Biologix · Physics-grounded polymer discovery
Biologix is an agentic discovery platform that reasons over chemistry the way a scientist does — proposing candidate polymers, testing them against real physics, and closing the synthesis route before it moves on.
01 — The problem
Insulin and the biologics that follow it must be kept refrigerated. Where refrigeration is unreliable or absent, the medicine degrades before it can help — a supply problem hiding inside a materials problem.
A protective polymer matrix could stabilize these proteins at ambient temperature. But the space of possible polymers is astronomically large, and no experimental campaign can screen more than a sliver of it — let alone check whether each candidate can actually be made.
500M+
people worldwide depend on insulin
2–8°C
the cold chain today's supply is trapped inside
02 — The shift
An LLM proposes hypotheses and reads the literature — acting as the acquisition function that decides what to try next.
Every hypothesis, simulation result, and literature claim accumulates into a shared memory that sharpens each iteration.
Explicit molecular simulation rejects infeasible packings and implausible structures before they can steer the search.
03 — How it works
Candidate generation, molecular simulation, retrosynthesis, and safety screening operate as one loop — grounded in physics rather than interpolation, with a human checkpoint on every iteration.
Reads the field to seed and constrain the search.
Proposes and mutates polymer repeat units in PSMILES.
Scores each candidate against the target with OpenMM & Packmol.
Closes a viable synthesis route before the candidate advances.
Screens ADMET and excipient rules on every iteration.
04 — Capabilities
Resolve an arbitrary biologic and the same workflow runs unchanged — insulin and adalimumab shared one pipeline.
Proposes and mutates candidate repeat units across the discrete PSMILES space under a fixed evaluation budget.
OpenMM and Packmol matrix screening scores each candidate against the real target structure, not a learned surface.
Agent-backed retrosynthesis and a four-tier precursor registry close a viable route before a candidate advances.
ADMET and excipient-compliance checks run inline on every loop — recorded as audit artifacts, not bolted on at the end.
The science stage is CPU-bound, compatible with local and edge LLM deployments — no specialized cluster required.
The only true measurement — but no campaign can synthesize and test more than a sliver of a combinatorial space.
Learned predictors are quick, yet they interpolate — they can't reject sterically infeasible packings or implausible structures.
An LLM decides what to try; explicit physics judges it; retrosynthesis and safety gate it — every iteration grounded, not interpolated.
05 — The evidence
Independent campaigns on insulin and adalimumab ran the exact same platform, unchanged — each converging on a synthesizable candidate polymer and outperforming reinforcement-learning and Bayesian-optimization baselines under matched budgets.
Lactate comonomers added the hydrogen-bonding the protein needed — a candidate the model reasoned toward and physics confirmed.
Bulky designs failed to pack; the platform learned that within the session and settled on a linear backbone that both packed and scored best.
The loop that stabilized insulin could hold for any biologic where a screening oracle exists. We're building toward a world where a medicine's reach is no longer decided by whether it can be kept cold.
Algonix AI
Edinburgh, Scotland — Algonix AI and NovoMCP have entered a partnership under which NovoMCP will use Biologix, Algonix's agentic platform for polymer-excipient discovery, within its own computational chemistry ecosystem.
NovoMCP operates an open engine that AI agents call through the Model Context Protocol. Through the partnership, its work gains Biologix's physics-grounded, synthesis-aware polymer discovery — the same workflow Algonix used to find candidate stabilizers for insulin and adalimumab.
Read the announcement →“Biologix was built to run anywhere a screening oracle exists. Seeing it put to work inside NovoMCP's engine is exactly the reach we designed for.”
Founder, Algonix AIQuestions
Physics checks reject infeasible packings and name–structure mismatches before they can steer the next iteration.
OpenMM Packmol-matrix screening, evaluating interaction energy between candidate polymers and the specified protein structure.
Yes — agent-backed retrosynthesis closes a route, seeded from a four-tier precursor registry, before a candidate advances.
Insulin (PDB 4F1C) and adalimumab (PDB 3WD5), each over five human-in-the-loop iterations of the same unchanged workflow.
The science stage is CPU-bound and runs on commodity hardware, compatible with local and edge LLM deployments.
Every iteration logs hypotheses, simulation outcomes, ADMET, and compliance alongside the structure files it produces.
06 — Research
Bring us a target and a screening oracle. We'll run the same loop that found candidates for insulin and adalimumab.
Request a demo →