DeltaFold

Protein research workbench

RESEARCH WORKSPACE
ModelSequence model

Train a Sequence Model

Give DeltaFold example sequences so it can learn which letters tend to follow each other. Use the trained model to generate new sequences. If your examples come from one protein family, those patterns give you a starting point for choosing new versions to test in the lab.

Synthetic example

Fill Example enters 64 invented versions of a short protein. Train Model makes the new model active. To switch back, choose Use Bundled Model.

FASTA / JSONL / plain text

Training Settings

This is the most recent stretch the model looks at when choosing the next letter. For example, 3 means it can use up to three preceding letters.

Active ModelLoading model...

About DeltaFold

Work from a reference protein or your own assay data. DeltaFold keeps reference evidence distinct from measurements and model predictions.

Library lets you find a documented protein and choose its canonical sequence or an isoform. Save the evidence, then open in Sequence. Use as Parent carries the exact saved sequence into Experiments; the project and its rounds in Campaigns retain that reference snapshot. A reviewed entry can include inferred annotations. Feature overlaps require exact coordinates and a declared substitution relationship; they do not establish that a mutated protein retains the reference function.

Design generates candidates with the active n-gram model, which you can train on your own sequences. Evolution explores variants using heuristic scores, distinct from structure confidence and experimental measurements. For assay-based ranking, Experiments and Campaigns use assay measurements to prioritize substitution variants of one parent. Campaigns keeps one assay, recorded batches, frozen selections, laboratory manifests, and follow-up outcomes together.

Structure starts with an artificial sequence illustration. Requesting a prediction sends your working sequence to Meta's public ESM Metagenomic Atlas service. Its pLDDT values describe local structure confidence, not measured function or experimental fitness.

Use Molecular Design to prepare binder plans, import structures, compare results and save shortlists. The execution pilot requires your own qualified worker and consent to transfer the job, and runs only four fixed public profiles: 1UBQ or 1VII at seed 42 or 43. Other targets support preparation and import.

Disease Research reviews published human target–disease associations. A candidate's activity needs its own evidence. Annotations, scores and plausible structures do not establish binding or therapeutic benefit. See the model evidence and limitations, or read about sources and methods.

Reopen saved work in Runs and export or restore your account research. Full service recovery from an independent backup is not yet verified. Help explains how to use each tool and recover saved work.

Questions about DeltaFold? Contact adamkaluza@berkeley.edu.