Flagship use case

From an antibody lead to a lab-ready optimization plate.

A campaign takes your lead antibody and antigen context, generates a constrained design space, screens it down through progressively more expensive checks, and returns a ranked set for the next wet-lab round — with binding, developability, and humanness considered together.

Discuss an antibody campaignWorking on enzymes instead?

The pipeline

Cheap checks first, expensive checks last.

The ordering is the point: each stage removes candidates that the next stage would waste compute on, so structural modelling only ever runs on variants that already passed everything cheaper.

  1. Input

    Lead and context

    Heavy and light chain sequences, the antigen or epitope context you can share, and the positions that are off limits. Structures are used when available and predicted when not.

  2. Baseline

    Numbering and liability scan

    Region assignment across frameworks and CDRs, then a baseline read on developability liabilities and germline distance, so the campaign knows what it is starting from.

  3. Generate

    Constrained variant proposals

    Structure-conditioned inverse folding and antibody language models propose substitutions at permitted positions, producing a design space far larger than a plate.

  4. Filter

    Fast sequence-level screening

    Cheap checks run first: liability motifs, charge and pI bounds, germline distance, and obvious expression risks remove most of the space before expensive modelling.

  5. Model

    Structural rescoring

    Survivors are modelled against the antigen and assessed on interface contacts and geometry, with antibody-specific checks on top of general structure quality.

  6. Select

    Multi-objective plate selection

    Binding, developability, and humanness are traded off explicitly rather than summed into one number, then the set is spread for diversity and sized to your plate.

Campaign shape

How the space collapses to a plate.

The numbers below show the structure of a campaign and the relative scale of each stage. They are not a measured result — see the benchmark status beneath them for what is actually being validated.

IllustrativeCampaign shape only — not a measured result

Screening funnel

  1. Generated1,800

    CDR-focused variants from inverse folding and language-model proposals, under your locked positions.

  2. Screened300

    Fast sequence-level filters: liability motifs, charge and pI bounds, germline distance.

  3. Structurally evaluated96

    Complex modelling and interface analysis against the antigen, plus antibody-specific geometry checks.

  4. Recommended24

    Pareto-optimal across binding, developability and humanness, then spread for diversity.

Candidate landscape

Scatter plot. The horizontal axis is predicted functional gain and the vertical axis is developability, both increasing toward the better outcome. Selected candidates cluster in the upper right.Predicted functional gain →Developability →
Selected for the plate (24)Generated, not selected (72)

Improving affinity often costs developability. Plotting both makes the trade-off visible, and makes it a decision you take deliberately rather than one a single combined score takes for you.

Candidate-level output

VariantRegionSignalsDecision
HC S103T / Y105FCDR-H3Interface contact gain · no new liability motifCore
HC T57ACDR-H2Predicted contact retained · germline-proximalCore
LC N92QCDR-L3Removes deamidation motif · neutral on predicted bindingCore
HC G55E / LC S31RCDR-H2 + CDR-L1Larger predicted gain · charge shift near interfaceExpanded
HC W47LFrameworkAggregation-risk flag · framework positionHeld

Benchmarks in progress

Foldry does not claim validated wet-lab outcomes it does not have. These are the public benchmarks currently being run and the metrics that will be published — including negative results.

AbBiBench

In progress

Antibody binding-variant ranking on held-out labels

Enrichment · precision@k · Spearman, against single-model and naive baselines

FLIP2 alpha-amylase

In progress

Held-out library selection on an official train/test split

Enrichment at fixed budget · nDCG@k · hits-vs-budget against random libraries

Deliverables

What lands at the end of a campaign.

Ranked variant set

24, 48, or 96 candidates sized to your next round, each with the signals behind its position.

Per-candidate rationale

Why a variant is in the core set, in the expanded set, or held back.

Liability and risk flags

Developability, humanness, and expression concerns surfaced before synthesis.

Handoff files

Sequences and structured outputs formatted for synthesis vendors or internal workflows.

Fit boundaries

Better qualification makes the work more useful.

Good fit

  • You have a lead antibody and can describe the antigen or epitope context.
  • You know which property needs to move and how you will measure it.
  • You can test a shortlisted set experimentally after the campaign.
  • You want a defensible plate rather than a larger one.

Not a fit

  • De novo binder discovery with no starting lead.
  • Guaranteed affinity improvements from a single round.
  • Fully outsourced multi-programme campaigns from day one.
  • Buyers looking for self-serve software rather than scoped technical work.

Start the conversation

Bring a real molecule and a real next experiment.

The first step is a scoped conversation about whether a campaign would change what you put on the next plate.

Useful to include

  • 01The lead molecule or variant family you are working from.
  • 02The property you want to improve and how you measure it.
  • 03Screening budget, timeline, and any hard constraints.

NDA available before sharing sequences or other sensitive detail.