Biology’s EDA Moment: Inside Enhe Tech’s BPL Language and Verification Architecture

In December 2025 OpenAI and partner labs released a wet‑lab report showing that GPT‑5’s multi‑round optimization of a molecular cloning protocol boosted efficiency by 79×, yet when the same protocol was handed to a robot the colony count dropped ten‑fold, highlighting a critical "translation" bottleneck between LLM‑generated natural language and deterministic machine instructions.

A 2016 Nature survey of 1,576 biologists (Baker) revealed that over 75% could not reproduce others' experiments and 60% could not reproduce their own, a problem traced to information loss during natural‑language communication. The article argues that biology lacks an equivalent to hardware description languages that eliminate such ambiguity.

Enhe Technology introduced BPL (Biology Protocol Language) and its supporting pipeline BPL‑COGEN. BPL is organized into six layers that progressively remove natural‑language fuzziness:

Layer 1 – Declaration: Precise material lists with exact physical attributes (solid vs. solution, concentration, container).

Layer 2 – Biological Primitive Type System: Nine physical dimensions (volume, mass, temperature, etc.) with ~40 units; dimensional analysis catches mismatches at compile time.

Layer 3 – Experiment Intent (14 atomic operations): Transfer, mix, incubate, run_pcr, centrifuge, pick_colonies, etc., plus a manual intent for human‑in‑the‑loop steps.

Layer 4 – Container State Engine: Real‑time tracking of each well’s volume, contents, temperature, and physical form.

Layer 5 – Trust Model & Compliance: Three‑tier trust (Declared, Calibrated, Verified) with native GLP/GMP/21 CFR Part 11 annotations.

Layer 6 – Control Flow: Conditional branches, loops, parallel blocks, and structured error recovery.

Above the layers, an "intent lowering" stage compiles high‑level intents into platform‑specific primitives, enabling the same BPL source to generate human‑readable guides, robot instruction files, or simulation inputs without modification.

BPL‑COGEN combines a fine‑tuned 30B‑parameter LLM (BPL‑Nano‑30B, Nemotron‑based) with a deterministic compiler. The pipeline runs five stages: input normalization, code generation (using a full 463‑line Lark PEG grammar), three‑stage compiler verification (syntax, semantics, planning), diagnostic feedback, and iterative repair (up to three loops). Diagnostics across a benchmark suite produced 343 messages, most frequently dimensional mismatches (142, 41.4%). The repair loop resolved 98.6% of variants within two iterations.

Two concrete case studies demonstrate physical validity:

GFP plasmid library construction: BPL‑COGEN generated identical protocols for human operators and a Biomek i7 liquid‑handling robot, yielding comparable transformation success and fluorescence gradients.

HPLC‑to‑UHPLC method migration: Automatic conversion reduced run time from 32 min to 2.1 min, cut solvent use by 95.8%, and achieved baseline separation for five compounds, while four compliance warnings were logged and manually approved.

In a large‑scale test on 300 Nature Protocols papers (2,992 generated variants), BPL‑COGEN achieved an average fidelity score of 95.1 ± 8.3 and structural consistency in 295 papers, preserving step order across ten independent generations.

Building on BPL, Enhe launched SAION AI, a Physical AI platform with a three‑layer Cognitive‑Control‑Execution (COE) architecture. The Cognitive layer aggregates millions of closed‑loop experiment data, literature, patents, and AI‑4‑Science models (AlphaFold, ProteinMPNN, etc.). The Control layer uses an Agent Harness to decompose research goals into structured task graphs and orchestrate 316 specialized tools. The Execution layer runs BPL‑derived protocols on liquid‑handling workstations, feeding back data into a Design‑Build‑Test‑Learn (DBTL) loop.

SAION AI’s benchmark performance surpasses leading LLMs (GPT‑5.3, Claude Opus 4.6) on LitQA+SuppQA (70.7% accuracy), Seq QA (88.2% vs. 81.9%), gene‑editing/cloning (84.9%), and scientific discovery (89.6%). Real‑world deployments include the FDA‑GRAS‑approved ZeaVida® corn‑yellow‑pigment production.

The article draws an analogy to electronic design automation (EDA), suggesting that BPL could enable a "Labless" model where experimental design and physical execution are decoupled, much like how EDA enabled fabless semiconductor manufacturing. The authors envision BPL evolving into an open, industry‑wide protocol standard akin to HTTP or MQTT, providing a precise, verifiable bridge between AI‑generated designs and wet‑lab reality.