AI for Bio · Model

DPLM-2

Multimodal discrete-diffusion protein language model generating matched sequence and structure

ProteinsMixedChinaArchived中文

Overview

NameDPLM-2 · aliases: DPLM-2
Full nameThe paper uses the model name only, with no expanded form
Ambiguous names (not auto-linked)DPLM, DPLM-2.1
FamilyNot a family entry. The sequence-only predecessor and DPLM-2.1 are separate items; this record covers DPLM-2 only
OrganizationByteDance Research, Nanjing University
Organization typeMixed · A joint project of an internet company's research arm and a university computer-science department, with no separate corporate entity
CountryChina
Model classDesign (joint sequence-structure diffusion)
DomainProteins (also: Molecule generation, Structure & co-folding)
First release · preprint
Basis: Preprint v1 submission source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Code release: † · Repository commit "release of dplm-2" source
  • Official announcement: † · Project page published with the preprint source
Latest update† partly verified · maintenance only
Basis: The final commit before the repository was archived
The final commit tweaked a sampling default, after which the repository became read-only (the archive date itself is unsourced); the successor DPLM-2.1 paper is listed separately and is not an update to this model
Note: No source found for the archive date (unverified) source
The most recent sourced event, not a claim that nothing changed since.
StatusArchived · Repository archived read-only; development moved to the successor
OpennessCode: Open · Apache-2.0 (per a mirror site, unverified)
Weights: Open
Code and weights are public and the repository is archived read-only; the licence label comes from a mirror site (unverified)
Benchmarks reported1 counted (retrospective 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, open_status.code_license, release_milestones.code_release, release_milestones.official_announcement, benchmarks_reported[0]
As of2026-11-16

What it does

Turns 3D coordinates into discrete structure tokens with lookup-free quantization and runs discrete diffusion over the joint sequence-structure distribution, generating matched sequence and structure in one pass while supporting folding, inverse folding and motif scaffolding; it is warm-started from the sequence-only predecessor. ESM3 is the comparable joint sequence-structure model on this site.

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. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
15 CAMEORetrospective dataset usefrozen 2022 target set, folding and inverse foldingThird-partyUses a frozen 2022 target set; no live submission, no service ranking. · evidence

Other evaluations (outside the 44)

  • Structure prediction (folding) (Public benchmark not among the 44) Self-reported: Sequence-to-structure on a frozen target set and a date-split structure database, against several folding and multimodal models
  • Inverse folding (Public benchmark not among the 44) Self-reported: Amino-acid recovery and self-consistency on the same frozen target set and date split
  • Unconditional co-generation (Comparison against baselines only) Self-reported: One hundred samples at each of five lengths, comparing designability and diversity against several generative models
  • Motif scaffolding (Public benchmark not among the 44) Self-reported: Twenty-four motif problems taken from prior work with 100 scaffolds each, against several generative models
  • Representation-learning tasks (Public benchmark not among the 44) Self-reported: Localization, thermostability, protein interaction and metal-ion binding against the predecessor and others, not winning on all of them

Developers report: Developers report that one multimodal sequence-structure diffusion model handles folding, inverse folding and motif scaffolding competitively, with diverse and structurally plausible co-generated samples, and that structure-aware representations also help predictive tasks.

Independent evidence: No independent evidence found. No third-party evaluation or independent replication found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPublic sequence and experimental-structure databases, an annotated protein database and synthetic structures source
b. Private or licensed dataNoNo private or licensed data reported
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: A big-tech lab and university joint project with no separate corporate entity. The research arm used one name on the 2024 paper and a different in-group research brand on the 2025 successor paper (same team); the partner is a university computer-science department.
  • Funding: Funded from the company's internal R&D budget; no external grants itemised
  • Products and partnerships: The predecessor is the sequence-only discrete-diffusion protein language model; the successor DPLM-2.1 (preprint v1 2025-04, v3 2025-06) explores the multimodal design space and is recorded separately from this model's repository events.
  • Related models: ESM3, La-Proteina, RFdiffusion2 / RFdiffusion3

Reading notes

  1. Its only site pairing is retrospective dataset use on a frozen target set: not participation, and with no service-assigned ranking.
  2. All evaluation is computational, with no wet-lab validation.
  3. The value is methodological: it shows discrete diffusion can unify sequence and structure generation.
  4. Synthetic structures enter training, so generation quality is capped by the model that produced them.

Caveats

  1. The repository is archived read-only, so expect no further fixes; the successor has its own entry.
  2. The archive date is unsourced (unverified), so the latest update is the final commit before archiving.
  3. Under this round's rules it leaves the zero-pairing list, but that pairing is retrospective reuse and must not be written as participation.
  4. The licence label comes from a mirror site (unverified).

Sources

  1. DPLM-2 preprint (preprint)
  2. Preprint v1 full text with the evaluation setup (preprint)
  3. Official repository (archived) (repository)
  4. Project page (official site)
  5. Successor DPLM-2.1 preprint (preprint)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update.date2025-04-15, the date of the successor paper DPLM-2.1 v12025-07-25, the final commit before this model's repository was archived, with the successor paper recorded separatelysource2026-11-16
latest_update.event_kindRecorded as a successor modelChanged to code maintenance only: the event is a repository commit, not a new modelsource2026-11-16
benchmarks_reportedThe site said it reported no site benchmark and listed it as zero-pairingUnder this round's rules it has a retrospective-dataset-use relation with one evaluation service and therefore leaves 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.