Tools showcase

Internal benchmark run. We ran BindCraft pilot against an internal kinase
target with two known hotspot residues annotated by Epitope Scout. The
pilot tier produced 47 design candidates over a single GPU session, with
the top design scoring ipTM 0.91 against the AF2 multimer scorer.

What the run delivered:

* 47 designs total, ranked by ipTM at the binder to target interface.
* Top ipTM 0.91, top pLDDT 89.4 on the binder body.
* Median binder length 92 residues. All designs in the 60 to 150 aa
  window BindCraft is calibrated for.
* Top 5 designs handed off to ProteinMPNN for sequence redesign on the
  same backbone, then to AlphaFold2 multimer for a sanity check.

What the run cost: roughly one hour of A100 GPU time, billed to the
wallet at standard tools.ranomics.com per second rates.

Read this entry as a typical BindCraft pilot shape, not a recommendation
for a specific target. Replace with real customer anonymized data before
launch.

Internal benchmark run. RFantibody pilot against a viral glycoprotein
epitope, with the hotspot residues taken from a published structure of
the antigen complex. Goal: generate VHH (nanobody) scaffolds that frame
the antigen footprint without prior antibody training data on this
target.

What the run delivered:

* 32 nanobody scaffolds with full CDR shaping over the hotspot patch.
* Top ipTM 0.82 at the VHH to antigen interface.
* Top pLDDT 86.1 on the scaffold body, lower on CDR3 as expected.
* Scaffolds threaded through ProteinMPNN for sequence design at
  sampling temperature 0.1, then run through AlphaFold2 multimer for
  fold validation.

Anchors for reading the result. ipTM above 0.7 on a VHH against a
shaped antigen is a tractable design. pLDDT above 80 on the scaffold
body indicates the model is confident in the framework fold. CDR3 pLDDT
is structurally noisier and a lower number there is normal.

Replace with real customer anonymized data before launch.

Internal benchmark run. Epitope Scout against an internal GPCR target.
The goal of Scout is to identify candidate epitopes before committing
GPU time on a downstream binder design tool. Scout scores each candidate
site on per dimension feasibility (surface area, conservation,
hydrophilicity, framework distance) and returns a ranked shortlist.

What the run delivered:

* Three predicted epitopes on the extracellular loops, with surface
  context.
* Top feasibility score 0.78 on the loop most exposed in the resting
  state of the receptor.
* Two of three sites with hotspot residues annotated, ready to hand
  off to BindCraft or RFantibody via the Scout to tools handoff flow.
* Predicted clash zones and framework proximity flagged on the
  remaining site so we deprioritized it before paying for design GPU
  time.

The Scout run cost was the smoke tier, billed at nominal CPU rates.
This is the recommended first step before any binder design campaign.

Replace with real customer anonymized data before launch.

Internal benchmark run. RFdiffusion pilot against an internal benchmark
target, followed by ProteinMPNN sequence design on the top backbone
candidates, then AlphaFold2 multimer scoring on the redesigned
sequences. This is the canonical RFdiffusion plus MPNN plus AF2 loop.

What the run delivered:

* 80 RFdiffusion backbones at the pilot tier.
* Top 10 backbones threaded through ProteinMPNN at sampling
  temperature 0.1 for 20 candidate sequences each.
* All 200 redesigned sequences run through AlphaFold2 multimer for
  ipTM and pLDDT scoring.
* Top combined design scored ipTM 0.88, with a clean predicted
  interface and pLDDT 87.3 on the binder body.

What this pattern is good for. Use RFdiffusion for general de novo
binder design when you do not yet have a strong scaffold prior and you
want AF2 multimer as the final scorer. For nanobody scaffolds use
RFantibody. For mixed mini protein and antibody scaffolds against the
same target use BoltzGen.

Replace with real customer anonymized data before launch.

Internal benchmark run. We took the top 5 designs from the BindCraft
kinase showcase entry above and ran each through ColabFold for a no MSA
sanity check fold. ColabFold completes in 1 to 2 minutes per run with
no MMseqs2 round trip on the local machine, which makes it the cheapest
spot check before ordering DNA.

What the run delivered:

* 5 designs folded through ColabFold in single sequence mode.
* All 5 produced foldings consistent with their AF2 multimer predicted
  interface, with a tightest pLDDT of 89.4 on the top design.
* Mean per design wall time 88 seconds on the dedicated GPU.
* One design flagged a CDR loop drift between AF2 multimer and the
  ColabFold single sequence fold. We deprioritized that design before
  ordering.

When to use ColabFold in the pipeline. As the cheapest fold sanity
check between an AF2 multimer scored design and ordering DNA. If you
need full MSA plus templates use AF2 standalone. If you only have one
sequence and want the fastest possible monomer fold use ESMFold.

Replace with real customer anonymized data before launch.

Every showcase entry above starts at Epitope Scout, runs through a design tool (BindCraft, RFantibody, RFdiffusion, BoltzGen, or PXDesign), and ends with a structure prediction sanity check (AlphaFold2, ColabFold, ESMFold, or Boltz-2).