Proteins & complexes

Mutation effects, folding and protein design. Explore models we could test and the test families that apply to them.

Models to test

These are candidates, not a ranked list. An applicable test may need a trained head, a scoring adapter, a specialist comparison, or access to a hosted service. No result is implied by appearing here.

Native prediction
Frozen embedding + trained head
Specialist / baseline
Adapter required
Conditional access

MIMIC

1.0foundation model

Public MIT code and 1.25B-parameter weights; large local memory requirement.

Official source ↗Paper-reported results →

Applicable tests

ProteinGym mutation effects

adapter required

Protein sequence outputs need a mutation-effect scoring rule.

ESM-2

8Mfoundation model

Public checkpoint; small 8M variant suits a local pilot.

Official source ↗Paper-reported results →

Applicable tests

ProteinGym mutation effects

adapter required

Predeclare a masked-marginal or likelihood-ratio variant score.

Monomer structure

adapter required

ESM-2 alone emits representations, not atomic coordinates.

ProteinMPNN

v_48_020specialist

Public code and checkpoints; requires a suitable protein structure.

Official source ↗Paper-reported results →

Applicable tests

ProteinGym mutation effects

specialist / baseline

Structure-conditioned amino-acid scores can be a comparator where structures exist.

Protein design / inverse folding

native prediction

Native structure-conditioned sequence design and scoring.

ESMFold

v1foundation model

Public checkpoint; materially larger than ESM-2 8M.

Official source ↗Paper-reported results →

Applicable tests

Monomer structure

native prediction

Native single-sequence 3D structure prediction.

Chai-1

released weightsfoundation model

Public code and weights under Apache 2.0; substantial compute required.

Official source ↗Paper-reported results →

Applicable tests

Monomer structure

native prediction

Native structure prediction, with protocol controls for sampling.

Test families

Each candidate test needs a defined dataset, split, metric and run protocol before it can become a benchmark.

ProteinGym mutation effects

Candidate test

Rank substitution effects within held-out deep-mutational-scanning assays.

Candidate models and comparators

MIMIC

adapter required

ESM-2

adapter required

ProteinMPNN

specialist / baseline

Monomer structure

Candidate test

Predict single-chain structure from sequence.

Candidate models and comparators

ESM-2

adapter required

ESMFold

native prediction

Chai-1

native prediction

Protein design / inverse folding

Candidate test

Score or design sequences conditional on a known structure.

Candidate models and comparators

ProteinMPNN

native prediction

Model sources and access descriptions checked 2026-09-16. Confirm licenses, terms and checkpoint revisions before any run.