β-catenin Inhibitors: AI’s Path to “Undruggable” Cancer

A single pathway has haunted oncology drug discovery for decades: WNT–β-catenin. The biology is strong, the patient need is obvious, and yet most programs aimed upstream (porcupine, frizzled, tankyrase) have struggled to survive the clinic—often because the pathway is essential in healthy gut, bone, and stem-cell compartments.

Now the industry is watching something different: direct or functionally direct β-catenin inhibition is finally producing credible clinical and preclinical signals. Parabilis and Sapience are testing cell-penetrating peptides that interrupt β-catenin’s protein–protein interactions. Dewpoint is preparing to test a small molecule designed to control β-catenin through biomolecular condensates.

This matters for more than β-catenin. It’s a live case study in how AI in drug discovery helps teams take on targets labeled “undruggable” by older playbooks—by finding binders faster, optimizing modalities, and connecting mechanistic biomarkers to trial decisions.

Why β-catenin has been so hard to drug (and why that’s changing)

β-catenin has been difficult because it sits at the intersection of pathway complexity, essential normal biology, and tough structural pharmacology. Those three problems stack.

First, WNT signaling behaves less like a straight line and more like a transit map with detours and redundancies. Blocking one upstream “station” can leave cancer cells plenty of alternate routes.

Second, the pathway isn’t a cancer-only switch. WNT–β-catenin activity supports normal tissue maintenance—especially gut epithelium and bone homeostasis. Many upstream inhibitors have shown on-target toxicities in these tissues. The therapeutic window is narrow, and “dose harder” is rarely the answer.

Third, β-catenin is a transcriptional hub that works through protein–protein interactions (PPIs) in the nucleus, partnering with multiple cofactors to drive gene programs. PPIs and transcription factors have historically been poor fits for traditional small-molecule screening.

So what’s changing? Two things are finally converging:

1. New modalities (stabilized/stapled peptides, intracellular delivery motifs, condensate modulators) that can engage PPIs and nuclear biology.
2. AI-driven discovery workflows that make these modalities practical at scale—especially for sequence design, structure prediction, property optimization, and biomarker strategy.

Three approaches in 2025: peptides vs condensates

The most credible progress right now comes from going downstream and being selective about what β-catenin does, not just whether it exists. Each company is picking a different “node” in β-catenin function.

Parabilis: block β-catenin–TCF transcriptional activity

Parabilis’s lead candidate FOG-001 (zolucatetide) is a stabilized peptide designed to inhibit the β-catenin–TCF complex, suppressing TCF-dependent transcription.

What stands out is the early clinical signal in desmoid tumors, where β-catenin or APC mutations are common and causally central. In a small phase I dataset:

• 12 patients were treated at three dose levels.
• Among 10 patients with at least one post-baseline scan, tumors appeared to shrink.
• Among 5 patients with more than one post-baseline scan, 4 (80%) met objective response criteria.

Desmoid tumors can regress spontaneously, so nobody should declare victory off phase I. But “every tumor is shrinking” is the kind of pattern clinicians pay attention to, especially when tolerability looks manageable.

Parabilis also reported broader phase I data across 94 patients with microsatellite stable colorectal cancer and other WNT-activating solid tumors. Signals included:

• In 23 non-colorectal patients with at least one post-baseline scan, an objective response rate of 43% (including desmoid patients).
• In 34 heavily pretreated colorectal cancer patients with at least one on-study scan, 15 (44%) had stable disease.

Safety is the other headline. Because earlier WNT-pathway drugs caused bone issues, Parabilis prophylactically doses a bone-strengthening agent in its trial. Alopecia showed up in 16 of 94 patients (17%), which some interpret as a “hair follicle is on-target” signal; the company says alopecia is minimal at the selected dose.

Where AI fits: a peptide program like this lives or dies on iteration speed and property balance—cell entry, stability, solubility, manufacturability, and target engagement. AI helps by:

• Prioritizing peptide designs that preserve binding while improving developability
• Predicting liabilities (aggregation, protease sensitivity, immunogenicity motifs)
• Optimizing formulation and PK/PD hypotheses before expensive clinical cycles

The lesson: AI doesn’t “discover β-catenin” as a target—biologists did that years ago. AI makes the engineering problem of drugging β-catenin less punishing.

Sapience: block β-catenin–BCL9 to keep β-catenin out of the nucleus

Sapience is aiming at a different interaction: β-catenin–BCL9. Their peptide ST316 is designed to disrupt this partnership, which the company says reduces nuclear localization and downregulates WNT–β-catenin gene targets.

In dose-escalation data presented in early 2025:

• 23 solid tumor patients (including colorectal cancer) received ST316.
• 4 (21%) had stable disease at that cutoff.
• The company reported good tolerability and no obvious gut/bone toxicities.

Here’s the strategic bet: if β-catenin–BCL9 activity is less essential in healthy intestinal homeostasis, then blocking it could offer a cleaner therapeutic index than upstream WNT inhibition. That’s an attractive hypothesis. But it also creates a diagnostic question: if side effects are absent, is the drug hitting the pathway hard enough in tumors?

Sapience’s answer is biomarker-driven: changes in immunosuppressive myeloid cells in blood, paired biopsies showing β-catenin retained in the cytoplasm, and upcoming tumor gene-signature data. They’re also running a colorectal cancer expansion combining ST316 with FOLFIRI and bevacizumab, with data expected in 2026.

Where AI fits: if your differentiation is “we’re selective enough to spare normal tissue,” then patient stratification and biomarker interpretation become mission-critical. AI can help by:

• Learning predictive signatures from multi-omics (tumor + blood) tied to response
• Identifying which WNT/β-catenin alterations correlate with nuclear dependency
• Optimizing trial enrichment (who to enroll) to avoid diluting a real effect

In practice, that means fewer “all-comers” disappointments and more focused signal detection.

Dewpoint: modulate condensates to inactivate β-catenin

Dewpoint’s approach is the most conceptually novel: a small molecule (DPTX3186) intended to force β-catenin into off-genome, inactivated nuclear condensates they call “depots.” Rather than binding β-catenin directly, the compound reportedly inhibits proteins regulating β-catenin movement, stability, and folding.

The core therapeutic-index idea is biophysical: cancer cells often maintain much higher β-catenin levels than healthy cells. If condensate formation has a threshold (“phase boundary”), then cancer cells might sit closer to it. A drug that nudges the system could push tumor cells into depot formation while leaving healthy cells alone.

Preclinical claims shown in 2025 include:

• Depot formation in cancer cell lines but not healthy ones
• Downregulation of β-catenin target genes
• Reduced tumor growth in patient-derived xenograft models
• No obvious GI or bone toxicity signals in early toxicology workups

A phase I/II trial is expected to start by the end of 2025 in gastric and other WNT-driven solid tumors, with readouts including safety, efficacy signals, plasma protein levels, gene expression, and depot formation in circulating tumor cells.

This program also comes with fair skepticism. Condensate biology is real, but some structures can be experimental artifacts, and independent validation matters. Dewpoint will need to show not only clinical tolerability but also clear, reproducible on-target biology.

Where AI fits: condensates multiply the complexity of mechanism-of-action work. AI becomes useful for:

• Integrating imaging + transcriptomics + proteomics into a unified MoA story
• Detecting subtle, early pharmacodynamic signals in small cohorts
• Prioritizing combination hypotheses based on network-level pathway rewiring

If this approach works, it opens a playbook for other “hard nuts,” including transcriptional oncogenes.

What AI adds across the whole β-catenin problem

AI’s real value in β-catenin programs is speed plus decision quality. When the biology is complex and the therapeutic window is tight, faster iteration isn’t enough—you need fewer wrong turns.

1) Finding the right node: not “WNT off,” but “oncogenic program off”

The industry’s upstream failures taught a painful lesson: shutting down the entire pathway often injures normal tissue. AI-enabled causal modeling helps teams ask a sharper question:

• Which β-catenin partnerships (TCF, BCL9, CBP, other cofactors) actually drive tumor fitness in a given context?
• Which partnerships are dispensable in critical healthy compartments?

This is target validation, not just target discovery.

2) Modality engineering: peptides and delivery stop being guesswork

Cell-penetrating peptides aren’t new, but repeatable intracellular delivery and scalable manufacturability are still differentiators. Modern AI-assisted design supports:

• Sequence-to-property prediction (stability, solubility, clearance)
• In silico screening for toxicity and immunogenicity risk flags
• Multi-objective optimization (binding potency plus developability)

The goal isn’t a pretty model. It’s fewer late-stage surprises.

3) Trial optimization: biomarkers as the steering wheel

Both peptide approaches emphasize paired biopsies, gene expression shifts, and phenotype readouts (like alopecia as a proxy for pathway suppression). This is where many oncology trials either sharpen or collapse.

I’m opinionated here: β-catenin trials that don’t treat biomarkers as first-class endpoints will waste years. AI helps by:

• Automating signal detection in longitudinal multi-modal data
• Identifying responder subgroups earlier
• Suggesting rational combinations (VEGF inhibitors, PD-1 blockade, chemo backbones) based on observed pathway effects

Parabilis, for example, noted gene-expression shifts tied to stemness, angiogenesis, and inflammation—exactly the kind of triangulation that should guide combination strategy.

Practical takeaways for pharma and biotech teams

If you’re building (or evaluating) an “undruggable” oncology program, β-catenin offers a checklist worth stealing.

1. Pick a mechanistic claim you can measure in patients. “We block β-catenin transcription” or “we keep β-catenin out of the nucleus” is testable. “We affect WNT” is too vague.
2. Design the biomarker plan before the trial starts. Paired biopsies, circulating tumor cell assays, gene-signature panels, and imaging endpoints should be tied to go/no-go rules.
3. Assume combination therapy is the destination. Even if monotherapy works in a narrow indication (like desmoid tumors), colorectal and gastric tumors will likely require combinations.
4. Use AI where it changes decisions, not where it looks impressive. The highest ROI is usually: molecule optimization cycles, patient selection, and early pharmacodynamic signal detection.

Where this goes next

β-catenin inhibition is no longer a “someday” idea—it’s a competitive race with real data emerging and more coming in 2026. The bigger story for this AI in Pharmaceuticals & Drug Discovery series is simpler: the boundary between druggable and undruggable is moving, and AI is one reason it’s moving faster.

If you’re leading R&D strategy, the question to ask isn’t “Should we use AI?” It’s: Which part of our discovery-to-clinic pipeline still relies on slow intuition when the data is already there?

What would happen if your next hard target had a biomarker strategy as rigorous as its chemistry—and an AI stack built to learn from every patient, not just every assay?