← Back to Home

Antibodies from Thin Air

Five AI Platforms Are Rewriting the Rules of Cancer Drug Discovery—And One Just Tested Immunogenicity

January 2026 (Updated)

UPDATE: Was made aware of Latent-X2 by the Latent Labs team. Incorporated their results and re-ran my analysis pipeline to generate this new blog post & technical report.

Key additions: First-ever immunogenicity data on AI-designed antibodies, macrocyclic peptide design capability, and the tightest binding affinity reported (26.2 pM).

This post is intended to be brief and accessible to a lay audience. For a more scientific and detailed exposition, read the full technical report (PDF).

A Quick Primer: What Are Antibodies, Anyway?

Before we dive into the AI revolution, let's talk about what we're actually designing.

Antibodies are Y-shaped proteins your immune system makes to fight infections. Each antibody is exquisitely specific—it recognizes one target (an "antigen") and grabs onto it like a lock-and-key. Your body can generate billions of different antibodies, each with a unique binding site.

Variable Region (Fab) Constant Region (Fc) CDRs CDRs

Pharma companies figured out they could turn antibodies into drugs: blocking cancer signals (Herceptin), unleashing immunity on tumors (Keytruda), or neutralizing inflammatory molecules (Humira).

The problem? Making these antibodies has been stuck in the 1970s. Inject mice, wait, hope. This takes 12–24 months. Most campaigns fail.

The New Way: Design from First Principles

What if you could skip the mice entirely? That's what "de novo" antibody design promises. You specify exactly where on the target you want to bind, and software generates protein sequences that fold into antibodies hitting that exact spot. Until recently, this was science fiction. Now it's not.

The $200 Billion Shakeup

The antibody business has a dirty secret.

Behind the $200 billion market, there's a process that would embarrass a medieval alchemist. Inject mice, wait, hope. That's basically it.

Five groups just demonstrated something different.

Using generative AI—the same class of technology behind ChatGPT and Midjourney—they've shown you can design antibodies de novo. From scratch. No mice, no phage display, no hybridomas.

15-50%
Hit Rates
4-8 wks
Timeline
First
Immunogenicity Data

The most striking results?

The Five Contenders

JAM-2 "The Pragmatist"

Nabla Bio's JAM-2: 39% hit rate for VHH-Fc antibodies, 923+ designs on 26+ targets. 57% pass all developability filters. They've hit GPCRs (CXCR4 at 1.4 nM), checkpoint inhibitors (PD-L1), and RTKs (TrkA at <100 pM).

Chai-2 "The Precision Instrument"

Chai Discovery's Chai-2: structural accuracy below 1.7 angstrom RMSD. KRAS G12V antibody at 1.5 nM that discriminates single amino acids. The headline: a CXCR4 antibody that activates the receptor (EC50 = 164 nM)—first computational GPCR agonist.

Origin-1 "The Pioneer"

AbSci's Origin-1: built for targets where no structural precedent exists. "Zero-prior" epitope focus. Lower hit rates (~4%), but hitting targets that fail with conventional methods. Cryo-EM validated.

RFAntibody "The Open-Source Idealist"

Baker Lab's RFAntibody: MIT license, fully reproducible. Lower hit rates (~1-2%), needs 9,000+ designs per target, but rigorous validation. The genuine alternative for academic labs.

Latent-X2 "The Dark Horse"

Latent Labs' Latent-X2 brings several firsts:

The Head-to-Head

Hit Rate vs. Design Efficiency

Source: Platform publications 2024–2026

PlatformDesigns/TargetBest AffinityHit RateOpen Source
JAM-245–100<100 pM39%No
Chai-250–100453 pM48%No
Origin-1<10089 nM4%Partial
RFAntibody9,000+1.4 nM1–2%Yes (MIT)
Latent-X24–2426.2 pM50%No

The Immunogenicity Moment

Here's the elephant in every computational antibody room: all the hit rates in the world don't matter if patients develop anti-drug antibodies.

Until now, no one had tested whether AI-designed antibodies trigger immune responses.

First-Ever AI Antibody Immunogenicity Data

Latent-X2 tested 4 VHH binders in PBMCs from 10 healthy human donors:

Caveats: Ex vivo only, 10 donors, one target, short timepoints. This won't satisfy the FDA. But it answers: "Has anyone even checked?" Now someone has.

Beyond Antibodies: The Macrocycle Surprise

Latent-X2 designed macrocycles using the same architecture it uses for antibodies:

TargetLatent-X2RaPID mRNA DisplayComparison
PHD21.54 nM (10 designs)729 nM (>1012 library)470x better
K-Ras G12D5.43 uM (10 designs)5.53 uM (>1012 library)Comparable

Macrocycles can access intracellular targets that antibodies can't reach. If AI can design both antibodies and macrocycles from a single platform, the modality choice becomes a design parameter.

The Oncology Landscape

Best Affinities Achieved

TargetPlatformAffinityContext
HDAC8Latent-X226.2 pMEpigenetic; neuroblastoma
TrkAJAM-2<100 pMTumor-agnostic approval
CCR8Chai-2453 pMTumor Tregs; 0 approved drugs

Best Binding Affinities (nM, log scale)

Lower values = tighter binding

The Caveats

Reality Check

We've solved the "can we generate binders" problem. We haven't solved the "will they work in humans" problem.

What's Actually Been Demonstrated

Hit Discovery
Lead Optim.
Developability
Ex Vivo Immuno.
In Vivo
IND

Latent-X2 is the only platform with ex vivo immunogenicity data. In vivo validation remains ahead for all.

So What?

For big pharma: Computational antibody design will be table stakes in three years.

For biotechs: If your differentiation is "we have immunogenicity data," you now have exactly one competitor.

For VCs: Look for efficiency metrics, immunogenicity data, macrocycle capability, and structural validation.

For academics: RFAntibody exists. The barrier just dropped to "can you run Python and validate 9,000 candidates."

Where This Goes

Twenty-four months from now, I expect:

The $200 billion antibody market isn't going away. But how we populate it is about to change fundamentally.

Want more depth? For detailed methodology comparisons, additional oncology targets, and full citations, read the comprehensive technical report (PDF).