Glyco-Immune Checkpoints: AI’s Next Immunotherapy Target

Checkpoint inhibitors are now used in nearly half of US cancer patients, yet fewer than 20% respond—and at least 25% of responders eventually develop resistance. That gap is why the December 2025 Nature Biotechnology paper on antibody-lectin chimeras (AbLecs) matters: it proposes a practical way to block a whole class of “hidden” immune brakes driven not by proteins, but by tumor glycans.

Here’s the bigger point for our AI in Pharmaceuticals and Life Sciences series: glycosylation is messy, high-dimensional biology. It’s also exactly the kind of problem where AI in drug discovery and AI-driven biomarker development can move faster than traditional trial-and-error biology. AbLecs are a new therapeutic modality; AI can help decide which glyco-immune checkpoint axis to block, in which tumor, in which patient, and in which combination.

Why glyco-immune checkpoints are worth your attention

Cancer doesn’t only “hide” by changing proteins like PD-L1. It also alters its cell-surface glycosylation—the dense coat of sugars and glycoproteins on the tumor surface. Those altered glycans can bind inhibitory lectin receptors on immune cells, especially Siglecs (sialic-acid-binding immunoglobulin-like lectins) and galectins, effectively telling immune cells to stand down.

A crisp way to remember it:

Protein checkpoints are the obvious brakes. Glyco-checkpoints are the subtle traction control you didn’t know was on.

The paper focuses on Siglec-7 and Siglec-9, inhibitory receptors found on immune cells (macrophages, NK cells, neutrophils) that can be engaged by tumor sialoglycans. Multiple studies across cancers have implicated this axis in immune evasion, and recent work suggests sialoglycan/Siglec activity is upregulated in PD-1 non-responders—a practical signal that glyco-checkpoints can sit upstream of resistance.

The therapeutic bottleneck: antibodies aren’t great at “sugar targets”

The authors highlight a surprisingly blunt fact: there are far fewer high-quality anti-glycan antibodies than anti-protein antibodies. Glycans tend to be weakly immunogenic, structurally similar, and context-dependent.

Even when you avoid antibody generation and instead use a “decoy receptor” approach (lectin domain fused to an Fc), binding is usually too weak to be drug-like—often micromolar to millimolar affinity, which is a non-starter for most systemic biologics.

So the field has had a big unmet need: how do you block immunosuppressive glycans specifically, potently, and safely?

AbLecs in plain language: a tumor-targeted glycan blocker

AbLecs (antibody-lectin chimeras) combine:

• A tumor-targeting antibody arm (binding a familiar antigen like HER2, CD20, EGFR)
• A lectin receptor domain (like Siglec-7 or Siglec-9 extracellular domain) that binds tumor glycans

The logic is practical. The antibody arm binds tightly to a tumor antigen and “parks” the molecule on the tumor surface. That creates a very high local concentration of the lectin domain right where the glycans are—so even if the lectin domain is intrinsically low affinity, the overall functional binding becomes potent.

What’s different from “just combine two drugs?”

This is where the paper gets interesting. The authors didn’t just show AbLecs bind; they showed the chimeric architecture changes the biology at the immune synapse.

In immune killing, the tumor and the immune cell form a contact zone. If Siglecs are recruited into that synapse, inhibitory signaling rises and Fc-driven killing drops. AbLecs appear to occupy Siglec ligands locally, preventing Siglecs from being recruited into the synapse.

That local, spatial mechanism is why AbLecs outperformed:

• Trastuzumab + soluble Siglec decoy receptor (too weak)
• Trastuzumab + Siglec-blocking antibodies (more systemic, less synapse-specific)

What the data says: stronger ADCP/ADCC and in vivo signal

The proof-of-concept starts with a trastuzumab-based AbLec:

• T7 AbLec: trastuzumab × Siglec-7 domain
• T9 AbLec: trastuzumab × Siglec-9 domain

Key findings, simplified:

Binding potency: AbLecs reach nanomolar performance

Even though Siglec domains are weak binders on their own, AbLecs bound target cells at low nanomolar apparent affinity, comparable to the parent antibody. In other words, the platform converts “weak lectin biology” into therapeutically relevant binding.

Functional effect: immune cells kill better

Across multiple human immune cell types:

• Macrophages: increased antibody-dependent cellular phagocytosis (ADCP)
• NK cells: increased antibody-dependent cellular cytotoxicity (ADCC)
• Neutrophils/PMNs: increased tumor cell killing in chromium-release assays

The most convincing control is mechanistic: when they blocked Siglec-7/9 with antibodies or enzymatically removed sialic acids with sialidase, AbLec’s advantage largely disappeared. That’s a clean sign the effect comes from glyco-immune checkpoint blockade, not just Fc engineering.

In vivo signal: reduced lung metastasis burden

Mouse models are tricky because Siglec biology differs across species. The authors used humanized mice expressing human Siglec-7/9 and human FcγRs, then tested a HER2 lung colonization model. Treatment with the T9 AbLec reduced metastatic burden compared with trastuzumab.

That’s not a clinical result, but it’s the kind of translational “signal” you want before investing in serious development.

Where AI fits: making glyco-immunotherapy developable

AbLecs solve a design bottleneck (how to physically block inhibitory glycans). AI can solve the next bottleneck: how to choose the right glyco-target, right tumor antigen, right patient segment, and right combination.

1) AI for glycosylation pattern discovery

Glycosylation is heterogeneous by cell state, tissue, and therapy pressure. That makes it hard to use with classic biomarkers.

AI models can integrate:

• Glycoproteomics / glycomics features (even if partially observed)
• Tumor transcriptomics of glycosyltransferases (proxy signals)
• Spatial proteomics / imaging readouts (where ligands actually sit)
• Clinical response data to PD-1/PD-L1 or HER2/CD20 antibodies

A realistic output isn’t “the glycan structure,” but a predictive signature of Siglec-ligand activity that flags patients likely to benefit from glyco-checkpoint blockade.

2) AI-assisted AbLec design and optimization

AbLecs are modular, but not trivial to optimize. You still have choices:

• Which tumor antigen (HER2 vs EGFR vs CD20)
• Which lectin domain (Siglec-7 vs Siglec-9 vs Siglec-10 vs galectin-9)
• Geometry, linker length, Fc format, and Fc effector tuning
• Dual blockade configurations (PD-1 × galectin-9; PD-L1 × Siglec-7; CD47 × Siglec-7)

Protein design models and developability predictors can help reduce dead ends by forecasting:

• Stability and aggregation risk
• Immunogenicity flags
• Binding geometry constraints that influence synapse exclusion
• Manufacturability and purification yield

This is one of the most practical uses of AI in biologics development: fewer build-test cycles, better first candidates.

3) AI in clinical trials: smarter combinations, smaller bets

The paper reports synergy between glyco-checkpoint blockade and established checkpoints like CD47/SIRPα and PD-1/PD-L1. Combination therapy is where trials get expensive fast.

AI can make combinations more rational by:

• Predicting which tumors are driven by Siglec-7 vs Siglec-9 biology
• Identifying dose ranges where synergy appears (e.g., low-dose CD47 blockade)
• Enriching for patients with high Siglec-ligand signatures to reduce sample size
• Designing adaptive trials that learn quickly from early immune pharmacodynamics

If you’re working in a pharma or biotech environment (including Ireland’s global manufacturing and clinical operations footprint), that’s the difference between a “cool mechanism” and a fundable development plan.

Practical takeaways for pharma and biotech teams

If you’re exploring glyco-immune checkpoints or AbLec-like modalities, these are the near-term moves that pay off.

Start with questions you can operationalize

• Which immune effector is your therapy relying on? (macrophage ADCP, NK ADCC, neutrophils)
• Which Siglec is dominant in that compartment? (Siglec-9 can dominate in lung neutrophil-rich contexts)
• Do you need systemic checkpoint blockade—or synapse-local blockade?

Build a data plan early (this is where programs stall)

You’ll want aligned datasets across discovery and translational work:

• Glyco-proxy features (glycosyltransferase expression, sialylation markers)
• Immune cell receptor profiles in the tumor microenvironment
• Functional assays that read out FcγR and Siglec impacts together

Then use AI/ML to connect the dots, not as a buzzword, but as a way to move from “we can build it” to “we know who it’s for.”

Don’t ignore safety—use targeting to your advantage

The paper makes a strong case that targeted immunotherapies generally have better tolerability than untargeted checkpoint blockade. AbLecs inherit this advantage because the antibody arm anchors the molecule to antigen-positive cells.

For development teams, that translates into a design principle:

• Prefer tumor antigens with well-characterized expression and clinical precedent
• Use the glyco-checkpoint arm to boost efficacy without turning the whole immune system “on” everywhere

What happens next: from a platform to a pipeline

AbLecs shouldn’t be viewed as “a HER2 story.” The platform already demonstrated modularity across HER2, CD20, EGFR, plus dual blockade constructs including PD-1 × galectin-9 and PD-L1 × Siglec-7.

The next wave of progress will depend on two things:

1. Translational mapping: which tumors, which immune contexts, which Siglec axes dominate
2. AI-enabled decision-making: selecting targets, designing molecules, and running trials with sharper patient selection

If the first era of immunotherapy was dominated by protein checkpoints, the next era is going to be about combining immune synapse engineering with patient-level prediction.

The question for 2026 planning cycles is simple: will your pipeline treat glycosylation as noise, or as an addressable layer of immune regulation that AI can help decode?