Digital Skin

Discovery

Pipeline

Contact

PELLICLE

Live

Manifesto

Skin is the largest organ. We made it computable.

Pellicle is an end-to-end AI laboratory for skin science. We discover novel cosmetic ingredients, simulate how they permeate human skin, formulate them into products, and deliver lab-ready candidates with patent claims — in days, not years. From market signal to CRO submission. One system. Zero guesswork.

We built a digital replica of the human stratum corneum — the 15-micron barrier that stands between a molecule and the bloodstream. Our system reasons over 716,000 experimental data points from real Franz cell studies on human skin. It predicts permeation with R² = 0.91 accuracy. No molecule reaches a lab bench without passing through us first.

Discover

10-loop autonomous pipeline. From trending ingredients to novel molecular candidates. AI generates, searches patents, validates novelty, scores permeability, checks safety, estimates synthesis cost, evaluates formulation, assesses commercial viability, drafts patent claims, and produces lab protocols. 1,000 candidates in, 5 validated out.

Simulate

The Digital Skin. A reasoning engine that takes any SMILES string, finds the closest experimental analogs in our database, applies Potts-Guy corrections, nearest-neighbor blending, and physicochemical reasoning to predict logKp — the skin permeability coefficient. R² = 0.91 on 67 blind compounds. Matches state-of-art machine learning. Explains every prediction.

Deliver

Synthesis protocols. Formulation guidelines. YUKA-scored ingredient lists. Cost per unit calculations. Patent claim drafts. CRO submission packages ready for Charles River, Eurofins, or Tioga Research. Everything a brand needs to go from molecule to shelf. We hand you the keys.

Prediction Accuracy

0.91R²

Experimental Database

716Kdata points

Validation Pipeline

10sequential loops

Time to Candidate

<48hours

Live Simulation

Molecular Permeation Field

Stratum Corneum · 15μm

Permeated

SC Retained

Blocked

PERMEATED: 0
SC_RETAINED: 0
BLOCKED: 0
FLUX: 0.00 μg/cm²/h
[INTERACT TO ALTER PERMEATION FIELD]

First Discovery

N-Octyl Nicotinamide Patent Pending

A lipidated prodrug of niacinamide — the world's most popular skincare active — engineered for 100× improved skin penetration. Discovered by our pipeline, independently validated by expert formulation chemists, and currently entering synthesis. Effective at 0.5% instead of the standard 5-10%.

Novel lipidated prodrug · Series patent filed (C4–C14 alkyl chain variants) · 1-step synthesis · Currently entering lab validation

Molecular permeation through skin membrane

Lipidated molecules crossing the stratum corneum barrier

System Log

25/03

N-Octyl Nicotinamide greenlit for synthesis Breakthrough

Independent expert validation confirmed. "The synthesis of this ester seems much more opportune given the probability of increasing penetration." Entering lab phase at partner facility.

24/03

Digital Skin validated: R²=0.91 on 67 blind compounds

First LLM-based skin permeation prediction system. Matches XGBoost/LGBM state-of-art (2024). 94% of predictions within 1.0 log unit. Zero catastrophic failures. Outperforms fine-tuned models by 3×.

23/03

Synergy pipeline discovered 18 novel cosmetic combinations

8-loop combination discovery. 3 targets: anti-acne, retinol alternative, anti-aging. Top candidate: Carnosine + Fisetin + EGCG — first senolytic cosmetic complex. All YUKA-friendly, all under €1.33/unit.

22/03

Full discovery pipeline: 52 novel molecules → 5 CRO-ready candidates

10-loop pipeline completed first run. Market signal → chemical design → novelty search → permeation gate → safety → synthesis → formulation → commercial → patent → lab protocol. 3 provisional patent claims drafted.

The Approach

Why reasoning beats regression

Every existing model predicts a number. We built a system that reasons to a number.

Traditional QSPR models (Potts-Guy, Random Forest, XGBoost) treat skin permeation as a regression problem: molecular descriptors in, logKp out. Black box. No explanation. No context. They fail on novel scaffolds because they extrapolate from curves, not from chemistry.

Pellicle is fundamentally different. We developed a compound reasoning architecture built on a 235-billion parameter foundation model, fine-tuned on skin permeation science and augmented with the full experimental corpus at inference time. For every query molecule, the system identifies structurally similar compounds with real measured logKp values from human skin, applies physicochemical corrections informed by decades of published research, and reasons through the prediction like a senior scientist would. It explains why — which molecular features drive permeation, which pathway dominates, and crucially, when it doesn't know and real lab testing is required.

No model in the published literature has attempted this. We are the first team to apply frontier LLM reasoning to skin permeation science.

Benchmark

67 held-out compounds
Never seen during training

Validated on compounds the system has never encountered. Blind test. Real experimental ground truth.

Potts-Guy '92

0.67

Not explainable

Abraham LFER '04

0.80

Partial

Random Forest '20

0.85

Black box

XGBoost '24

0.90

Black box

Pellicle '26

0.91

Full reasoning chain

Infrastructure

What runs under the hood

235B

Parameter foundation model

Fine-tuned on skin permeation data via LoRA. Qwen3-235B-A22B architecture. Trained on Tinker distributed infrastructure across multiple GPU clusters. Specialized for structured JSON prediction with molecular reasoning.

716K

Experimental data points

Sourced from HuskinDB and SkinPiX — the two gold-standard databases for human skin permeation. Every measurement from real Franz cell studies on real human skin. 335 unique compounds across 15 experimental conditions.

Sequential validation loops

Market Signal → Chemical Design → Novelty Gate → Permeability Gate → Safety Gate → Synthesis Gate → Formulation Gate → Commercial Gate → Patent Gate → CRO Submission. Each loop is an autonomous protocol any AI agent can execute.

94%

Predictions within 1.0 log unit

On 67 blind compounds, 94% of predictions fell within 1.0 log unit of the experimentally measured value. 76% within 0.5 log units. Zero catastrophic failures. 100% within 1.5 log units.

We trained specialized models at 8B, 32B, and 235B parameters. All scored R² < 0.45. Then we developed a compound reasoning architecture — a foundation model that ingests the full experimental corpus and reasons through each prediction like a human expert. R² jumped to 0.91 overnight. The breakthrough wasn't more data or bigger models — it was a fundamentally different way of thinking about skin.

The Pipeline

From market signal to lab bench

LOOP 01 Market Signal
What ingredient is trending? What delivery problem exists?

LOOP 02 Chemical Design
Generate 3-10 molecular candidates per active

LOOP 03 Novelty Gate
Live patent search. PubChem. INCI. Google Patents.

LOOP 04 Permeability Gate
Digital Skin predicts logKp. R² = 0.91.

LOOP 05 Safety Gate
Structural alerts, PAINS, phototoxicity, irritation

LOOP 06 Synthesis Gate
Can it be made for <€500/kg? How many steps?

LOOP 07 Formulation Gate
Solubility, stability, sensory, product concept

LOOP 08 Commercial Gate
Would L'Oréal pay for this? Market size. Revenue model.

LOOP 09 Patent Gate
FTO analysis. Claim drafting. Competitive moat.

LOOP 10 CRO Submission
Full IVPT protocol. Budget. Timeline. Ready to send.

Each loop is a standalone protocol. Each is fully autonomous. The entire pipeline runs in under 48 hours. First run: 52 candidates in, 5 out. Three provisional patent claims drafted. Zero lab costs.

Built on real science

HuskinDB SkinPiX Potts-Guy Model Abraham LFER OECD TG 428 RDKit Tinker

Work With Us

You bring the molecule. We bring the skin.

For cosmetic brands, ingredient suppliers, pharmaceutical companies, and research laboratories. Discover what crosses — before you spend a dollar in the lab.

Science Discovery Pipeline Contact