Molecular research
Molecular Design
Choose a target structure and plan where another protein might attach. Compare candidate shapes and save the ones you want to investigate.
A target structure is the starting point for your binder plan. Once the constraints are recorded, you can import candidate structures for inspection and experimental follow-up. The listed public examples also support execution on your own qualified GPU worker, with job progress shown here.
The execution pilot supports exact public 1UBQ and 1VII profiles at seeds 42 and 43. Other targets and settings remain preparation and import workflows. No binding activity, therapeutic benefit, or reproduction of the paper’s laboratory findings is claimed.
Plan and inspect a candidate
Start with the protein another molecule might attach to. Sign in to keep its coordinates, the plan and any imported structures in your account.
- Choose a target
- Save a plan
- Inspect candidate shapes
To run an eligible public example, you need a GPU worker you control. You review the destination and consent before sending a job.
1. Choose a Target
Choose a coordinate file (PDB or mmCIF). The saved plan keeps the original file and its residue numbering.
No structure selected.
2. Prepare a Binder Plan
PUBLIC-REFERENCE PILOT
Run a Public Example on Your GPU
Each version specifies a fixed public target and seed. Setup fills a new preparation form and leaves saved plans intact, without submitting a job or downloading a model.
Use these small, fixed examples to check the full path from a saved plan to an automatically imported native result. Your worker serves your account only. DeltaFold does not supply a shared GPU or transfer private research targets in this pilot.
Sign in to check your own GPU connection.
Connect and Manage Your Own Worker
You must already operate a Linux GPU machine with the qualified, pinned RFD4 runtime installed. Enrolling it creates a credential limited to your account’s molecular jobs. The worker connects outbound over HTTPS, so it needs neither an inbound port nor your browser password.
The credential stays masked and is not saved by this page. Treat it as a secret. If it is lost or exposed, revoke this connection and enroll a new one.
Start an Installed Worker
Extract the client on your own Linux or WSL GPU host and open its folder in a terminal. Replace the two absolute paths below. Use a private directory inside Linux/WSL for worker state, not a mounted Windows folder; see the included README. Paste the credential into the hidden prompt; do not put it in a command, URL, or notebook.
python3 -m deltafold.molecular_worker --server https://deltafold.org --runtime-root /absolute/rfd4-runtime --state-dir /absolute/private-linux-state --prompt-tokenThis command does not install RFD4, download model weights, or qualify a different GPU. Keep the worker running for queued jobs. Temporary scratch is removed after confirmed completion or a pre-upload failure. Bounded private failure diagnostics remain for operator review; the client refuses new work after ten files until they are reviewed or removed. An uncertain completion retains result data and scratch until resolved or removed by the operator.
Submit an Eligible Saved Plan
Open a saved preparation plan to check its eligibility.
Your GPU Jobs
Attempt History and Execution Receipts
Review a New Attempt
3. Saved Plans & Imported Results
Each saved revision is preserved. Importing results creates a new revision.
Import an Externally Generated Candidate
Record an Experimental Measurement
4. Compare & Shortlist
Compare up to eight candidates from saved results for the exact same target. Each retains its structures, measurements, and production history. Reviewing or saving a selection stays within your private DeltaFold account, without starting computation or sending data to a worker.
Choose candidates in a saved result and use Add Selected to Comparison. Repeat from another result to compare across runs.
The ZIP includes selected originals and complete source snapshots, including unselected candidates and measurements in those runs. Review the contents before sharing. Downloading exports research to your device, without sending data to a laboratory or other service.
Saved Shortlists
Structure Inspection
What the Paper Supports — and What Is Available Here
RFD4-Proteina research ↗ describes joint sequence and all-atom molecular design. The supplied paper is a preprint, not peer reviewed. Its experiments are the authors’ findings; DeltaFold has not reproduced them.
| Research mode | Available here | Interpretation | Release / validation status |
|---|
Software verification: DeltaFold tests input validation, account isolation, preservation, import, and export behavior.
Model execution: Six offline native executions across public 1UBQ and 1VII targets at seeds 42 and 43 completed and imported successfully. The repeated seed-42 outputs agreed at saved coordinate precision on the tested runtime. The own-worker website pilot is restricted to the listed exact profiles; your client must report support before a job can start. These engineering checks do not establish general design performance or binding. Other settings, hotspots, larger targets, and the paper’s other modes remain unqualified. No shared GPU is provided.
Experimental validation: None performed by DeltaFold for this integration. Imported measurements are user-supplied observations. The paper’s authors’ experimental results remain separate.
Read the Evidence in Context
- Computational pass rates use task-specific structural and confidence filters. They are not the probability that a candidate binds.
- Reported hits vary by target, including zero-hit screens. Selected examples do not establish a success rate for a new target.
- Extra search can reduce throughput. Compute presets and evaluation conditions need matched-budget comparisons.
- Ensemble ranking was retrospective and cohort-dependent. It is not a validated general predictor of enzyme activity.
- Experimental affinity, avidity, enrichment, conversion, and turnover are distinct endpoints. Controls, replicate types, uncertainty, and detection limits matter.
Native RFD4 code and weights are released under BSD-3-Clause; optional evaluators have their own terms. This page does not download weights or accept external services’ terms.
Software checks verify input handling, preservation, permissions, and reported geometry. Scientific validation requires held-out experimental data and laboratory testing. No therapeutic efficacy or safety is inferred.
Your Protein Library
Look up a protein by name or ID. Sign in to keep its sequence and source together so you can find them again.
Find a Protein Reference
A UniProt entry describes a protein. Some entries have several versions. Open a match and choose the main sequence (canonical) or a named alternative (isoform) before saving.
Selecting Search References sends your public search terms and organism filter to UniProt through DeltaFold. Inspect and preview requests use the identifiers you select here. Keep private sequences and confidential research details out of these fields. Searches start only at your request; your working sequence is never used as search input.
Your saved sequences
Your account keeps saved proteins with their notes and sources, along with runs and trained models. Sign in to reopen that work on another device. Existing experiment accounts work everywhere on DeltaFold.
Your anonymous runs stay available in this browser. After signing in, you can copy them to your account.
Saved Proteins
Plan Your Next Experiment
Compare lab measurements for different versions of one protein and make a shortlist for your next test.
Save Your Recovery Code
This recovery code is shown only once. Keep the code private to restore access if you forget your password.
FROM LAB RESULTS TO A SHORTLIST
Choose the Next Variants to Test.
Use measurements from your lab to choose which versions of one protein to test next. The shortlist carries the model’s estimates alongside the values you supplied.
- Name the starting protein and the test you used.
- Import the variant sequences, including at least 12 with measured results.
- Use those results to estimate how the untested versions might perform in the same test.
Compare versions of the same starting protein. A lab test can show how a shortlisted version actually performs. A ranking alone cannot show whether it treats a disease.
Your Projects
Keep versions of one starting protein together. Use the same lab test and units so their measurements can be compared.
Create a Project
Restore a Project from an Export
Synthetic example controls are available only in a separate project created with Fill Example.
Example controls are available only in the matching synthetic demo project. To repeat the example workflow, create a new demo project; saved research stays in place.
Project Variants
Keep these sample IDs in your assay files.
Recorded Measurements
Rank Your Unmeasured Variants
DeltaFold estimates how much each letter change contributes to this lab result and adds those effects together. It uses your project’s measurements on the DeltaFold server. Rank the untested versions to choose what to measure first.
Fitting requires at least 12 distinct measured variants. Predictions are rough estimates within the sequence space covered by those measurements. The model cannot learn interactions between substitutions, so keep experimental controls in your testing plan.
Candidate Shortlist
Fill the shortlist size, then choose Select Top Candidates. Nothing is selected automatically.
New data has been added since this run. These predictions belong to the saved input revision; choose Run Ranking again to use the current data.
Validation, Model Details, and Limitations
Saved Ranking Runs
Each run preserves its input data, settings and checks so you can trace how its predictions were produced.
Project Storage, Backups, and Data Transfers
Experimental projects are private to your account and separate from the original anonymous workbench. Signing into Experiments does not move old runs or make the other tools part of your private project.
Variant fitting runs on the DeltaFold server. This page does not send sequences or measurements to external model providers. Requesting an ESMFold prediction in Structure sends the sequence to that provider. Searching for target evidence in Disease Research sends the query to its provider.
Export project backups regularly and keep your recovery code private. Exported files contain your research data; store them securely. Restore an exported project from the projects page. This creates a new project from the saved input data; prior ranking records remain in the export.
ONE PROTEIN · ONE ASSAY
Campaigns
Record which versions of a protein you chose before each round of testing. When the results come back, use them to plan the next round.
Keep Each Round of Research Together
A campaign keeps each decision with the results that follow. Sign in to keep your rounds in your private workspace.
- Choose samples
- Freeze the decision
- Record what happened
Uploads and fitting stay on the DeltaFold server. No external model provider receives your research. Downloads contain private data; keep them secure. Laboratory testing is separate.
Your Campaigns
Create a Campaign
Start with versions made by replacing letters in one parent protein. Keep the test, units and conditions consistent across rounds.
Prepare a separate synthetic campaign, then choose Create Private Campaign. All example results are simulated.
Start from an Existing Experimental Project
Start a new campaign from the project's parent and assay, including its variants and numeric measurements. The original project and its saved models stay unchanged. Each campaign round records a separate model snapshot.
Variant Sequences
FASTA identifiers link each sequence to its measurements. Existing sequences remain unchanged; import additional candidates when needed.
Starting Measurements
Record which batch each reading came from and whether it used a separate biological sample or a repeat measurement of the same sample. Keep failures, missing results and readings reported only as a limit. The model uses numeric measurements.
CSV Fields and Data Rules
Required columns: variant_id,batch,replicate_type,replicate,status,raw_value. Use biological or technical for replicate type; use measured, failed, missing, or censored for status. For censored observations, retain values such as <0.1. Optional endpoint, unit, and condition fields must match this campaign. Frozen-round results also require the exported sample_id.
Saved Observations
Choose the Next Batch
Choose one batch of readings and one type of repeat. When your selection method makes predictions, they estimate what the same test might measure. Count the samples, repeats and controls before you save the round.
Review Before Freezing
A frozen round preserves the candidate pool and training snapshot used to make the selection, including predictions and controls. Later results are added separately; they cannot rewrite the original decision.
Freezing saves a record without ordering synthesis, purchasing supplies, or starting laboratory work.
Frozen Rounds
Compare what you chose before testing with what the lab later measured.
ZIP includes the frozen record, sequences, and sample manifest. All sample identifiers, controls, and replicate positions are retained.
Import This Round’s Outcomes
Use sample IDs from the handoff. Give this round a distinct batch name. Preview checks catch conflicts before any data is saved.
What Happened in This Round
These outcomes describe this recorded batch. They do not establish general protein function, binding, or therapeutic benefit.
Recovery and Limitations
Frozen rounds and outcomes are included in your full account export. Restoring a backup preserves historical decisions and does not run experiments or recompute predictions. Use Account for full export and restoration.
The current model supports substitution variants of one parent protein. A single selected assay batch and replicate type is used for each model fit. There is no automatic batch correction, censored-value model, or calibrated prediction interval. Selecting a different batch prepares a new model snapshot; earlier rounds remain unchanged.
Regular account exports help you retain your research. Local service snapshots are not a guarantee of independently protected off-site backups.
Working Sequence
Paste or edit the letters that spell out a protein. See how long it is and which building blocks it contains. Checking the length can help you catch a piece you missed when copying the sequence.
Working Sequence Ubiquitin example0 residuesSequence & Metrics
This editor starts with the ubiquitin example on a fresh visit. Copy or save letters you want to keep.
Fill Example enters a 12-letter example. Choose Analyze Sequence to calculate its numbers and update the illustration. Clear Example brings back your earlier input.
Copy your letters before signing in. In your Library, paste them into Save a Protein and give it a name.
Sequence Metrics
- Length
- --
- Water-avoiding letters
- --
- Charge-count estimate
- --
- Letter variety
- --
- Cysteines
- --
- Flagged stretches
- --
A higher score means the letters fit more of DeltaFold’s rules for sequence patterns. The counts below the score describe those letters. A lab test can tell you how the protein behaves.
How the numbers are calculated
The hydrophobic percentage counts amino acids that tend to avoid water. The charge-count proxy adds K, R and H and subtracts D and E. It leaves out pH and terminal groups. Letter variety ranges from 0 for a sequence with just one kind of amino acid to 1 when all 20 kinds occur equally often. Researchers call this normalized composition entropy. Flagged stretches count overlapping pattern matches. Several flags can come from the same stretch.
Design Candidates
Give DeltaFold a few starting letters. Compare the protein sequences it writes and choose one to inspect.
Candidates
Your starting letters
Each letter stands for an amino acid, one of the building blocks of a protein.
Compare candidates
A higher score means fewer or smaller penalties for letter patterns, such as long repeats. A lab test can tell you what the protein does. Comparing candidates leaves your working sequence alone. Choose Use to make one active in Sequence.
Run Settings
| Rank | Score | Sequence | Hydrophobic % | Charge-count proxy | Cysteines | Select |
|---|
Evolve a Sequence
Try small changes to a protein's sequence and see which versions DeltaFold scores highest. A lab test can tell you how those changes affect the actual protein. You can use the scores to choose a change to test first.
Evolution results
Score by Generation
Select a point or use the arrow keys to compare rounds.
Starting sequence
Select a letter to see which amino acid it stands for.
A higher score means fewer or smaller penalties for patterns in the letters, such as long repeats. A lab test can tell you how the protein behaves.
Read the chart values
The average covers distinct sequences scored in that generation. Several versions can tie. The run stores one top sequence per generation.
| Generation | Highest score | Average score |
|---|
Sequences and run settings
Starting sequence
Recorded top sequence for this generation
Best sequence across the run
Structure
View a protein's sequence as a picture or request a prediction of its 3D shape. If you already have coordinates, open Molecular Dynamics to inspect them. A 3D shape can show which parts sit close together even when their letters are far apart.
Sequence illustration · Not predicted
Stages the shared synthetic working sequence without changing the displayed structure. Analyze it in Sequence for the illustrative view. Predict Structure is a separate external request. For an experimental structure, choose Public ubiquitin · 1UBQ below, then Prepare for dynamics. No account needed.
Working sequence
Molecular Dynamics
Open a structure to see which parts have recorded positions. Choose an archived simulation inside the viewer to follow the same protein across snapshots. Checking for gaps helps you decide whether this structure is a useful starting point.
Displayed Sequence
Confidence by Residue
- Scored residues
- --
- Unknown
- --
- Below 50
- --
- 90 and above
- --
Select a position to see how confident the model was about that part of the predicted shape (pLDDT, 0–100). Higher values mean greater confidence in the local shape. Gaps mark missing scores or breaks between chains. Percentages include only positions with a score.
No residue selected.
Atoms in PDB counts all atom records. Only C-alpha positions are displayed. Mean C-alpha pLDDT averages valid scores for displayed residues, counting each once and excluding unknowns. Stored run summaries retain their separate all-atom averages.
Disease Research
Look up the protein your candidate might act on. See which diseases have been linked to that target and where the evidence comes from. Those sources can help you decide which disease link is worth a closer look.
The report follows published evidence about a target. Whether your candidate acts on that target is a separate question for experiments.
Start with What You Know.
Bring what you know about the candidate. If you have a target in mind, look up the sources linking that target to a disease.
- Bring a generated, evolved, or working sequence.
- Choose the human protein you want to investigate, or leave the target blank.
- Follow the sources behind each research direction and identify the evidence still missing.
If the connection cannot be supported, the report will say so.
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.
Fill Example enters 64 invented versions of a short protein. Train Model makes the new model active. To switch back, choose Use Bundled Model.
Saved Runs
Reopen earlier work to see the settings you used and the results you got.
Saved runs belong to the current workspace.
Choose Open Run to inspect a result or Use Model to activate a saved model. To create an example here, run the synthetic example in Design, Evolution or Model.
| Created | Type | Summary | Files | Open |
|---|
