# Rexis: The L1 for DeSci

*The first end-to-end platform for Biomedical Decentralized Science (BioDeSci).*

## Introduction to Rexis

**Rexis** is designed for a new era of biomedical discovery—one where large-scale machine learning on sensitive patient data is not just possible, but secure, verifiable, and decentralized.

It introduces the first **end-to-end (e2e) platform** that unifies:

* **Secure data sharing**
* **Privacy-preserving model fine-tuning and training**
* **Decentralized evaluation**

within a single, privacy-first ecosystem (van Rossum et al., 2023).

***

### Why Centralized Biomedical AI Falls Short

Today’s biomedical workflows remain fragmented, insecure, and siloed. Researchers often rely on ad hoc data transfers—via email, FTP, or cloud storage—which are incompatible with modern privacy standards like HIPAA and GDPR (Mora et al., 2022; Kim et al., 2021).

Clinical and genomic datasets are too sensitive to share, yet large-scale AI models require data diversity far beyond what any single institution can provide. Partial solutions—like federated learning (FL), differential privacy (DP), and secure enclaves (TEE)—offer incremental gains but suffer from high overhead, limited scalability, or weak guarantees (Abadi et al., 2016; Zhu et al., 2019; Costan & Devadas, 2016).

Without a purpose-built system for secure, collaborative computation across institutional boundaries, biomedical AI cannot reach its full potential.

***

### The Rexis Architecture

At the core of Rexis lies **Equivariant Encryption (EE)**—a cryptographic method that transforms both model parameters and data into an encrypted space. EE supports operations like matrix multiplication and non-linear activation in the encrypted domain, enabling full ML pipelines to run without decrypting inputs, activations, or gradients (Gilad-Bachrach et al., 2016; Brutzkus et al., 2019).

#### 🔐 Secure Data Sharing

Rexis enables a decentralized biomedical data marketplace. Institutions tokenize datasets as NFT-like digital assets with embedded on-chain metadata for provenance, ownership, and access policy enforcement.

Instead of exchanging raw files, researchers submit **compute-to-data queries**, which execute securely over encrypted data within trusted smart contract frameworks (Hasselgren et al., 2020). This ensures regulatory compliance and robust auditability—without ever revealing patient-level information.

#### 🧠 Privacy-Preserving Model Fine-Tuning

Rexis supports secure fine-tuning of biomedical foundation models (e.g., BioBERT, ClinicalBERT) using encrypted electronic health records or genomic data (Lee et al., 2020; Alsentzer et al., 2019).

Each institution performs encrypted training locally and shares only encrypted gradient updates with a decentralized aggregator. No patient records, model weights, or activations are ever exposed—satisfying both fidelity and privacy requirements (Gu et al., 2021; Cho et al., 2022).

#### ✅ Decentralized Evaluation

To support reproducibility and regulatory audit, Rexis implements **verifiable inference**. Models are evaluated on encrypted test data across multiple sources. The results are certified using lightweight cryptographic commitments and stored on-chain (Zhang et al., 2023; Mora et al., 2022).

This ensures transparency and trust—without compromising confidentiality.

***

### Why Rexis Is Different

Rexis is the **first platform** to integrate sharing, training, and evaluation within a **unified, cryptographically secure stack**:

| Capability            | Rexis                              | Other DeSci Platforms           |
| --------------------- | ---------------------------------- | ------------------------------- |
| Data sharing          | Tokenized assets + compute-to-data | IPFS or metadata-only solutions |
| Model fine-tuning     | Fully encrypted via EE             | Not supported                   |
| Verifiable evaluation | Encrypted inference + proofs       | Not supported                   |
| Regulatory compliance | HIPAA/GDPR aligned design          | Varies; typically incomplete    |

While other DeSci projects focus on governance or data funding (e.g., Molecule, VitaDAO), Rexis directly supports **computation** on sensitive biomedical data—making it usable in real-world research and clinical pipelines.

***

### Representative DeSci Projects

| Platform                                       | Focus Area             | Sharing Mechanism      | Privacy            | Example Use Case                      |
| ---------------------------------------------- | ---------------------- | ---------------------- | ------------------ | ------------------------------------- |
| [DeSci Labs](https://www.desci.com/)           | Research storage       | Cloud + AI             | N/A                | Secure data repository                |
| [Molecule](https://www.molecule.xyz/)          | Genomics collaboration | Blockchain peer review | N/A                | Tokenizing IP for review              |
| [VitaDAO](https://vitadao.com/)                | Longevity research     | DAO funding            | N/A                | Crowdsourcing trials                  |
| [BIO Protocol](https://www.bio.xyz/)           | Biotech IP             | Tokenized IP + funding | N/A                | Funding research projects             |
| [Fleming Protocol](http://flemingprotocol.io/) | Patient data ownership | IPFS                   | Patient-controlled | Sharing data for rare disease studies |

Rexis builds on these foundations by enabling not just collaboration and funding—but **actual AI computation over encrypted biomedical data**.

***

### The Future of BioDeSci

Decentralized science (DeSci) offers a compelling vision: open, transparent, and equitable research infrastructure. But biomedical domains demand more—they require privacy, reproducibility, and regulatory integrity (van Rossum et al., 2023; Mora et al., 2022).

**Equivariant Encryption (EE)** makes this possible. By enabling compliant, cross-institutional computation without revealing raw data, Rexis supports:

* Multi-site clinical research collaboration
* Genomic and EHR-driven personalized medicine
* Federated pipelines for drug discovery
* On-chain model certification and reproducibility auditing

***

### What Comes Next

This documentation covers:

1. **The systemic barriers** to trustworthy biomedical AI
2. **The cryptographic design** of Equivariant Encryption
3. **Real-world case studies** showing how Rexis enables private training and decentralized evaluation

Rexis sets a new standard for biomedical AI—**making privacy the default**, not the exception.

***

#### References

* Abadi, M., et al. (2016). Deep learning with differential privacy. *CCS*.
* Alsentzer, E., et al. (2019). ClinicalBERT embeddings. *Clinical NLP Workshop*.
* Brutzkus, A., et al. (2019). Low latency privacy-preserving inference. *ICML*.
* Cho, H., et al. (2022). Privacy-preserving genomic analysis. *ARBMDS*.
* Costan, V., & Devadas, S. (2016). Intel SGX explained. *IACR ePrint*.
* Gilad-Bachrach, R., et al. (2016). CryptoNets: Neural networks on encrypted data. *ICML*.
* Gu, Y., et al. (2021). Biomedical domain-specific pretraining. *ACM TCH*.
* Hasselgren, A., et al. (2020). Blockchain in healthcare. *J Med Internet Res*.
* Kim, M., et al. (2021). Federated learning in medicine. *JAMIA*.
* Lee, J., et al. (2020). BioBERT: Biomedical language models. *Bioinformatics*.
* Mora, C., et al. (2022). Promises and pitfalls of DeSci. *Science*.
* van Rossum, T., et al. (2023). Decentralized biomedical research. *Nature Biotech*.
* Zhang, Y., et al. (2023). ZK proofs for privacy-preserving ML. *ACM Computing Surveys*.