AI for Bio · Model

Chemprop v2

Open-source molecular property prediction library on directed message passing, the most common deep-learning baseline

Small moleculesAcademic labUnited StatesActive中文

Overview

NameChemprop v2 · aliases: Chemprop v2, Chemprop-RDKit, Chemprop-MT
Full nameThe software name has no expanded form
Ambiguous names (not auto-linked)Chemprop
FamilyThis record covers the v2.x rewrite only; the older version's benchmark results must not be attributed to v2
  • Chemprop v1 (2019 D-MPNN release) · †
  • Chemprop v2.x ·
OrganizationMIT (chemical engineering and electrical engineering groups)
Organization typeAcademic lab
CountryUnited States
Model classProperty prediction (graph neural network, directed message passing)
DomainSmall molecules
First release† partly verified · code release
Basis: GitHub v2.0.0 stable release (first public software release)
Note: The site treated paper dates (preprint 2025-09-15, journal 2025-12-26) as the release; a public pre-release v2.0.0-rc.1 (2024-03-04) is unverified source
Other milestones
  • First public evidence: † · v2.0.0 stable release source
  • Preprint v1: † · v2 paper preprint source
  • Journal online: † · Journal version online source
  • Code release: · v2.0.0 stable release source
Latest update · new version
Basis: GitHub v2.3.1 patch release
v2.3.1 patch, preceded by v2.3.0 (2026-07-08) and the v2.2.x series source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Several releases within 12 months
OpennessCode: Open · MIT
Weights: Closed
No general pretrained weights; users train on their own data, so any "Chemprop score" depends on who trained it and on what data
Benchmarks reported0 counted (cited only 1, unconfirmed 2)
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.first_public_evidence, benchmarks_reported[0], benchmarks_reported[1]
As of2026-11-16

What it does

Treats a molecule as a graph, builds a molecular vector by message passing on directed edges and feeds a feed-forward head for regression or classification; v2 is a full rewrite with a Python API and multi-GPU training, and the paper claims algorithmic equivalence to the previous generation with higher speed and lower memory rather than new accuracy records. It is not a pretrained foundation model but a train-on-your-own-data tool, so it usually appears as a baseline: ADMET-AI builds on it and KERMT uses its multitask variant as a comparison. Next to pretrained models such as MolE, MiniMol and Uni-Mol2 it is small, fast to train and easy to retrain on in-house data.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 0 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
27 Polaris / ASAP / OpenADMETCited as comparison only (not counted)—Third-partyAnonymous baseline in the organiser report; comparison only †partly verified · evidence
03 MoleculeNet (+OGB)Unconfirmed (not counted)—Unconfirmed sourcePage marks it unconfirmed; numbers belong to the 2019 original †disputed · evidence
04 TDC (ADMET + DTI)Unconfirmed (not counted)—Unconfirmed sourceUnversioned board entries, linked indirectly via ADMET-AI †disputed · evidence

Other evaluations (outside the 44)

  • Tasks in the v2 paper (reaction barriers, absorption wavelengths, a large quantum-property set) (Public benchmark not among the 44) Self-reported: The abstract claims speed and memory gains with accuracy unchanged from the previous generation, and gives no per-item values
  • Downstream applications such as antibiotic discovery (Case study) Third-party: Several studies use the library as a model component; applications rather than evaluations of the library

Developers report: v2 is algorithmically equivalent to the previous generation while roughly halving runtime and cutting memory to about a third, with a Python API and multi-GPU training.

Independent evidence: Third parties use it as a baseline: a competition organiser report lists it as an anonymous baseline and a public ADMET board carries entries built on it, but neither is an independent accuracy evaluation of v2.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesNo fixed training set; the downstream ADMET build uses 41 ADMET datasets from TDC source
b. Private or licensed dataNoThe library ships no data; users may train on in-house data, as corporate comparisons do source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic open-source project from two engineering departments at a US technical university, not a company; the code lives under a same-named GitHub organisation under the MIT licence. The first version shipped with a 2019 methods paper, a software paper followed in 2023 and v2 is the rewrite.
  • Funding: The v2 paper acknowledges a pharmaceutical machine-learning industry consortium, two US federal graduate fellowship programmes, a defence research programme and fellowships from a software company and a pharma company.
  • Products and partnerships: No commercial entity; the ecosystem includes an ADMET prediction platform built on it, corporate research using its multitask variant as a comparison, and applications such as antibiotic discovery.
  • Related models: ADMET-AI, KERMT, MolE, MiniMol

Reading notes

  1. The paper's claim is engineering - equivalent, faster, easier - not a new accuracy record.
  2. Old MoleculeNet numbers belong to the first version and must not be attributed to v2.
  3. With no general pretrained weights, a "Chemprop score" depends on who trained it and on what data, so cross-paper comparison requires reading the setup.
  4. Under this round's rules none of the three pairings count - two unconfirmed, one comparison-only - so the model enters the zero-pairing list.

Caveats

  1. Enters the zero-pairing list this round, recomputed under the new relation rules.
  2. The site treated paper dates as first release; the software v2.0.0 actually shipped 2024-04-23.
  3. An earlier public pre-release (2024-03-04) is unverified, so first release is flagged.
  4. Entries on the public ADMET board carry no version label, so they cannot be confirmed as v2.

Sources

  1. v2 journal paper (journal)
  2. v2 preprint (preprint)
  3. Code repository (repository)
  4. Release list (package registry)
  5. Original methods paper (2019) (preprint)
  6. Public ADMET leaderboard overview (leaderboard)

Corrections

3 corrections
FieldWasNowBasisFound
first_release2025-09-15 preprint / 2025-12-26 journal2024-04-23 (software v2.0.0 stable release); the paper date is listed separately as a milestonesource2026-11-16
benchmarks_reportedTDC recorded as an official leaderboardUnconfirmed: the board entry has no version label and appears only indirectly through a downstream modelsource2026-11-16
n_benchmarks_reportedCounted on 3Counted on 0; the model joins the zero-pairing listsource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.