AI for Bio · Model

Pearl

Genesis Molecular AI's closed protein-ligand co-folding model, trained on large-scale physics-based synthetic structures

Structure & co-foldingAI-native biotechUnited StatesActive中文

Overview

NamePearl · aliases: Pearl, Pearl-2026.1.dev, Pearl system, DeCAF-Pearl
Full nameNo expanded name; official materials use Pearl
Ambiguous names (not auto-linked)Pearl
FamilyNot a family entry. Single model line; Pearl system and DeCAF-Pearl are its inference system and distilled variant
OrganizationGenesis Molecular AI
Organization typeAI-native biotech
CountryUnited States
Model classCo-folding (protein-ligand, SO(3)-equivariant diffusion)
DomainStructure & co-folding (also: Small molecules)
First release · preprint
Basis: arXiv v1, with the Genesis release post the same day source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv 2510.24670 v1 source
  • Official announcement: · Genesis release post, with the same-day company rename press release source
Latest update · product update
Basis: Genesis blog post: DeCAF-Pearl distillation
Announced DeCAF-Pearl (flow-map distillation on the Pearl-2026.1.dev checkpoint), after the 06-03 Pearl system with inference-time scaling and pose ranking
Note: Blog page dates taken from search results; the model remains proprietary, with no weights or API found source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Official product updates within 12 months; the model itself is closed
OpennessCode: Closed
Weights: Closed
Neither the paper nor the release page offers weights or inference code; no API found
Benchmarks reported2 counted (self 1, modified protocol 1)
Wet-lab validation1 recorded campaign
Unverified fieldsNone
As of2026-11-16

What it does

Predicts protein-ligand complex structures, with pocket conditioning and templates at inference. The paper says performance scales with synthetic data, and concedes out-of-distribution accuracy drops and confidence-based pose selection is unreliable. Unlike Boltz-2 it targets pose accuracy, not affinity; unlike AlphaFold 3 it relies on synthetic 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. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
30 PLINDER / PINDER / Runs N' PosesDeveloper report, official protocolRuns N' Poses, best@5 with PB-valid filterSelf-reportedRuns N' Poses best@5; baselines rerun by the Pearl authors †partly verified · evidence
05 PoseBustersDeveloper report, modified protocolPoseBusters V2, best@5 with PB-valid filterSelf-reportedMixes its own best@5 with another model's top-confidence result; one model excluded for its cutoff †partly verified · evidence

Other evaluations (outside the 44)

  • InternalXtals internal crystal set (Own task set) Self-reported: 111 internal crystal structures, unconditioned and pocket-conditioned; private and not reproducible
  • OpenBind zero-shot evaluation (Pearl system) (Public benchmark not among the 44) Self-reported: Company blog says Pearl system beats all co-folding models zero-shot; company account only

Developers report: it beats AlphaFold 3 and open models on PoseBusters and Runs N' Poses (best@5 with PB-valid), with performance scaling with synthetic data.

Independent evidence: No independent evidence found. No third-party replication; weights unreleased

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesPDB (to 2021-09-30) plus OpenFold / AFDB monomer distillation, and physics-generated synthetic structures source
b. Private or licensed dataNoThe paper says no proprietary experimental data was used source
c. Own wet-lab data used for trainingNoSynthetic structures are physics-generated (582,065 in the scaling experiment), not wet lab; flagship set size undisclosed source
d. Wet lab used only for validationYes111 existing internal crystal structures used for evaluation only, not new wet lab source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Internal crystal-structure holdout (existing data) (arXiv v2)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: 111 internal crystal structures used for evaluation only; training strictly on public data
Developer-reported; the set is private and cannot be reproduced externally, not an independent replication; the synthetic set is not wet lab

Self-reported · No sourced date · verified · source

Background

  • Organization: Genesis Molecular AI (formerly Genesis Therapeutics), founded in 2019, is based in Burlingame, California, with its own lab in San Diego and an office in New York; an AI-native small-molecule drug discovery company, it renamed itself and released Pearl on 2025-10-28.
  • Funding: $52M Series A (2020-12); $200M Series B (2023-08, a16z co-led, NVentures among participants); further NVentures equity investment in 2024-11. A cumulative figure above $280M is second-hand only.
  • Products and partnerships: Pearl is the structure engine of the GEMS platform; NVIDIA engineering work speeds training and inference (company figures); pharma partners include Gilead and Incyte, whose 2026-05-20 expansion brings $120M upfront and proprietary experimental data to train Genesis's next models.
  • Related models: AlphaFold 3, Boltz-2, Boltz-1 / Boltz-1x, Chai-1, Protenix

Reading notes

  1. Both headline results are author-reported under best@5 with PB-valid; with single predictions or without PB-valid, both numbers and ranking change.
  2. The comparison takes AlphaFold 3 at its official max-confidence protocol, below other papers' re-runs, which inflates the margin.
  3. The synthetic-data route is notable, but the paper admits out-of-distribution decline, matching the Runs N' Poses pattern of lower similarity, worse results.
  4. Gains on the internal crystal set cannot be reproduced externally and are company claims only.

Caveats

  1. Other models' results in the paper were re-run or cited by the Pearl team; this page does not reproduce them.
  2. Neither weights nor inference code are public, so third parties cannot check the results.
  3. Pearl is a common word; searches and auto-linking can hit unrelated projects of the same name.

Sources

  1. Pearl paper (arXiv 2510.24670) (preprint)
  2. Pearl paper v2 full text (preprint)
  3. Genesis: Introducing Pearl (company blog)
  4. Pearl technical report PDF (official site)
  5. Genesis: DeCAF-Pearl model distillation (company blog)
  6. Genesis: Pearl system zero-shot on OpenBind (company blog)
  7. Genesis Molecular AI Pearl launch press release (2025-10-28) (press)
  8. Genesis Series B announcement (press)
  9. TechCrunch: Genesis Series A (press)
  10. Genesis NVIDIA investment and collaboration announcement (press)
  11. Incyte and Genesis collaboration expansion press release (press)

Corrections

None recorded.

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