SwissTargetPrediction batch prediction — DeepSeek Harness plugin & agent skill
English | 中文
Want to know which proteins a molecule — or a whole series — might act on? Hand it
the SMILES. It submits each one to
SwissTargetPrediction, waits for the
prediction, and lays the results out as one table: target and common name, UniProt
and ChEMBL ids, target class, probability, and the known actives behind each target
(3D/2D). The same table is written to a CSV in the session workspace.
Note: This tool is unofficial and not affiliated with the
Swiss Institute of Bioinformatics (SIB). Please
respect their server usage policies. If the predictions go into research, cite
Daina, Michielin & Zoete, Nucleic Acids Research 47(W1), W357–W364 (2019).
What you get
- A whole batch in one go. Paste one SMILES per line; submitting, waiting and
collecting the results is the plugin's job.
- Everything the service reports, in one table. Target and common name, UniProt
and ChEMBL ids, target class, probability, known actives (3D/2D).
- Salt forms handled for you. A hydrochloride, a sodium salt or another ionic
form is desalted before submission — and neutralised where that is safe — with
every change listed for you. SwissTargetPrediction takes one molecule per query,
so pasting a salt as-is simply returns nothing.
- Three ways to run it. Ask the agent; type
/swiss in the composer; or open
the SwissTarget panel and paste your list.
- No blind waiting. Progress and elapsed time while it runs, results still there
if you switch away, and a written reason for every molecule that failed.
- Human, mouse or rat.
- Results you can keep. The full table goes to a CSV in the session workspace,
and the panel exports it on demand.
- English and Chinese UI.
Install
From npm, or from a release — nothing to clone:
dsh plugin --profile <profile> add @yumeayai/dsh-swiss-target
dsh plugin --profile <profile> add https://github.com/YumeAyai/dsh-swiss-target/releases/latest/download/dsh-swiss-target.tar.gz
Or from the repository, or from a copy you already have:
dsh plugin --profile <profile> add github:YumeAyai/dsh-swiss-target
dsh plugin --profile <profile> add /absolute/path/to/this/directory
As an agent skill (the same directory is the skill bundle):
git clone https://github.com/YumeAyai/dsh-swiss-target.git ~/.dsh/skills/dsh-swiss-target
Usage
Ask the agent
swiss_target_batch(smiles=["CC(=O)OC1=CC=CC=C1C(=O)O", "CN1C=NC2=C1C(=O)N(C(=O)N2C)C"])
| Parameter | Default | Meaning |
|---|
smiles | — | One SMILES string per molecule, at most 25 per call |
organism | Homo_sapiens | Also Mus_musculus, Rattus_norvegicus |
output_csv | swiss_target_results.csv | Relative to the session workspace, or absolute |
From the composer
/swiss CC(=O)OC1=CC=CC=C1C(=O)O
/swiss {"smiles":["CCO","c1ccccc1"],"organism":"Homo_sapiens"}
One SMILES per line (# starts a comment), or the JSON payload the panel sends.
From the panel
Open a session, choose SwissTarget in the conversation view tabs, paste one
SMILES per line, pick the organism and run. The results appear as a table, and
Export CSV downloads them; the host also writes the whole batch into the
workspace.
While it runs the panel shows progress and elapsed time — 5–60 seconds per molecule
is normal. Switching tabs or sessions and coming back keeps the draft and the
results.
Example output
| Target | Common name | Uniprot ID | ChEMBL ID | Target Class | Probability* | Known actives (3D/2D) | input_smiles |
|---|
| Cyclooxygenase-1 | PTGS1 | P23219 | CHEMBL221 | Oxidoreductase | 0.756 | 13 / 2 | CC(=O)OC1=CC=CC=C1C(=O)O |
Salt forms and mixtures
Most active pharmaceutical ingredients are weighed as a salt — a hydrochloride, a
sodium salt, a tartrate — and SwissTargetPrediction takes one molecule per query:
submitting [Cl-].CC(=O)Oc1ccccc1C(=O)O as-is returns nothing at all. So every line
is converted before it is submitted:
| You submit | Actually submitted | What happened |
|---|
[Cl-].CC(=O)Oc1ccccc1C(=O)O | CC(=O)Oc1ccccc1C(=O)O | chloride counter-ion dropped |
CN(C)CCOC(c1ccccc1)c1ccccc1.Cl | CN(C)CCOC(c1ccccc1)c1ccccc1 | neutral HCl adduct dropped |
CC(=O)[O-].[Na+] | CC(=O)O | sodium dropped, carboxylate neutralised |
[Na+].[O-]C(=O)c1ccccc1O | OC(=O)c1ccccc1O | sodium dropped, phenolate/carboxylate neutralised |
[NH3+]CC(=O)[O-] | NCC(=O)O | zwitterion neutralised |
C[N+](C)(C)CCOC(c1ccccc1)c1ccccc1.[Cl-] | quaternary parent | chloride dropped, the charge stays (no hydrogen to lose) |
CCO.c1ccccc1 | c1ccccc1 | a real mixture: largest fragment kept, the rest reported |
[Na+].[Cl-] | — | refused: counter-ions only |
Your structure is never rewritten beyond that: only counter-ions are dropped, and
only charges that can be neutralised safely are touched — a nitro or a quaternary
ammonium group is submitted as written. The original line and the submitted form
each keep a CSV column (input_smiles, submitted_smiles), and every change is
listed in the tool result and in the panel.
Good to know
- The service throttles. After a run of submissions it refuses new jobs for a
while. The client retries each molecule three times, then marks that molecule as
failed with its reason and finishes the rest. Retry the failed ones later.
- Caps. At most 25 molecules per call; at most the service's own top 100 targets
per molecule.
- Timing. Submissions are sequential with a 2-second pause; one molecule usually
takes 5–60 seconds.
- Organism. Human (default), mouse, rat.
- Predictions, not evidence. Do not present them as measurements.
License
This code is released under the ISC License.
The data and predictions are owned by SwissTargetPrediction / SIB.
A Node.js port of the Python client
swiss_target_batch, which keeps
the original source and its own install instructions.
