Biology Models

Forecast the virus's next move, then redesign the antibody for it: Arc Institute connects the two into one pipeline

November 19, 20261 views阅读中文原文
Forecast the virus's next move, then redesign the antibody for it: Arc Institute connects the two into one pipeline

TL;DR

  • Who: Brian Hie's lab at the Arc Institute, working with Stanford. On November 17, 2026 the group released two tools together, both with open code.
  • What: antiGen does one job — it reads a virus's evolutionary tree and estimates which mutations are likely next. It has roughly 500,000 parameters. Germinal Multi also does one job — it scores a single antibody design against several versions of a target at once, and keeps the design only if every version passes.
  • Why it matters: neither tool is a breakthrough on its own. The interesting part is the handoff. The team used antiGen to shortlist three mutations that would defeat a given antibody, assembled a variant that does not exist in nature (XFG-Next), then asked Germinal Multi to redesign the antibody so it covers both the old and the hypothetical version.
  • Result: of 60 redesigns, the lead construct halved infection at 2.1 ng/mL against XFG and 6.3 ng/mL against XFG-Next.
  • The catch: a second design, X58, bound both variants but only blocked the original. Binding is not neutralization, and the authors put that counterexample in their own announcement.

The short version

Antibody drugs stop working not because the design was poor, but because the target keeps changing. Arc Institute's approach is to compute where the target is heading first, then fold that not-yet-existing version into the design objective. Instead of chasing the virus, you reserve a seat for its next move. The limitation is built in: a forecast can compress tens of thousands of possibilities into a shortlist, but it cannot name the one that will actually take over.

What the two tools actually do

antiGen reads trees, not sequences

Most mutation-forecasting models run statistics over massive sequence databases and look at which positions change often. antiGen takes a different input: a phylogenetic tree.

The team wrote its own tree-building algorithm, inspired by UShER but adapted for protein sequences and for ambiguous bases in sequencing reads. They used it to place up to 9 million public SARS-CoV-2 spike sequences on a tree seeded from Nextstrain. The model then learns how mutations happen directly from that parent-to-child structure.

It is a small model — roughly 500,000 parameters, not much larger than a phone keyboard dictionary.

Trained on data through 2023 (about 7.8 million sequences and 240,000 mutations), it beat earlier methods overall. Its largest margin came on the hardest category: mutations never seen in training that are reachable through a single nucleotide change.

It also called real events. More than half of receptor-binding-domain escape mutations landed in antiGen's top 10% of predictions. Along the actual BA.2.86-to-KP.3 trajectory, all but one were ranked near the top — KP.3 being the lineage that dominated the summer of 2024.

What happens with less data? Using only sequences through early 2021 (about 350,000 sequences and 14,000 mutations), it still outperformed prior methods at the task public health cares about most: predicting which novel mutations would show up in 100 or more samples.

The authors are candid about a weakness. In the first months of a pandemic, simply counting how often a mutation has already appeared can beat antiGen. So they built a hybrid that leans on raw frequency when data are sparse. Across every stage tested from 2020 to 2024, the hybrid matched or exceeded either component alone.

It also transfers: influenza, RSV and dengue all improved, despite far thinner sequencing coverage than COVID-19.

Germinal Multi scores one antibody against several target versions

Last year's Germinal designed antibodies against a single target. COVID-19 taught the field that an antibody tuned to one version can fail on the next.

Germinal Multi makes a straightforward change: put several versions of the target in the computer at once, and keep a design only if it is predicted to bind all of them. A second mode starts from an existing antibody and stays close to its original sequence and binding pose — a rescue path rather than a replacement.

Three cases, with the numbers

Case one: rescuing an antibody that KP.3 escaped

Sipavibart worked against earlier Omicron variants and then stopped working against KP.3. The team designed antibodies from scratch against its binding site on the spike, targeting BA.2 and KP.3 simultaneously.

They tested 57 designs. The best, R50, blocked infection by both variants. It also bound variants left out of the design, which suggests the breadth extends past the two inputs.

Case two: antibody resistance in cancer

Resistance is not a viral problem only. In colorectal cancer, resistance to cetuximab has been traced to mutations in EGFR, the protein it targets.

The team picked six EGFR mutations reported in patients and cell models. Rather than designing against each separately, they built a single target carrying all six and designed against it alongside unmutated EGFR.

Five of 60 designs bound unmutated EGFR and every one of the six single-mutation variants — versions the model never saw individually.

Case three: closing the loop (the important one)

This is the only fully closed loop in the announcement:

  1. BD55-1205 is a broadly neutralizing antibody that blocks the XFG variant.
  2. antiGen ranked every possible change to the XFG spike. Published experimental data then filtered out changes the virus could not tolerate, leaving three: A475V, E493R and Y501T.
  3. Any one of them alone left the pseudovirus infectious. All three combined — dubbed XFG-Next — made BD55-1205 completely ineffective.
  4. Germinal Multi was asked to redesign BD55-1205 to cover both XFG and XFG-Next. Of 60 redesigns, the lead, X51, blocked both, halving infection at 2.1 ng/mL (XFG) and 6.3 ng/mL (XFG-Next).

The counterexample matters more than the win

In the same redesign campaign sat X58: it bound both XFG and XFG-Next, but blocked only XFG.

A binding assay says "it recognizes the target." A neutralization assay says "it actually stops infection." By publishing that pair side by side, the authors flagged the failure mode most likely to fool a pipeline like this: the design-stage score is a binding prediction, not a function prediction.

Who it fits

  • Teams working on fast-evolving targets with dense sequencing coverage: SARS-CoV-2, influenza, RSV, dengue.
  • Anyone holding an antibody that has already failed but whose structure is well characterized — Germinal Multi's rescue mode exists for exactly that.
  • Programs facing multi-mutation resistance (EGFR-style cancer targets), where one design should cover a known set of variants instead of one at a time.
  • Groups that already maintain trees with mutation annotations — that tree plus its mutations is antiGen's entire input. Raw reads are not required.

Who it does not fit

  • Sequences but no tree. antiGen's core input is a phylogeny. Without a sampled history there is nothing for it to read.
  • Anyone asking which currently circulating variant will win. The team states plainly that they cannot answer this; the labels are too scarce and it probably needs a different class of model.
  • Anyone who needs clinical-grade evidence. Both tools rest on preprints plus pseudovirus and binding assays. No animal models, no human data.
  • Stable targets. If the target barely moves, paying for multi-version joint optimization is wasted compute.

How to try it

  1. Read the code. evo-design/antiGen (MIT) and twidatalla/germinal-multi (Apache-2.0) are both on GitHub, and both saw commits on November 17.
  2. Prepare your tree. antiGen needs a phylogeny with mutations mapped onto it. If all you have is FASTA, build the tree and annotate substitutions first.
  3. Wire the two together. Turn antiGen's shortlisted mutations into multiple target versions and hand them to Germinal Multi for joint optimization. Without antiGen you can still run the same loop over variants you already know — you just lose the forecasting layer.

Pitfalls

  • Treating binding scores as neutralization. X58 is the textbook case: bound both, blocked one.
  • Treating antiGen as a next-variant oracle. It outputs a ranked shortlist, not a name.
  • Ignoring the sparse-data regression. Early in an outbreak, frequency counting wins; the authors built a hybrid to cover that.
  • Reading preprints as settled science. Germinal Multi's November 13 preprint and antiGen's August preprint have not been peer reviewed.

Cost and time

  • Money: both tools are free and open source (MIT / Apache-2.0).
  • Model size: antiGen is tiny at roughly 500,000 parameters. Germinal Multi's cost scales with how many target versions each design must be scored against — more versions, more compute.
  • Time: the authors' own scale is "tens of designs per campaign" (57 for R50, 60 for X51 and for the five EGFR cross-binders). That is one to two orders of magnitude smaller than conventional antibody discovery campaigns that screen thousands of clones.

Next step

  • If your program is stuck on an antibody that a new variant has escaped, pull both repositories and run Germinal Multi's joint optimization over the variant set you already have. Lowest cost, fastest signal.
  • If you work on influenza or RSV, antiGen still helps despite thinner data. Reproduce its top-10% hit rate on your own tree before wiring it into a pipeline.
  • Watch for the missing piece: someone testing X51-class designs in real animal models. That layer does not exist yet.

Sources

  • Arc Institute announcement (November 17, 2026): https://arcinstitute.org/news/antigen-and-germinal-multi
  • Germinal Multi preprint (bioRxiv, November 13, 2026): https://doi.org/10.64898/2026.11.10.765734
  • antiGen preprint (bioRxiv): https://doi.org/10.64898/2026.08.27.741102
  • JClub paper page (November 16, 2026): https://jclub.com.au/a/10.64898/2026.11.10.765734
  • Bloomer report (November 18, 2026): https://bloomer.news/stories/arc-institute-antigen-germinal-multi-future-variants/
  • BioInfoSoul Chinese write-up (November 18, 2026): https://bioinfosoul.com/items/qbxhzl5qcxrtw4win6qzhrkjk
  • GitHub evo-design/antiGen: https://github.com/evo-design/antiGen
  • GitHub twidatalla/germinal-multi: https://github.com/twidatalla/germinal-multi

This article is compiled from multiple media reports. Both papers are preprints and have not been peer reviewed; the wet-lab work described is at the level of pseudovirus and binding assays, not clinical evidence.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...