v0.1 — Protein Function Reasoning
// protein input
sequence MVLSPADKTNVKAAWGKVGAHAGEYGAEAL...
organism Homo sapiens
// model reasoning
Analyzing primary sequence homology... Conserved domain detected: Globin-like fold. His63 and His92 coordinate heme Fe(II) binding. Context: hemoglobin alpha chain, oxygen transport. Confidence: 0.97
// expert RL signal
Reasoning chain validated. Evidence trace: GO:0006810, GO:0021700

Models that
think like
biologists.

Multimodal reasoning LLMs trained on biology tasks with expert RL feedback. Not just predicting function — reasoning through evidence the way a PhD scientist would.

What the model does

Chain-of-thought reasoning over biological context

Sequence → Structure → Function

Combines ESM2 protein embeddings, AlphaFold structural data, and biological literature into a single reasoning context. Not three models — one chain.

Expert RL Feedback Loop

PhD biologists rate reasoning chains, not just predictions. The model learns what constitutes sound scientific thinking — not pattern-matching to training data.

Explainable Trace Output

Every prediction comes with a human-readable reasoning trace: which domains, sequences, and evidence the model used. Auditable. Citable. Trustworthy.

Why current models fail at scientific reasoning

ProteinGPT, ESM, AlphaFold — these are exceptional at structure and function prediction. But they are descriptive, not explanatory. They tell you what a protein does. They don't tell you why, or how, or what the evidence for that conclusion is.

A biologist doesn't just classify — they trace evidence across sequence, structure, literature, and mechanistic context. AxiomBio trains models to replicate that chain of reasoning, then uses expert biologist feedback to reinforce it.

3
modalities in reasoning chain
×
expert RL over standard fine-tuning
verifiable reasoning traces
reasoning pipeline
01 Protein sequence input
02 ESM2 embeddings + AlphaFold structure
03 Biological literature grounding
04 Chain-of-thought reasoning generation
05 Expert biologist RL feedback
06 Reasoning-validated model weights

What makes AxiomBio different

Standard PLM
Multimodal LLM
AxiomBio
Sequence + structure input
Chain-of-thought reasoning
~
Expert RL feedback (not crowd)
Human-readable reasoning traces
Mechanistic explanation, not just label

The virtual cell is the target. Reasoning is the path.

Protein function is the first task. The reasoning architecture we're building extends to full cellular context — predicting how a protein behaves in a specific cellular environment, how it interacts with other molecules, how it responds to perturbation.

When we can model a virtual cell with mechanistic reasoning instead of statistical pattern-matching, we change what drug discovery looks like. Every experiment faster. Every hypothesis more grounded. Every failed candidate understood before it's synthesized.

2026
Protein function reasoning
Expert RL on biological reasoning chains. First validated benchmarks.
2027
Cellular context modeling
Protein-in-context: reasoning over pathways and subcellular localization.
2028
Virtual cell simulation
Mechanistic reasoning across full cellular network. In silico perturbation at scale.