DeepMedChem Python SDK

The official, chemistry-thin Python client for the DeepMedChem hosted chemical-space platform. It contains no RDKit, models, databases, or proprietary search implementation.

Beta: deepmedchem 0.4 is available for early use. APIs may still change before the stable release.

Installation

pip install deepmedchem

For account features, authenticate once, or set DEEPMEDCHEM_API_KEY in automation:

dmc login
dmc status

dmc login prints a short code and an approval URL. On a desktop it opens the URL in your browser; on a headless server, container, or SSH session it only prints the URL, which you can open on any device. Sign in or create a CHEESE account there, approve the connection, and the CLI finishes on its own. The key goes to the OS keyring when one is available, otherwise to a credentials.json file (mode 0600) next to the SDK config. Use --no-browser to force the print-only behaviour and --token-stdin to paste an existing key from a pipe.

Search without an account

When no API key is configured, the SDK uses guest access for catalog, unfiltered single-database similarity search (Morgan, shape or ESP), and sampling. Pass guest=True to ignore saved/environment credentials explicitly; it cannot be combined with an explicit api_key or credential_provider.

import deepmedchem as dmc

# Use the production API without signing in.
result = dmc.search(
    "CC(=O)Oc1ccccc1C(=O)O",
    database="enamine",
    guest=True,
)
print(result.mode)       # "guest"
print(result.warnings)   # server limit/field warnings, if any
result.to_csv("guest_search.csv")

Client(guest=True), AsyncClient(guest=True) and dmc.aio.search(..., guest=True) work the same way. client.is_guest reports whether the client has no credentials. The server enforces the guest policy: by default at most 20 results, no synthon/property acquisition, and per-IP/shared request budgets. Larger requests are clamped with warnings; guest_limit errors retain the server's details (including retry_after and login_url) and are returned immediately without automatic retries. Invalid keys remain authentication errors; the SDK never retries them anonymously.

dmc search "CCO" -d enamine --guest -n 20
dmc sample -d enamine --guest -n 20
dmc databases --guest

Exact substructure search, filtered/conditioned similarity, automatic multi-database routing, selections, batches and optimization require an account. Guest clients fail locally for these operations; sign in and create a client without guest=True. Guest access is available on the production and development APIs. Other deployments must enable guest access explicitly; installing this SDK alone does not enable it on a server.

Command line

The dmc command (also installed as deepmedchem) covers the everyday operations without writing Python:

dmc databases                        # abbreviations, sizes, prices, order emails
dmc databases --detailed             # full IDs, BioSolveIT mappings, availability, links
dmc usage                            # account plan and CHEESE Credits remaining today
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine -m shape -n 10
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine -o aspirin.csv
dmc substructure "[N;R0][N;R0]C(=O)" -d enamine -n 10 --timeout-seconds 45 -o hydrazides.sdf
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d synple --subset amino-acids
dmc sample -d freedom -n 100 --seed 7 -o sample.smi
dmc order aspirin.csv --get-quote
dmc optimize run my-run -d enamine --minimize --score-cmd './dock.sh {input} {output}'
dmc optimize status                  # your optimizations; also ask, tell, results, cancel

databases lists every searchable space with its size, whether per-compound price estimates are available, and the vendor address for orders and quotes. Output captured on September 11, 2026; the catalog can change. For Enamine, the size is an estimated number of source reagent combinations, marked with ~, rather than unique molecular graphs. --detailed adds full IDs, BioSolveIT mappings, type, availability, success estimates, and vendor links. --json always returns the unmodified API catalog:

$ dmc databases
abbreviation  molecules  prices  orders
------------  ---------  ------  ------------------------
enamine          ~93.4B  yes     info@enamine.net
freedom          296.4B  yes     sales@chem-space.com
explore            9.5T  yes     sales@emolecules.com
synple             7.6T  yes     sales@emolecules.com
vast               6.8B  yes     VAST@XtalPi.com
cheminfinita     794.2B  -       sales@otavachemicals.com
spacem1            1.5B  -       hello@molecule.one

7 databases. Lead times vary by database; see --detailed.
Order or request quotes by email, or run `dmc order results.csv`.

Searches print a table of rank, similarity score, price, and SMILES, followed by what was searched and the score range. Substructure hits show exact instead of a score, and samples have no score column. Product ids and the other API fields are kept in --json and in exports.

Example output from the production API (September 11, 2026):

$ dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine -n 3
rank   score  price  smiles
----  ------  -----  -----------------------------
   1  0.7037   $245  O=C(O)Oc1ccccc1C(=O)O
   2  0.6667   $163  COC(=O)Oc1ccccc1C(=O)O
   3  0.6061   $163  CC(C)(C)OC(=O)Oc1ccccc1C(=O)O

Searched ~93.4B source combinations (Enamine REAL v5a) in 396 ms.
Similarity range: 0.61-0.70 ECFP4 Tanimoto.

-o/--output saves the hits as CSV, SDF, SMILES (.smi), or JSON, inferred from the file suffix (--format overrides it). CSV and SDF carry the score, price, product id, and every other field from the response. SDF output needs RDKit (pip install "deepmedchem[sdf]"); the other formats have no extra dependencies. Every command accepts --json for the raw API response and --profile to pick a configured profile.

Enamine filtered population

The production and development APIs use the filtered Enamine release 2026-09-06.2 by default for database="enamine", including sampling, similarity, selections, runs, and substructure search. It applies MW ≤500, logP ≤5, HBA ≤10, HBD ≤5, rotatable bonds ≤10, and TPSA ≤140 Ų on the assembled product after selecting the first valid source topology. Price estimates remain available on the returned products.

The updated estimate is 93.41B source combinations (95% interval: 93.19–93.63B), based on 18.7 million draws. This is the historical approximately 95B population, replacing the larger 336.72B normalized population as the default. The catalogue reports the estimate and its confidence interval. The CLI marks estimated counts with ~; this population counts source reagent combinations, which may produce the same molecular graph. Sampling is not uniform over unique molecular graphs. No client-side property preset is needed to select this population.

from deepmedchem import Client

with Client() as client:
    sample = client.sample(database="enamine", count=20, seed=7)
    matches = client.search_substructure("C(=O)N", database="enamine", limit=20)

DeepMedChem All Chemical Spaces

Snapshot: September 11, 2026. Use the abbreviation in database="enamine" or dmc search ... -d enamine. Full database IDs remain supported. Abbreviations resolve to the releases listed by dmc databases --detailed.

Abbreviation	Type1	Size	Availability	Success rate	Prices2	Orders	Link
enamine	Make-On-Demand	~93.41B source combinations	3–4 weeks	>80%	yes	info@enamine.net	🔗
freedom	Make-On-Demand	296.4B	5–6 weeks	>80%	yes	sales@chem-space.com	🔗
explore	Make-On-Demand	9.5T	3–4 weeks	>85%	yes	sales@emolecules.com	🔗
synple	Make-On-Demand	7.6T	3–4 weeks	>85%	yes	sales@emolecules.com	🔗
cheminfinita	Make-On-Demand	794.2B	5–8 weeks	55–85%	-	sales@otavachemicals.com	🔗
spacem1	Make-On-Demand	1.5B	2–6 weeks	>85%	-	hello@molecule.one	🔗
vast	Make-On-Demand	6.8B	2–4 weeks	85%+	yes	VAST@XtalPi.com	🔗

Migrating from BioSolveIT InfiniSee

Use the DeepMedChem abbreviation in Python or with -d. The BioSolveIT slugs below identify analogous collections; they are not accepted as DeepMedChem database IDs. Releases and molecule counts differ. BioSolveIT download catalog.

DeepMedChem abbreviation	BioSolveIT slug
enamine	REALSpace_95bn_2026-04**
freedom	FreedomSpace_296bn_2026-03
explore	eXplore_8tr_2026-06
synple	Synple_8tr_2026-06
cheminfinita	CHEMriya_55bn_2025-10*
vast	VAST_4bn_2026-05

* No direct mapping: CHEMriya is a related Otava collection, not ChemInfinita. But it might be very similar. Contact Otava for details.

** No direct mapping: our Enamine version is built from public Enamine building blocks, so it is not identical to the BioSolveIT version. But it is very similar.

Onepot CORE 2.0 Lite

Use database="onepot" (canonical ID onepot-core-2-lite) for Onepot searches. Availability and supported capabilities come from the connected API catalog. The supplier quotes $250 for two-component products and $375 for three or more components, per 1 mg. Shipping cost is unspecified. Reported synthesis success is 60–70%, synthesis takes 5–7 business days, and US delivery is overnight. These are supplier estimates; request a final quote from orders@onepot.ai.

dmc search "CC(=O)Nc1ccccc1" -d onepot -n 10 -o onepot.csv
dmc substructure "[N;R0]C(=O)" -d onepot -n 10 -o onepot-substructure.csv
dmc order onepot.csv --get-quote --amount-mg 1 --no-open

Use the existing properties= and max_price= search arguments for descriptor and price constraints. ADMET availability is listed in the API capabilities; CP16 acquisition scores are approximate and final properties use full-product models. Onepot result exports retain the supplier reaction ID and ordered vendor synthon IDs in offers. Order drafts carry those identifiers into the supplier CSV, together with the product SMILES and DeepMedChem reference.

Requesting quotes and orders

Prepare vendor-ready requests directly from an exported result CSV:

dmc order results.csv --get-quote          # confirm prices and availability
dmc order results.csv --amount-mg 1        # initiate a 1 mg order request
dmc order results.csv --no-open             # files only; useful over SSH

The command groups molecules by vendor email, creates one directory per recipient, and then asks the operating system to open a pre-filled email draft. It never sends email or places an order. Every request remains available as email.txt plus molecules.csv if no graphical mail client is available or a draft fails to open. DeepMedChem is CCed so vendors can attribute the request.

Vendor-facing molecule files contain only the database ID, a -DMCH reference ID, and SMILES. Search scores, properties, and non-binding SDK price estimates are deliberately omitted. The message asks the vendor to confirm final pricing, availability, lead time, and order details before processing. Use --to ADDRESS for a private database without a configured procurement contact, and --database ID for older CSV files that do not carry a database column.

Quickstart

import deepmedchem as dmc

result = dmc.search(
    "CC(=O)OC1=CC=CC=C1C(=O)O",  # Aspirin
    database="enamine",
    method="shape",
    limit=3,
)

print(repr(result))
for hit in result.hits:
    price = f"${hit.price}" if hit.price is not None else "unavailable"
    print(f"{hit.rank}  score={hit.score:.4f}  price={price}  {hit.smiles}")

Example output from the production API (September 11, 2026):

SearchResult(3 molecules, method='shape', database='enamine-real-v5a')
1  score=0.9726  price=$245  O=C(O)Oc1ccccc1C(=O)O
2  score=0.9719  price=$163  COC(=O)Oc1ccccc1C(=O)O
3  score=0.9551  price=$163  COC(=O)Oc1ccccc1C(C)=O

Prices are estimates in whole US dollars for delivery to the United States. They are returned in the original search response, so both hit.price and the aligned result.prices list are available without another API request. Databases without price estimates return None; run dmc databases for the current list and the vendor address to request a binding quote.

Optimize with explicit chemistry and optional custom scores

Put structures and property limits in constraints=, and preferences in objectives=. The immutable Python objects provide named bounds, local validation and IDE completion. Navigator guides synthon tuple generation and ranking, then verifies assembled products. Docking and proprietary models remain callbacks.

import deepmedchem as dmc

result = dmc.optimize(
    database="enamine",
    name="qed-amino-acids",
    budget=500,
    batch_size=50,
    constraints=[
        dmc.PropertyRange("MolWt", max=400),
        dmc.PropertyRange("MolLogP", min=0, max=3),
        dmc.RequireStructure("alpha-amino-acid/v1"),
    ],
    objectives=[
        dmc.Maximize("rdkit.qed", scale=(0, 1)),
    ],
)
result.to_csv("qed_demo.csv")

All constraints in the list must hold. min and max are inclusive; omit either for an open end. scale=(lower, upper) normalizes objective values; it does not set eligibility limits. Use scale=dmc.Scale(lower=0, upper=1) when named bounds are clearer. Tuples are validated and stored as immutable Scale objects. Each objective has an optional positive weight (default 1). The aggregate is maximized automatically, even when individual terms use Minimize.

Intent	Python object
Descriptor bounds	PropertyRange("MolWt", max=400)
Required registered pattern	RequireStructure("alpha-amino-acid/v1")
Required inline SMARTS	RequireStructure(smarts="C(=O)N", min_count=1)
Excluded inline SMARTS	ExcludeStructure(smarts="[N+](=O)[O-]")
At least one of several patterns	RequireAnyStructure([RequireStructure(smarts="C(=O)N"), RequireStructure(smarts="C(=O)O")])
ADMET bounds	PredictedPropertyRange(endpoint, units=units, max=6)
Descriptor objective	Maximize("rdkit.qed", scale=(0, 1))
ADMET objective	Minimize(PredictedProperty(endpoint, units=units), scale=(3, 8))
Similarity objective	Maximize(Similarity("CCO"), scale=(0, 1))
Callback in a mixed run	Minimize(ExternalScore(), scale=(-12, -4))

All names above are available from deepmedchem (use the dmc. prefix with the import above). The objects are frozen Pydantic models: invalid bounds, nonfinite numbers, blank IDs and misspelled fields fail locally. Known descriptor aliases such as MolWt, mw and rdkit.mol_wt share units; a custom property ID needs explicit units in PropertyRange. The service validates ID availability and chemistry. Only one RequireAnyStructure group is supported per specification; its minimum defaults to 1. ExcludeStructure rejects products with at least min_count matches (default 1). Exclusions require a positive min_count and do not accept an upper count bound, so more matches can never make an excluded product pass. Use RequireStructure(..., max_count=...) when you want an inclusive match-count window.

Objective sources must be unique after alias resolution. There can be at most one external score and one similarity objective, even with different reference molecules. An external-only run uses a callback and direction= without objectives. Local validation enforces the service limits: 16 objectives, 32 descriptor and 32 ADMET conditions, and 16 structures per group (all, any, none).

Existing properties= dictionaries, structured constraints= dictionaries and objective dictionaries remain supported. An objective list can mix typed objects and dictionaries; a constraint list contains typed objects only. Both lists and tuples are accepted. Typed property ranges serialize to constraints.properties, and typed objectives include explicit direction and weight. When converting an existing campaign from a differently shaped legacy request, use a new name or dmc.resume(old_name) to continue its saved specification: resume identity is based on the serialized request.

No callback or client RDKit installation is needed for built-in-only runs. Check dmc.catalog()["optimization"] for the deployment's enabled constraints, objectives and ADMET endpoints. Typed structure requirements use synthon anchors, including reaction junctions; released descriptor caches and reaction margins guide property-feasible sampling. Final product checks enforce counts, chirality, exclusions and exact descriptor bounds. The engine reports its actual acquisition path and any fallback in optimization.engine.constraint_acquisition.

For docking, keep the callback focused on docking:

result = dmc.optimize(
    docking_score,
    direction="minimize",
    database="enamine",
    budget=500,
    batch_size=50,
    constraints=[dmc.PropertyRange("MolWt", max=400)],
)

A callback takes a SMILES list and returns one score per molecule, optionally with metrics. name= resumes an unchanged specification; local score journals prevent repeated external scoring after an interrupted upload. The sync/async create APIs support the same parameters. The CLI provides repeatable --require-pattern, --exclude-pattern, --require-smarts and --exclude-smarts flags. Combine them with --property MolWt=:400 for common constraints; --constraints constraints.json remains available for advanced conditions. Use --objectives objectives.json and omit --score-cmd for built-in-only runs. See optimization examples for normalized mixed objectives, molecular similarity, ADMET and docking.

ADMET availability is specific to the database and its installed release. With a client, inspect client.optimizations.capabilities("enamine"); the returned predicted_properties maps endpoint IDs to units. The equivalent catalog field is optimization.database_capabilities["enamine-real-v5a"].predicted_properties. The legacy optimization.constraints.predicted_properties contains only endpoints shared by every advertised database, so it can be empty while Enamine supports ADMET. Also check the catalog's constraints.enabled and objectives.enabled flags.

with dmc.Client() as client:
    capabilities = client.optimizations.capabilities("enamine")
    endpoint = "openadmet-herg-pchembl"
    units = capabilities["predicted_properties"][endpoint]
    result = dmc.optimize(
        client=client,
        database="enamine",
        budget=100,
        batch_size=20,
        constraints=[dmc.PredictedPropertyRange(endpoint, units=units, max=6)],
        objectives=[
            dmc.Minimize(
                dmc.PredictedProperty(endpoint, units=units),
                scale=(3, 8),
            ),
        ],
    )

The bounds above illustrate the API; choose thresholds for your experiment. CP16 ranks synthon tuples before assembly. The matching full-product model supplies the final eligibility check and objective value. These are model predictions, not experimental measurements. A database with no advertised endpoint rejects that term before a campaign is created. Adding an external docking term keeps docking in the callback while the same internal ADMET guidance and checks remain active.

Reuse chemistry constraints in search and selections

Use the same objects across workflows:

mw = dmc.PropertyRange("MolWt", max=400)
amino_acid = dmc.RequireStructure("alpha-amino-acid/v1")
no_nitro = dmc.ExcludeStructure(smarts="[N+](=O)[O-]")

hits = dmc.search(
    "CC(=O)Oc1ccccc1C(=O)O", database="enamine",
    where=mw, require=amino_acid, exclude=no_nitro,
)
selection = dmc.Selection.from_database("enamine").constrain(mw, amino_acid, no_nitro)
# Selection.where(mw) also accepts a PropertyRange.

Selection.constrain(*constraints) accepts descriptor/ADMET ranges and required, excluded or alternative structures. It returns a new selection, retaining existing conditions. Search where= accepts a PropertyRange or a list mixed with existing text filters; require= and exclude= accept matching typed structure objects or existing SMARTS strings. Legacy Pattern and PropertyCondition objects remain supported.

Filters and virtual subsets

A similarity search can be restricted to a virtual subset: a definition rather than a stored catalogue, so it works on every DeepMedChem chemical space. No enumerated amino-acid version of Synple exists, yet this returns only amino acids:

dmc search "CC(C)C[C@H](NC(=O)c1ccccc1)C(=O)O" -d synple --subset amino-acids
dmc search "CC(C)C[C@H](NC(=O)c1ccccc1)C(=O)O" -d synple:amino-acids        # same search
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine --subset ro5 --where "tpsa=40..90"
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine --require "c1ccncc1" --exclude "[N+](=O)[O-]"
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine --require-count 1-2 "[NX3;H2]"
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d enamine --max-price 200                # USD per 1 mg
dmc search "CC(=O)Oc1ccccc1C(=O)O" -d synple:amino-acids --explain           # no credit used
dmc sample -d enamine --subset druglike -n 100
dmc subsets                                                               # the registry

--subset, --where, --require, --require-count and --exclude can be repeated and are combined with AND. --where takes mw, logp, hbd, hba, rotb, tpsa, fsp3 and arom (or the rdkit.* property IDs) as mw<=450, logp<4, tpsa=40..90 or 1<=hbd<=3. --max-price USD keeps molecules whose catalogue price (whole USD, 1 mg, US delivery, the price column) is at most that amount; molecules without a price are dropped unless --keep-unpriced is given. --explain prints the selection document with the server's validation and execution estimate, and runs nothing.

subset	definition
amino-acids	alpha-amino-acid residue, carboxyl free or esterified; N free, alkylated, acylated, carbamate-protected, sulfonylated or in a ring
amino-acid-building-blocks	alpha-amino acid with N free, alkylated or carbamate-protected (Boc, Fmoc, Cbz)
beta-amino-acids	beta-amino-acid residue, carboxyl free or esterified
carboxylic-acids, primary-amines, acrylamides	the named group
covalent-warheads	any of acrylamide, chloroacetamide, vinyl sulfonamide, sulfonyl fluoride, propiolamide
ro5	Lipinski: MW <= 500 Da, cLogP <= 5, HBD <= 5, HBA <= 10 (server preset lipinski-ro5/v1)
druglike	MW <= 500, cLogP -10..5, HBA <= 10, HBD <= 5, RotB <= 10, TPSA <= 140, Fsp3 0.1..0.8, 1..5 aromatic rings
beyond-ro5	MW > 500 Da
ro3	MW <= 300, cLogP <= 3, HBD <= 3, HBA <= 3, RotB <= 3

amino-acids has the scope of the Mcule amino-acid catalogues: it matches 98-100% of each residue file of MCULE-AA (2026-08-01) except glycine and alanine (about 90%), whose remaining entries attach the residue through an aromatic ring nitrogen. Matching ignores stereochemistry, so natural and unnatural (D) amino acids are not told apart. dmc subsets --json prints every SMARTS.

In Python the same filters are keyword arguments of search and sample:

import deepmedchem as dmc

hits = dmc.search(query, database="synple", subset="amino-acids", limit=50)
hits = dmc.search(query, database="enamine:ro5", where=["mw<=450"], require=["c1ccncc1"])
hits = dmc.search(query, database="enamine", require=[("[NX3;H2]", 1, 2)])  # 1-2 matches
picks = dmc.sample(database="enamine", subset="druglike", count=100, seed=7)
cheap = dmc.search(query, database="enamine", max_price=200)  # price <= $200 per 1 mg

A filtered search is one molecule-selection/1 document: a similarity objective plus the filters as constraints, sent to POST /api/v2/selections and returned as a SelectionResult. Selection.subset(), .require_structure(), .exclude_structure(), .require_any_structure(), .where("mw<=450") and .max_price(200) build the same constraints by hand. The server checks every constraint on every assembled product, so each returned molecule satisfies all filters; nothing is filtered on the client. Engines that support it anchor the synthon search on synthons able to carry a required pattern instead of filtering nearest neighbours. If fewer molecules than requested pass the filters among the candidates examined, the result is shorter and carries a partial_results warning, which the CLI prints.

Limits: a filtered search returns at most 200 hits. Sampling accepts property filters only, because the API applies structure constraints to similarity searches alone. Batch search (--input) and substructure search do not take filters yet. The API does not report how many molecules a subset contains.

A price limit is enforced inside the synthon search, not by dropping expensive hits from an unfiltered list: the server skips building-block combinations priced above the limit before it ranks them, then prices every returned molecule again. It works for similarity search and sampling on databases whose catalog entry lists max_price under pricing.constraints (dmc databases --json); elsewhere the server answers capability_unavailable for constraints.price. A limit below the cheapest price of a database returns no molecules and a price_unattainable warning. Prices are planning estimates, not quotes. With RDKit installed (pip install "deepmedchem[rdkit]"), SMARTS are checked locally before any request.

SMILES and SMARTS substructure search

Enamine substructure search is enabled in production on release 2026-09-06.2. All 326 route partitions have native indexes. Returned products satisfy the filtered population rules and are checked against the original query, including bond order, recursive predicates, and chirality.

Use format="smiles" for a concrete molecular graph, including the existing junction-spanning examples. Use format="smarts" for atom lists, ring constraints, recursive expressions, and other SMARTS query features:

junction = dmc.substructure("CNC(=O)N1CCC1", format="smiles", database="enamine-real-v5a", limit=10)
hydrazides = dmc.substructure(
    "[N;R0][N;R0]C(=O)", format="smarts", database="enamine-real-v5a",
    limit=10, timeout_seconds=45, timeout=60,
)

See the runnable substructure example for several SQC-derived SMARTS queries. More demanding motifs can take tens of seconds. timeout_seconds sets the server search budget; timeout sets the SDK's HTTP timeout and should leave room for serialization and transport. A deadline can return fewer than limit hits; inspect result.raw["timed_out"], coverage, and warnings. Exact returned matches do not imply an exhaustive search of every possible product.

Any result writes itself with result.to_csv(path), result.to_sdf(path), or result.to_file(path) (format inferred from the suffix). dmc.usage() and Client.usage() return the account plan and today's CHEESE Credit balance (plan, limit, used, remaining, unlimited, reset_at, seconds_to_reset) from GET /api/v2/usage. The balance is shared by all of the account's keys and the CHEESE web app, resets at midnight UTC, and reading it is free.

Module-level search, substructure, sample, catalog, and usage operations create and close a small internal client. The explicit Client remains available for connection reuse and advanced selections/runs. Search results behave as ordered SMILES sequences (result[0], result[:3], list(result)) while retaining typed hits, scores, prices, metadata, warnings, and the complete raw response locally.

The default profile calls https://api.deepmedchem.com. Keys created at https://cheese.deepmedchem.com work on both the legacy and v2 APIs. All keys for an account share one daily CHEESE Credit balance: one successful synchronous execution or durable-run item costs one credit. Ordinary synchronous work has a 10-second execution limit; use the Runs API for longer work, where a basic item has a 60-second limit. Substructure search uses its separate timeout_seconds budget.

Credentials resolve from an explicit api_key, DEEPMEDCHEM_API_KEY, compatibility environment variables, a custom credential provider, the selected profile's OS-keyring entry, or the credentials.json fallback file. Set DEEPMEDCHEM_CREDENTIAL_STORE=file or =keyring to force one store. Use dmc login --profile dev for the development service; profiles never share credentials. Dev login opens https://cheese-dev.deepmedchem.com, reusing your session in that browser. If an older config.toml explicitly sets [profiles.dev].web_url to https://cheese-new-dev.deepmedchem.com, change it to https://cheese-dev.deepmedchem.com. Explicit profile URL overrides are preserved.

Every request identifies its source with X-DMC-Client, X-DMC-Client-Version, and X-DMC-SDK-Version. The default values attribute direct SDK use to deepmedchem-python; an application such as Navigator can override application and application_version while retaining the installed SDK version separately.

Selections and durable runs

Selection and Run are immutable, chemistry-thin builders. They produce the public molecule-selection/1 and run/1 documents; all chemistry and capability validation remains on the API.

Exact RDKit constraints, fast predicted-property acquisition, and hard assembled-product predicted ranges are available through Selection. The simple search and sample helpers and the CLI accept the common filters directly; see Filters and virtual subsets.

from deepmedchem import Client, Selection

selection = (
    Selection.from_database("enamine-real-v5a")
    .reference(
        "query",
        smiles="CCOc1ccc(C(=O)N2CCN(C)CC2)cc1",
    )
    .maximize_similarity("rdkit.ecfp4_tanimoto", reference="query")
    .require_preset("lipinski-ro5/v1")
    .where("rdkit.mol_wt", lte=450, units="Da")
    .acquire_predicted_property(
        "openadmet-herg-pchembl",
        direction="minimize",
        keep_fraction=0.25,
    )
    .include("properties", "objective_components")
    .limit(100)
)

with Client() as dmc:
    result = dmc.selections.create(selection)

Every returned RDKit value is calculated on the assembled product and enforced literally. For predicted-property acquisition, factorized CP16 scores cheaply narrow the candidate pool before assembly and the pinned OpenADMET teacher predicts every unique surviving assembled product. The response exposes the two stages separately as hit.acquisition.approximate_value and hit.acquisition.predicted_value. These remain model predictions rather than assay measurements.

A hard predicted-property range is enforced only by the assembled-product teacher:

selection = (
    Selection.from_database("enamine-real-v5a")
    .reference("query", smiles="CC(=O)Oc1ccccc1C(=O)O")
    .maximize_similarity("rdkit.ecfp4_tanimoto", reference="query")
    .where_predicted_property(
        "openadmet-herg-pchembl",
        lte=5.0,
        units="pChEMBL",
    )
    .limit(20)
)

Property-filtered random sampling uses the same selection contract without acquisition:

selection = (
    Selection.from_database("freedom-space-5")
    .sample(seed=42)
    .require_preset("lipinski-ro5/v1")
    .where("rdkit.mol_wt", lte=450, units="Da")
    .include("properties")
    .limit(100)
)
result = Client().selections.create(selection)

The authenticated catalog is the source of truth for each database's available properties, presets, predicted-property endpoints, and supported acquisition operation.

from deepmedchem import Client, Run, Selection

template = (
    Selection.from_database("enamine-real-v5a")
    .ranked()
    .maximize_similarity("rdkit.ecfp4_tanimoto", reference="query")
    .limit(10)
)

run_spec = Run.selection_batch(
    template=template,
    items={
        "lead-001": {"query": "CCO"},
        "lead-002": {"query": "CCN"},
    },
)

with Client() as dmc:
    run = dmc.runs.create(run_spec, idempotency_key="lead-set-v1")
    terminal = dmc.runs.wait(run.id)
    results = list(dmc.runs.iter_results(terminal.id))

AsyncClient offers matching asynchronous operations and iterators. DMCClient and AsyncDMCClient are compatibility aliases for code written against the pre-split Navigator SDK.

Navigator

The navigator terminal application is distributed separately as dmc-navigator. It depends on this SDK and adds file handling, login commands, terminal presentation, and Navigator-specific workflows.

Use with AI coding agents

The repository ships an Agent Skill in skills/deepmedchem that teaches Claude Code, Codex, Cursor, and other skill-aware agents how to search chemical space with this package and the dmc command.

# Claude Code: add the marketplace once, then install the plugin
/plugin marketplace add Deep-MedChem/deepmedchem-python
/plugin install deepmedchem@deep-medchem

# Any agent that supports the open skills format
npx skills add Deep-MedChem/deepmedchem-python

You can also copy skills/deepmedchem/ into .claude/skills/ of a project (or ~/.claude/skills/ for every project), or zip the folder and upload it as a custom skill in Claude.ai or through the Skills API.

Development

python -m venv .venv
. .venv/bin/activate
python -m pip install -e ".[test]"
ruff check .
pytest
python -m build
twine check dist/*

API documentation: https://docs.deepmedchem.com/docs/guides/python/quickstart

Runnable authenticated examples using the established Enamine query panels are in examples/live.

For interactive RDKit visualization of similarity and SMARTS substructure queries, open the Enamine search notebook.

Links

• DeepMedChem website
• CHEESE UI and database overview
• API and API reference
• Documentation and Python quickstart
• Python package on PyPI
• Python SDK on GitHub

Seeded optimization and analog harvesting

client.optimizations.create(..., hit_threshold=-7, start_paused=True) initializes without proposing. After status becomes paused, use run.add_seeds(rows, mode="synthon") with measured scores and Navigator product IDs, or mode="external" with SMILES and scores. SMILES-only seeds train the surrogate; in-space product seeds also initialize analog search. Neither spends the scoring budget. If a scorer was pinned, pass the same scorer metadata with seeds.

After the seed operation finishes, run.resume() starts proposals. During an ask/tell loop, run.transition("analog_harvest_accurate", idempotency_key="harvest-1") queues a switch after scoring the current batch. analog_harvest_fast is also supported. Poll run.refresh() and inspect resource.raw["pending_operation"]; controls apply asynchronously and do not replace pending molecules. Invalid seeds pause the run with a reason and preserve its checkpoint.

dmc optimize seeds NAME seeds.json --mode synthon --idempotency-key seeds-1 and dmc optimize transition NAME analog_harvest_fast --idempotency-key harvest-1 expose the same operations. dmc optimize run and dmc.optimize() accept a hit threshold (--hit-threshold / hit_threshold). See the seed example.

Descriptor bounds such as constraints=[dmc.PropertyRange("MolWt", max=350)] are checked on assembled products, including when combined with filters="druglike". Check the catalog's optimization property capabilities before requesting them.

Long-running optimizations

Give each independent campaign a name and pin scorer metadata when reproducibility matters. The scoring callback executes locally; keep one coordinating client per campaign.

result = dmc.optimize(
    score, direction="maximize", name="campaign-1", budget=50_000, batch_size=500,
    score_chunk_size=25, score_workers=2, result_limit=0,
)
# After interruption, resume using the stored specification (no new job is created):
result = dmc.resume("campaign-1", score, score_chunk_size=25, result_limit=0)

with dmc.Client() as client:
    run = client.optimizations.get("campaign-1")
    run.export_csv("observations.csv")
    # Optional: pip install 'deepmedchem[parquet]'
    run.export_parquet("observations.parquet")
    for observation in run.iter_results(order="best", limit=100):
        print(observation)

score_chunk_size journals each finished scoring chunk before the next upload. A crash can still lose an unfinished chunk; work on another machine needs the same local cache to reuse unsubmitted scores. Changing chunk size on resume is supported. score_workers runs callbacks in bounded threads and requires a thread-safe scorer. Running callbacks finish before thread shutdown; this is not a mechanism for killing a hung docking process.

result_limit=0 returns an OptimizationResult with an empty observations list and the server summary in .optimization; a positive value fetches only the best N observations. Omitting it preserves the existing behavior of collecting every observation. Streaming exports are atomic and encode arbitrary metrics in one JSON column instead of discovering all metric columns in memory. Async handles also support iter_results().

run.pause() requests a pause after the current server step; run.resume() releases a server pause. dmc.resume(id_or_name, score) also resumes a server-paused job before continuing scoring. For built-in-only objectives, use dmc.resume(id_or_name) without a callback. These pause controls require the corresponding backend release. Cancelled jobs remain terminal. A local advisory lock rejects a second high-level scoring client sharing the same cache; it cannot coordinate clients on different machines. Use one coordinator with scoring threads for one run, or different campaign names for independent jobs.

New backend result pages pin the snapshot at the first page so advancing campaigns do not shuffle later pages. With older servers, export completed or paused jobs for a stable view.

Exact molecular retrieval

Morgan searches use the shared platform engine, including inverse-assembly identity recovery enabled by default in core 0.4.3. No client flag or larger result limit is required. Synchronous and asynchronous clients preserve server ordering, including exact isomeric identity ahead of equal-score fingerprint collisions. Existing filters still apply; a bounded search miss is not proof that a molecule is absent. Several supplier routes may produce the same molecule.

Automatic database selection

Omit the database to search orderable synthon spaces automatically. Use --all (or Python include_catalogues=True) to also search orderable in-stock and enumerated catalogues. This is separate from scope="all", which explicitly admits research and hypothetical sources. The default never admits those sources.

dmc search 'CC/C(/C1C=CC=CC=1)=C(/C1C=CC(OCCN(C)C)=CC=1)\C1C=CC=CC=1' --synthesis 7d -n 5
dmc search 'CCO' --all --availability in_stock
dmc search 'CCO' --synthesis '7 days' --max-price 300 --explain

--synthesis 7d means a maximum vendor-declared synthesis time of 7 business days, excluding shipping. The planner compares the declaration's upper bound, excludes unknown or non-comparable estimates, and reports why each source was skipped. In-stock offers need no synthesis wait, but availability is a catalogue snapshot. Onepot CORE 2.0 Lite currently has a supplier-declared 5–7 business-day estimate; these are estimates, not guarantees. Calendar/unspecified timing and months are not silently converted into business days.

The table always includes synthesis next to price, with full words such as 5–7 days, 1–2 months, in stock or unknown. Auto results also show the database immediately after rank. SMILES are printed in full. Database, price and timing refer to the same representative offer; other retrieved offers remain available in hit.sources. hit.synthesis is a typed SynthesisEstimate with bounds, units and declaration provenance; CSV/JSON retain that metadata. Timing for explicit single-database CLI results comes from the catalogue when available.

hits = dmc.search(query, synthesis="7d", max_price=300)
hits = dmc.search(query, include_catalogues=True, availability="in_stock")
with dmc.Client() as client:
    plan = client.plan_search(query, synthesis=7)  # no search credit

These additions need the matching backend synthesis-routing release. For development testing, append --profile dev or pass profile="dev". An older server returns an error; the SDK never silently discards a requested filter or falls back to unfiltered search.

hits = dmc.search(query, database="auto", include_catalogues=True, availability="in_stock", limit=100)
hits = dmc.search(
    query, database="auto", max_price=200, where=dmc.PropertyRange("MolWt", max=450)
)
matches = dmc.substructure("C(=O)N", database="auto", limit=100)

with dmc.Client() as client:
    plan = client.plan_search(query, max_price=200)  # no search credits
    print(plan["selected"], plan["excluded"])

print(hits.routing)
for hit in hits.hits:
    print(hit.smiles, hit.sources)

Requires a backend with /api/v2/search_auto; older servers return an error and are not silently searched through a client-side fallback. Explicit database selection keeps its existing behavior. Sync/async clients and module-level search facades accept auto (the default for similarity search). scope="orderable" is the default; scope="all" includes research/hypothetical collections. An optional databases=["enamine", "vast"] restricts candidate sources. availability describes catalogue snapshots, not live stock or a delivery guarantee. Unsupported filters exclude sources with a reason in the routing report; the current enumerated runtime cannot apply prices.

The global result limit applies after molecular deduplication. Each hit keeps its representative database/product/price together; sources retains alternatives retrieved by the search. CSV export preserves the per-hit database and serializes sources as JSON. all_sources_searched is separate from exhaustive molecular coverage; check routing.sources and warnings for failures and timeouts.

One federated request costs one credit, independent of selected source count; all-source failure is refunded. The default shared execution deadline is 30 s. For timeout_seconds=120, configure a client HTTP timeout above 120 s too.

Adaptive optimization across databases

result = dmc.optimize(
    docking_score,
    database="auto",
    databases=["enamine", "freedom"],  # optional candidate allowlist
    availability="make_on_demand",
    direction="minimize",
    budget=1000,
    batch_size=40,
    name="multi-space-docking",
    constraints=[dmc.PropertyRange("MolWt", max=400)],
)
result.to_csv("candidates.csv")  # includes source routes as JSON

The backend selects up to eight eligible spaces, pins their releases, and runs one resumable campaign with one total evaluation budget. Each space keeps its own Navigator state. The coordinator explores each source, then uses observed score ranks with an exploration allowance to allocate later proposals. Scores are reused for duplicate canonical isomeric molecules across sources; different salts, tautomers and stereoisomers remain distinct. Failed evaluations also use the budget. Source IDs, releases and product routes are retained in each observation's sources and in exports.

client.optimizations.create and its async equivalent accept the same routing options. scope="orderable" is the default. Existing property/structure constraints exclude spaces that cannot enforce them. Inspect result.optimization.specification["resolved_databases"] for the pinned selection and result.optimization.engine["constraint_acquisition"] for evaluation counts. Resuming the same named intent retains its original releases.

Requires the matching backend and worker implementation. Auto optimization currently supports external scoring callbacks only; built-in multi-objective runs and seed import are unsupported. Price filters are available for auto search, not optimization. Exhausting child proposal budgets can stop a campaign before the global budget; it does not prove exhaustion of the chemical union.

In local real-Navigator benchmarks (two small amide spaces, three seeds, 24 evaluations), adaptive allocation raised mean top-five QED from 0.816 to 0.875 and tied equal allocation on a target-molecular-weight objective. These are small-space checks, not evidence of docking or trillion-scale performance.

Reusable chemical queries

The same immutable query works in substructure search, filtered similarity, selections, and optimization. The SDK validates the document without requiring RDKit; the service checks chemistry and database capabilities before executing it.

import deepmedchem as dmc

query = dmc.StructureQuery(
    expression=dmc.AllOf(conditions=(
        dmc.Contains(pattern_id="alpha-amino-acid/v1"),
        dmc.AnyOf(conditions=(dmc.Contains(smarts="c1ccccc1"), dmc.Contains(smarts="C1CCCCC1"))),
        dmc.Not(condition=dmc.Contains(smarts="[N+](=O)[O-]")),
    )),
)

hits = dmc.substructure(query, database="enamine")
neighbours = dmc.search("NCC(=O)O", database="enamine", structure_query=query)
result = dmc.optimize(
    database="enamine", name="amino-acid-qed", budget=500, batch_size=50,
    constraints=[query, dmc.PropertyRange("MolWt", max=400)],
    objectives=[dmc.Maximize("rdkit.qed", scale=(0, 1))],
)

Selection.constrain(query) and both Client and AsyncClient accept these objects. client.validate_query(query) checks chemistry without running a search. Automatic routing (database="auto") retains the complete query and excludes incompatible sources.

Mapped atoms make attachment positions explicit:

linker = dmc.Linker(
    scaffold="[N:1].[O:2]", start=1, end=2,
    min_length=2, max_length=4, atom_smarts="[#6]", bond_types=("single",),
)
r_groups = dmc.RGroups(
    scaffold="[c:1]1cc[c:2]cc1",
    groups=(dmc.RGroup(site=1, label="R1"), dmc.RGroup(site=2, label="R1")),
)

Linker lengths count internal atoms; length_unit="bonds" counts bonds instead. path="any" accepts any allowed path; path="shortest" requires the molecular shortest-path length. allow_ring_atoms=False excludes ring atoms from the interior. R groups are pendant substituents rooted at the first atom of their smarts pattern. allow_hydrogen=True permits H at that position. Equal labels require identical whole rooted substituents. exact=True prohibits additional substituents elsewhere on the scaffold; bridging groups use Linker. MatchRelation expresses disjoint, overlapping, directly connected or shared mapped-site matches. Independent conditions may overlap.

Stereo is honored by default. Use StructureQuery(..., stereo="ignore") explicitly for relaxed matching. Queries do not change product charge, tautomer or protection state. Exact hit validity does not imply exhaustive search: inspect completeness and timeout metadata. Complexity limits fail explicitly, including inside exclusions.

Save and reuse the document with the CLI:

from pathlib import Path
Path("query.json").write_text(query.model_dump_json(indent=2))

dmc substructure -d enamine --query-file query.json
dmc search 'NCC(=O)O' -d enamine --query-file query.json
dmc optimize run amino-acid-qed -d enamine --query-file query.json --objectives objectives.json

The query format is structure-query/1. Synthon sampling does not support structural constraints; catalogue sampling can use them through a selection. These features require a compatible server and remain subject to the account/guest capabilities it advertises.

1. All rows are Make-On-Demand spaces served by the platform API with its full feature set: similarity, substructure search, sampling, selections and runs. ↩

2. Prices are planning estimates per compound, before any discounts are applied, and generally assume a low number of molecules ordered and price per 1 mg. For instance VAST H2 2026 uses 1 mg per compound in a 50-compound order. Other databases similarly assume small orders; the exact basis varies by vendor. Request a final quote for your actual quantity and order size. Counts and price support come from the live API catalog. ↩