The Billion-Dollar Blind Spot: Why Nobody Is Using AI to Accelerate Tooth Regeneration
The world's first drug to regrow human teeth is in clinical trials. AI has designed antibodies that are already in human testing. These two revolutions have never met. Here's why — and why that's about to change.
Teeth That Grow Back: It's Actually Happening
In October 2024, something remarkable began at Kyoto University Hospital: the world's first clinical trial of a drug designed to make humans grow new teeth.
The drug, TRG-035, is a monoclonal antibody that targets a protein called USAG-1 (gene name: SOSTDC1). USAG-1 acts as a molecular brake on tooth development — it simultaneously blocks both the BMP and Wnt signaling pathways, which are essential for tooth formation. By neutralizing USAG-1, the antibody releases dormant tooth buds that most humans carry but never develop, allowing new teeth to emerge.
The science behind this is elegant. In 2007, researchers led by Dr. Katsu Takahashi at Kyoto University showed through targeted gene knockout experiments that mice lacking USAG-1 grow supernumerary teeth — extra teeth that develop from rudimentary tooth germs that are normally suppressed. By 2021, they had published in Science Advances showing that a single injection of an anti-USAG-1 antibody could induce whole new teeth in mice, with the key insight that blocking USAG-1's interaction with BMP (rather than Wnt) was sufficient to trigger tooth regeneration.
The Phase I trial enrolled 30 men aged 30-64, each missing at least one molar, over an 11-month study period. Toregem BioPharma, the Kyoto University spinoff driving commercialization (a team of about 13 people), has partnered with WuXi Biologics for manufacturing and has already received FDA Pre-IND feedback for US clinical trials. In September 2025, TRG-035 received orphan drug designation from Japan's Ministry of Health for severe congenital tooth agenesis. The target: commercial availability by 2030.
Meanwhile, in China, a multicenter randomized controlled trial published in 2025 demonstrated that allogeneic dental pulp stem cell injections significantly improved periodontal bone defects (26.81% improvement vs. 17.43% for controls). At the Air Force Medical University, Professor Jin Yan's team achieved the world's first functional full-length dental pulp regeneration using stem cells from exfoliated deciduous teeth.
Tooth regeneration is no longer science fiction. It's in human trials.
The AI Drug Discovery Revolution: Proven and Accelerating
At the same time, AI-driven drug discovery has crossed a critical threshold — from theoretical promise to human clinical validation.
Absci has used its generative AI platform, Origin-1, to design full-length monoclonal antibodies entirely from scratch — no prior known binders needed. Two of these AI-designed antibodies are already in human clinical trials:
- ABS-101 (anti-TL1A for inflammatory bowel disease): Phase 1, first participants dosed May 2025
- ABS-201 (anti-PRLR for androgenetic alopecia): Phase 1/2a, first participants dosed December 2025
Insilico Medicine achieved another milestone: its AI-discovered drug Rentosertib (ISM001-055), a novel TNIK inhibitor for idiopathic pulmonary fibrosis, became the first drug with both target and molecule discovered by generative AI to enter Phase II clinical trials. The entire journey from target discovery to Phase 1 took under 30 months — roughly half the traditional timeline. Positive Phase IIa results were announced in late 2024.
The money follows the results: Generate Biomedicines signed a $1B+ deal with Novartis in September 2024. Isomorphic Labs (Alphabet's AI drug discovery spinoff) raised $600M in March 2025. The field is exploding.
As of early 2026, no AI-designed drug has yet received FDA approval. But the first is widely expected within 1-2 years.
The Gap: Two Revolutions That Have Never Met
Here is the striking fact at the center of this article:
There is not a single project — on GitHub, Kaggle, or Hugging Face — that applies AI drug discovery techniques to tooth regeneration.
I conducted a systematic search across all three platforms. Here's what I found:
GitHub: Only 3-4 repositories are even tangentially related to tooth regeneration biology:
- Ruohola-Baker-lab/Tooth_sciRNAseq — Single-cell atlas of human tooth development (the most relevant, from a Developmental Cell 2023 paper that generated enamel proteins from iPSCs)
- TheMoorLab/Tooth — Single-cell atlas of periodontal and dental pulp tissue
- fberio/Genetic-building-of-teeth-and-odontodes — Gene expression database for vertebrate tooth development
Zero projects involving USAG-1/SOSTDC1 computational analysis, molecular docking, or AI-driven antibody design for tooth regeneration targets.
Kaggle: ~30 dental datasets found — all focused on X-ray image segmentation and cavity detection. Zero datasets related to tooth regeneration, dental stem cells, or USAG-1.
Hugging Face: Mature dental imaging models exist (OralGPT-Omni, Dental-GPT-OSS-20B, OralSeg). But for tooth regeneration? Nothing.
The pattern is clear:
Dental imaging AI: ==================== Very mature
Dental diagnostic LLMs: ======== Emerging
Tooth regeneration AI: Empty
Why the Gap Exists
This isn't because people are lazy or unaware. There are real structural reasons:
1. The Discovery Didn't Need AI
The USAG-1 target was identified through classical developmental biology: gene knockout mice, phenotype observation, and hybridoma antibody screening. The five monoclonal antibodies (#12, #16, #37, #48, #57) were generated by immunizing USAG-1 knockout mice and fusing lymphocytes with SP2/0 myeloma cells — a technique from the 1970s. Most of this work happened before AlphaFold2 matured in 2020.
AI excels at optimization within known frameworks. It's less suited for the kind of serendipitous biological discovery that led to the USAG-1 breakthrough.
2. Data Desert
AI needs data. Tooth regeneration has almost none in machine-readable form:
- No standardized datasets of tooth developmental gene expression
- No large-scale screens of USAG-1 binding compounds
- No public repositories of dental stem cell differentiation outcomes
- The entire field produces perhaps a few dozen papers per year
Compare this to oncology, where millions of compounds, thousands of protein structures, and decades of clinical data are available for AI training.
3. Organ-Level Complexity
A tooth is not a single tissue. It's five distinct tissues — enamel, dentin, pulp, cementum, and periodontal ligament — each derived from different embryonic cell lineages (enamel from ectoderm, the rest from ectomesenchyme), requiring precise spatial organization and temporal coordination.
Bone regeneration = regenerating one tissue type. Skin regeneration = 2-3 layers. Tooth regeneration = coordinating five different tissues simultaneously. This approaches the complexity of whole-organ regeneration.
4. The Amelogenin Problem
The key protein for enamel formation — amelogenin, which comprises ~90% of the enamel matrix — is an intrinsically disordered protein (IDP). It has no fixed 3D structure. This is precisely AlphaFold's weakest area: the model cannot reliably predict disordered or dynamic substructures, and AlphaFold3 still shows ~22% "hallucination" rates for IDP regions. Since enamel regeneration depends on understanding amelogenin's self-assembly behavior, this creates a fundamental computational bottleneck.
5. Signaling Pathway Nightmare
Tooth development involves the interplay of four major signaling pathways — BMP, Wnt, FGF, and Shh — with extensive crosstalk. The same signal can be synergistic or antagonistic depending on cellular context, timing, and spatial location. This context-dependent combinatorial explosion is beyond current computational modeling capabilities.
6. Small Market, Low Priority
Dental implants already work well (>95% success rate over 5-10 years). Congenital tooth agenesis affects perhaps hundreds of thousands of people globally, compared to tens of millions for cancer. AI drug discovery companies like Absci and Insilico rationally prioritize oncology and autoimmune diseases — that's where the market is.
What AI Could Actually Do
Despite these challenges, there are clear opportunities where AI could meaningfully accelerate tooth regeneration — not by replacing wet-lab biology, but by augmenting it:
1. Better Antibodies Against USAG-1
TRG-035 was found through traditional hybridoma screening. It may not be optimal. Absci's Origin-1 platform has demonstrated the ability to design antibodies de novo against zero-prior epitopes. Applying similar AI antibody design to USAG-1 could yield:
- Higher affinity antibodies (tighter binding = lower dose = fewer side effects)
- Better selectivity (blocking BMP interaction without affecting Wnt, which is the key therapeutic mechanism)
- Improved pharmacokinetics (longer half-life, better tissue distribution)
Important caveat: USAG-1 itself (unlike amelogenin) is a structured protein belonging to the sclerostin family with a cystine knot domain. Its AlphaFold-predicted structure is available. This means AI-driven antibody and small molecule design against USAG-1 is technically feasible.
2. Small Molecule Alternatives
An antibody drug like TRG-035 is expensive — perhaps thousands of dollars per treatment. If AI could design a small molecule inhibitor of USAG-1 (blocking its interaction with BMP), the cost could drop by 100x, transforming it from a niche orphan drug into a mass-market therapy.
Tools exist for this: AlphaFold 3 for protein-ligand complex modeling, generative chemistry models like DiffDock and RFdiffusion for candidate generation, and ADMET prediction models for drug-likeness screening. The pipeline is well-established in other therapeutic areas; it just hasn't been pointed at USAG-1.
Note: For small molecule docking against USAG-1, tools like AutoDock Vina are applicable. However, for modeling USAG-1's protein-protein interactions with BMP or LRP5/6, specialized tools like HADDOCK or PatchDock would be needed.
3. Mining the Dental Single-Cell Atlas
Rich single-cell transcriptomic data for human tooth development already exists (e.g., the Developmental Cell 2023 census). Foundation models like scGPT and GeneFormer — pre-trained on 33M+ cells — could be applied to:
- Identify whether adult jawbone tissue retains dormant tooth progenitor cells (which would expand the addressable patient population from congenital agenesis to all adults with missing teeth)
- Predict optimal reprogramming factor combinations to convert jawbone mesenchymal cells into tooth bud cells
- Discover new regulatory targets in the BMP/Wnt/FGF/Shh signaling network
4. Accelerating Clinical Trials
AI could compress the Phase I-III timeline through:
- Synthetic control arms using historical patient data (reducing required enrollment)
- Adaptive trial design with real-time ML analysis of interim data
- Digital twins predicting individual patient response
- AI toxicology modeling potential side effects of USAG-1 inhibition on BMP/Wnt pathways (both are implicated in cancer — understanding the safety profile is critical)
What's Available Right Now
For anyone who wants to start working on this, here's the public data and tooling:
Data
| Resource | What It Contains |
|---|---|
| AlphaFold DB: Q6X4U4 | Predicted 3D structure of USAG-1/SOSTDC1 |
| Murashima-Suginami et al., 2021 | Antibody binding sites, functional classification, animal experiment data |
| Tooth_sciRNAseq (GitHub) | Human tooth development single-cell atlas + iPSC enamel organoid analysis |
| TheMoorLab/Tooth (GitHub) | Periodontal and dental pulp single-cell atlas |
| UniProt, KEGG, Reactome | BMP/Wnt pathway data (decades of accumulated knowledge) |
Tools
| Tool | Application |
|---|---|
| AlphaFold 3 | Protein-ligand complex modeling for USAG-1 |
| AutoDock Vina | Small molecule docking against USAG-1 binding pockets |
| HADDOCK / PatchDock | Protein-protein docking (USAG-1 ↔ BMP / LRP5/6) |
| DiffDock / RFdiffusion | Generative molecular design |
| scGPT / GeneFormer | Single-cell transcriptomic foundation models |
| ProteinMPNN / ProtGPT2 | Protein sequence design |
| Scanpy / Seurat | Single-cell data analysis |
The Competitive Landscape
Number of open-source projects combining AI drug discovery with tooth regeneration: zero.
The Opportunity
Let me be direct about what this is and isn't:
What this isn't: A guaranteed path to regrowing teeth faster. The biological challenges are real. Amelogenin is an IDP that AlphaFold can't model well. The signaling pathway complexity is daunting. Wet-lab validation is irreplaceable.
What this is: A clear case of cross-domain information asymmetry. The people who understand USAG-1 biology (developmental biologists in Kyoto, stem cell researchers in China) don't have AI drug discovery expertise. The people who build AI antibody design platforms (Absci, Isomorphic Labs) don't know this target exists. The first person or team to bridge this gap — even with basic computational analysis — will be doing something no one else in the world has done.
The most likely path to impact:
- Use AlphaFold to model the USAG-1/BMP binding interface — understand where TRG-035 binds and how it could be improved
- Apply generative models to design candidate small molecules against the same binding site — potentially replacing an expensive antibody with an affordable pill
- Mine the dental single-cell atlas with foundation models — answer whether adults retain activatable tooth progenitor cells
Each of these is a publishable result. Together, they could meaningfully accelerate the timeline from 2030 to 2027-2028 — not by skipping biological validation, but by giving biologists better starting points.
We're Building This in the Open
I've started an open-source project to begin this work: Open Tooth Regeneration (link will be live soon).
The initial scope:
- Curated dataset of all public USAG-1/SOSTDC1 structural and functional data
- AlphaFold structure visualization and binding site analysis
- Basic molecular docking experiments with AutoDock Vina
- Integration of public dental single-cell transcriptomic data
- A comprehensive literature review connecting tooth regeneration biology with AI capabilities
What I need: Collaborators with structural biology, medicinal chemistry, or dental developmental biology expertise. The AI tooling is ready. The data is public. What's missing is the interdisciplinary bridge.
If you work in any of these areas and find this interesting, open an issue or reach out. This is the kind of problem that no single discipline can solve alone — but the pieces are all there, waiting to be assembled.
All facts in this article have been individually verified against primary sources. See the fact-check document for detailed source attribution for every claim made above.
Sources:
- Murashima-Suginami, A. et al. (2021). Anti-USAG-1 therapy for tooth regeneration through enhanced BMP signaling. Science Advances, 7(7), eabf1798. Link
- Toregem BioPharma. Clinical trial initiation (October 18, 2024). Link
- Toregem BioPharma. TRG-035 Orphan Drug Designation (September 29, 2025). Link
- Toregem BioPharma. FDA Pre-IND Response (November 2025). Link
- WuXi Biologics & Toregem BioPharma MOU (October 2022). Link
- Absci. ABS-101 Phase 1 First Dosing (May 2025). Link
- Absci. ABS-201 Phase 1/2a HEADLINE Trial (December 2025). Link
- Insilico Medicine. First Generative AI Drug Enters Phase II. Link
- AlphaFold Protein Structure Database: SOSTDC1 (Q6X4U4). Link
- Pagella, P. et al. Single-cell atlas of healthy human teeth. Developmental Cell, 2023. Link
- Applications of AI in regenerative dentistry. Frontiers in Cell and Developmental Biology, 2024. Link
- EBI. AlphaFold: Strengths and Limitations. Link
Top comments (0)