Rare Earth × Protein Design AI

Rare Earth Separation, Rewritten with Designed Proteins.

DFRN designs proteins that selectively bind rare earth elements with AI, then validates them in our own Wet Lab. Our fully aqueous process replaces organic-solvent extraction, and is built to produce and supply neodymium and dysprosium oxides in Korea.

DFRN / 2026 SEOUL · DAEJEON
1–3 stagestarget
Nd/Dy Separation Stages
2N–4N
Supplied Oxide Purity
0 solvent
Organic Solvent · Fully Aqueous

Heavy rare earths are not a technology problem.
They are a supply chain problem.

NdFeB permanent magnets sit at the core of EV traction motors, wind turbines and robotic actuators. Yet the heavy rare earths that give those magnets their thermal stability, dysprosium and terbium, are separated and refined in only a handful of countries, and Korea has virtually no commercial-scale separation capacity of its own. Securing strategic resources is no longer a question of price; it is a question of access.

99%
Global Heavy REE Processing

Separation and processing capacity for Dy/Tb-class heavy rare earths sits with Chinese firms. It is effectively a monopoly.

SRC. Covington, Heavy Rare Earth Elements (Feb 2026)
89.4%
Korea's REE Feedstock from China

89.4% of Korea's rare earth feedstock imports come from China, and for some high-value elements the figure exceeds 95%.

SRC. Korea Customs Service, Jan–Sep 2025
SX
Organic-Solvent Separation

Conventional solvent extraction requires tens to hundreds of stages and large volumes of organic solvent. In Korea, permitting and environmental load then become a direct barrier to entry.

SRC. hydrometallurgical practice

We design proteins that bind
only the target metal,
and rebuild separation around them.

Lanmodulin (LanM) is a natural protein that binds rare earth ions with remarkable selectivity. DFRN generates LanM variants with a protein-sequence AI model, then validates binding strength and Nd/Dy selectivity in our own Wet Lab. Every result feeds back into training, closing a design–validate–learn loop.

STAGE 01 AI Protein Design LanM Variants STAGE 02 Wet Lab Binding · Selectivity STAGE 03 Oxide Supply Nd₂O₃ · Dy₂O₃ AI-BIO LOOP circl:o PLATFORM

Design → Validate → Learn → Recover

It does not stop at AI-generated candidates. Expression and binding data from our in-house Wet Lab is fed back into the model, driving selectivity and hit rate upward cycle by cycle. The proteins that survive go into a real separation process as a working unit operation. The loop is not bound to one target. It is a transferable engine.

01
Generative AI for Protein Sequences
A proprietary model (Llama + stacked LoRA) trained on 460M protein data points generates target sequences.
02
Selective by Design
LanM binding sites are re-engineered to distinguish between rare earth ions, specifically Nd and Dy.
03
AI × Wet Lab Closed Loop
Synthesis, expression and binding results return to the training set, raising the hit rate every cycle.
04
Sustainable Recovery of High-Value Resources
Rare earths recovered from end-of-life NdFeB magnets and process scrap, supplied as oxide materials.
Positioning
Statement
DFRN does not license technology or sell reagents. We are a materials supplier, delivering Nd₂O₃, Dy₂O₃ and related oxides at 2N–4N purity. LanM-based Nd/Dy separation is the core unit operation that makes that supply possible.

A proven loop.
A transferable engine.

DFRN's technology rests on three things: the validation record proving the platform actually works, the position we occupy as a unit operation inside the rare earth process, and the competitive edge of the AI model underneath it all.

PART A Validated Engine

We proved the loop works on a hard enzyme target first.

Designing a PET-degrading enzyme took DFRN through a full cycle of sequence generation, screening, synthesis, activity validation and retraining in under twelve months. The target material has changed. The model, the data and the integrated Dry & Wet Lab infrastructure carried straight over to rare-earth-binding protein design.

Jul 2025
Sequence-Based Model Development
Dry Lab established
Jan 2026
1st Synthesis · 32 Variants
38%Protein Expression
Mar 2026
2nd Synthesis · 20 Variants
Advanced Screening
75%Protein Expression
Apr 2026
3rd Synthesis · 120 Variants
3+ High-Activity Enzymes
15%Activity Expression
2–3months
Multiple functional candidates out of a single cycle.
The cycle itself is the asset, and it runs the same whether the target is PET or a rare earth ion.
* Part A figures come from the PET-degrading enzyme program and serve as platform validation data. DFRN's current commercial target is rare earths.
PART B Nd/Dy Separation Unit Operation

An aqueous unit operation that replaces solvent extraction.

Splitting Nd from Dy in a leached rare earth solution is the hardest and most expensive step in the whole chain. DFRN handles exactly that step with a LanM-based protein instead of organic solvent. Upstream and downstream steps (leaching and calcination) run on partner technology; DFRN drops its separation unit into the middle.

STEP 01
Feedstock
End-of-life NdFeB magnets
Process scrap
STEP 02
Leaching
REE into solution
Partner upstream tech
STEP 03 · DFRN
LanM Selective Binding & Release
1–3targetbind / release stages
STEP 04
Precipitation & Calcination
Conversion to oxide
STEP 05
Oxide Shipment
2N–4NNd₂O₃ · Dy₂O₃
Aqueous Process

Separation in water,
not in solvent.

SX
Organic
multi-stage
DFRN
Aqueous
water-based
Lower permitting burden
  • No organic solvent storage or handling systems
  • No spent-solvent disposal stream
  • A real advantage for siting and permitting in Korea
Selectivity & Capacity

Tell them apart.
Hold more of them.

Wild-type LanM
4
EF-hand sites
DFRN design
6target
EF-hand sites
Selectivity and capacity together
  • Re-engineered binding sites for Nd/Dy selectivity
  • The wild type binds only 2–3 ions in practice. Raising molar capacity per protein is what makes the process commercial
  • Continuously improved through the AI and Wet Lab loop
Process Footprint

Fewer stages.
Smaller plant.

Hydromet (magnet)
~20
stages
DFRN
1–3target
stages
Reduced capex & footprint
  • Refining primary ore takes tens to hundreds of stages, but end-of-life NdFeB magnets reach high purity in roughly 20 hydrometallurgical stages
  • Selective protein binding aims to bring that down to a single digit
  • Small modular units, operating at ambient temperature and pressure
Reusability

Use the protein
again and again.

Reported to date
10–20
bind / release cycles
DFRN
100target
bind–release cycles
The variable that sets unit cost
  • Cycle count is the denominator under protein cost. Going from 10 to 100 cuts that cost contribution to a tenth
  • The immobilisation support and the release conditions are designed together for durability
  • A target set for commercial operation, not for a research yield
* Figures marked target are goals DFRN is working toward, not demonstrated results. The comparison baseline is hydrometallurgical processing of end-of-life NdFeB magnets, not primary ore refining. Measured values will replace these once demonstration is complete.
PART C AI Model's Competitive Edge

Faster. More versatile. A measurable advantage.

Compared with competing protein generation models, training cost is 1/9 and memory usage 1/1.5. Protein length generation is unlimited, and LoRA fine-tuning switches targets quickly. That is the structural reason a pivot of this kind was possible at all.

Cost-effective

No large-scale
hardware infrastructure needed.

Competitor
40,968
GB
DFRN
4,560
GB
8.9× Training Cost Reduction
  • Low-power, high-efficiency model with a low carbon footprint
  • Operable on compact infrastructure
Parallel Processing

Multiple projects
running simultaneously.

Competitor
3,100
MB
DFRN
2,090
MB
1.5× Memory Efficiency
  • Concurrent multi-project execution
  • Optimal per-target batch operation (Design → Synthesis → Retrain)
Universal & Scalable

Unlimited protein
length generation.

Competitor
300–500
AA
DFRN
UNLIMITED
AA
Versatility & Expandability
  • Rapid target switching via LoRA
  • Extends from degrading enzymes to metal-binding proteins
We are Differian.

From an AI-Bio platform
to a rare earth materials supply.

The path from selling technology to supplying material. Three phases, executed in order and proven by results rather than promises.

Ph#1
Validated

AI-Bio Platform Validation

via PET-degrading enzyme

We proved on a genuinely hard enzyme target that the design–validate–learn loop produces functional proteins in the real world.

  • High-difficulty PETase designed and validated
  • 15% activity expression rate achieved
  • Integrated Dry & Wet Lab infrastructure built
Ph#2
In progress

Rare Earth Binding Proteins & Demonstration

LanM · Nd/Dy Separation

Designing and validating LanM variants, then inserting DFRN's separation unit into a demonstration line that runs from leaching through calcination.

  • LanM variant design and Nd/Dy selectivity validation
  • Upstream (leaching to calcination) technology transfer partnership signed with Mokpo National University, Sep 2026
  • DFRN's separation unit inserted into a demonstration line running from leaching through calcination
Ph#3
Planned

Commercial Oxide Supply

Nd₂O₃ · Dy₂O₃ · 2N–4N

Scaling the demonstration line to commercial capacity and supplying rare earth oxides directly to domestic magnet and component manufacturers.

  • Pilot facility → commercial plant scale-up
  • Tolling model for end-of-life magnets and process scrap
  • A new pillar in Korea's rare earth supply chain

Who we work with.

Our business is supplying material. That opens three directions of partnership: those who send us feedstock, those who receive the oxide, and those who validate the process alongside us.

Materials
Magnet & Component Makers

Nd₂O₃ and Dy₂O₃ supplied to NdFeB magnet and component manufacturers. Purity (2N–4N) and form factor are open to specification.

Tolling
Scrap & Waste Processors

A tolling model that recovers rare earths from end-of-life NdFeB magnets and process scrap. The aqueous process keeps the domestic permitting burden low.

R&D · Public
Research & Public Institutions

Joint demonstration combining upstream leaching and calcination technology with DFRN's separation unit. Universities, national labs and public agencies are welcome.

Milestones & History

Rare Earth Program
Sep 2026
Upstream (leaching to calcination) technology transfer partnership signed with Mokpo National University
Platform Validation · AI-Bio
Apr 2026
Target material degraded successfully by an enzyme generated with our own AI model
Jan 2026
Successful PET depolymerization using an in-house biocatalyst
Certifications · Investment · Awards
Jun 2026
Selected for TIPS (Tech Incubator Program for Startup Korea) · KRW 800M total, approx. USD 0.6M (Ministry of SMEs and Startups)
May 2026
Selected for the Youth Startup Academy, 16th cohort (KOSME)
Dec 2025
Selected as KIBO (Korea Technology Finance Corp.) Star Valley Company
Dec 2025
Approved as Daejeon Corporate R&D Center
Aug 2025
Industrial PET workwear resource recovery research contract signed (Lindström)
Aug 2025
Venture company certification
Jul 2025
Seed investment secured (Korea Investment Accelerator)
Jun 2025
KIBO Venture Camp 16th cohort — Top 10 Outstanding Companies
Nov 2024
1st Asan University Demo Day — Encouragement Award (Climate Tech Track; now the Chung Ju-yung Entrepreneurship Competition)
Nov 2024
5th KAIST Lean Startup Camp — Excellence Award
Sep 2024
KAIST Convergence Capstone Design — Selected Company
Aug 2024
KAIST Deep-Tech Startup Rapid Prototype Support Program — Selected
— DIFFERENT FUTURE THROUGH REACTION IN NATURE —

Looking for partners to rewrite
the rare earth supply chain with us.

Oxide supply, scrap tolling, joint demonstration, investment. Every inquiry is welcome.
Before this year is out, we intend to prove it with results.

value@circlo.kr