype-based prediction and compound–target interaction profiling together. Start with your compounds in SMILES format and choose the predictions that fit your research.
Three prediction modes · Batch input · Downloadable CSV results
Known antiviral targets
Explore interactions with curated antiviral-related targets and their biological annotations.
Explore known targetsAll-target profiling
Screen across the full available target library. Supports up to 5 compounds per task.
Run all-target profilingCite DeepAVC
Kang, B., Zhao, Y., Chen, X. et al. Bridging antiviral drug discovery with a large language model-powered framework. Commun Biol (2026).
Read the publication · About the platformNew prediction
Predict your compounds
Enter SMILES and select the modes you need. You can combine modes in a single task.
Resources
Models & source code
Download the pre-trained prediction models, feature extraction models, and source archive for local research.
These are the files currently provided by the server. File dates describe the download artifacts, not model version numbers. No published release version or checksum is implied by the file date.
DeepPAVC
Model checkpoint
343.96 MiBDownload
DeepTAVC
Model checkpoint
354.76 MiBDownload
CADTI
Model checkpoint
364.77 MiBDownload
KPGT
Model checkpoint
426.40 MiBDownload
ESM-2 650M
Model checkpoint
2.43 GiBDownload
ESM-2 650M contact regression
Model checkpoint
3.60 KiBDownload
ESM-2 15B
Model checkpoint
28.18 GiBDownload
ESM-2 15B contact regression
Model checkpoint
8.48 KiBDownload
Source code archive
Source archive
269.98 KiBDownload
Documentation
Help & interpretation
Get started, understand the three prediction modes, and explore example outputs.
Getting started
- Enter compounds. On the Predict page, paste one SMILES per line, or use “Load 3 examples”. Do not include column headings or compound names.
- Select modes. DeepPAVC predicts phenotype-level antiviral activity; DeepTAVC profiles known antiviral-related targets; CADTI profiles the full available target library. Multiple modes can be combined.
- Run prediction. Save the task link shown after submission. You can revisit it to check status and retrieve your results.
- Explore results. Compare scores within each prediction mode and download the CSV files for further analysis.
Each task supports up to 100 compounds, or 5 when all-target profiling is selected. Each compound is limited to 4096 SMILES characters and 256 atoms.
Start a predictionScore interpretation
- DeepPAVC · Phenotype-based antiviral activity
- One model score in [0, 1] per compound. Higher values indicate stronger predicted antiviral potential at the phenotype level and help prioritize compounds for experimental testing. This score does not identify a specific molecular target.
- DeepTAVC · Known antiviral target interactions
- One model score in [0, 1] per compound–target pair in the curated antiviral-related target set. Higher values indicate stronger model support for an interaction, not a measured binding strength. Target ranks are calculated separately for each compound.
- CADTI · All-target interactions
- One model score in [0, 1] per compound–target pair across the available protein library. Higher values prioritize potential interactions for follow-up. A high target score alone does not establish antiviral activity; target ranks are calculated separately for each compound.
A score of 0.8 does not mean an 80% chance of experimental success. These are model outputs, not calibrated probabilities, measured affinity or potency values (such as Kd, Ki or IC50), or experimental activity measurements. Compare scores within the same model; experimental validation is required.
What do the activity colors mean?
DeepPAVC retains the original display bands: Strong (green) for 0.8–1, Moderate (gold) for 0.5–<0.8, and Weak (gray) for 0–<0.5. These are visual aids, not experimentally validated activity classes or recommended decision thresholds. They do not apply to DeepTAVC or CADTI.
Tables show four decimal places, but colors use the unrounded score. Hover over a score or download its CSV to inspect the underlying value. Historical pChEMBL results use a different scale and must not be interpreted using these bands.
How are scores calculated?
Each model's output is transformed with a sigmoid function to produce a score between 0 and 1. Scores are not rescaled to the highest value in your job and do not have to sum to 1 across targets. Sorting and Top K filters change the display, not the underlying model scores.
Which mode should I use?
Choose DeepPAVC to explore phenotype-level antiviral potential, DeepTAVC to investigate interactions with known antiviral-related targets, and CADTI for broader target profiling. These modes can be used independently or together.
Example prediction outputs
Current results interface, captured from completed demonstration jobs. These screenshots illustrate the layout and controls, not model performance. Open an image to view the full-size screenshot in a new tab.
Phenotype-based activity
Open full-size screenshot
Known antiviral targets
Open full-size screenshot
All-target profiling
Open full-size screenshot
About DeepAVC
Viral infections continue to pose major threats to global health, underscoring the need for effective antiviral therapeutics. Antiviral drug discovery traditionally follows two paradigms: phenotype-based drug discovery (PBDD) and target-based drug discovery (TBDD).
DeepAVC is a large language model-powered unified framework integrating DeepPAVC for PBDD and DeepTAVC for TBDD, enabling complementary antiviral compound prediction. The CADTI module supports large-scale compound–all-target interaction profiling.
Kang, B., Zhao, Y., Chen, X. et al. Bridging antiviral drug discovery with a large language model-powered framework. Commun Biol (2026). Read the publication.
Contact
Peking University Health Science Center
Prof. Qinghua Cuicuiqinghua@bjmu.edu.cn
Dr. Boming Kangkangbm@bjmu.edu.cn