Research and Disclosure at Axon Agentic

Validation at Axon Agentic means applying our systems to publicly available experimental datasets and measuring how well the predictions align with measured biology. The datasets are peer-reviewed deep mutational scanning A lab technique that measures the functional effect of every possible single amino acid substitution across a protein. The gold standard for variant effect data. Full definition assays, saturation genome editing A CRISPR-based assay that introduces every possible single-nucleotide variant at a genomic locus in its native context, then measures functional impact through cell viability or other selection. The gold standard for clinical variant classification in cancer genes. Full definition studies, and variant abundance measurements — experiments that exist independently of us and predate our analysis. We contribute the computational predictions; the experimental ground truth is external.

This is not a regulatory validation. Nothing here constitutes a clinical study or regulatory submission. Results are designated Research Use Only Research Use Only — a regulatory designation meaning the tool provides research scores, not clinical diagnoses. The same label used by REVEL, CADD, AlphaMissense, and PolyPhen-2. Full definition , the same designation used by REVEL, CADD, AlphaMissense, and PolyPhen-2.

The limitation disclosures below are not footnotes. The calmodulin result is listed at the same level as the positive results because accurate failure disclosure is part of what this page is for.

ESM-2 A protein language model by Meta AI trained on 250 million protein sequences. Predicts how amino acid mutations affect protein function from sequence alone — no structure required. Full definition 650M is the protein language model A deep learning model trained on millions of protein sequences to predict how mutations affect function. NeuroAutomata uses ESM-2, a PLM developed by Meta AI. Full definition at the core of NeuroAutomata. It was developed by Meta AI Research (Lin et al. 2023, Science) and predicts variant effects from sequence alone using masked marginal scoring ESM-2's zero-shot method for scoring variant effects. It masks each position and measures how surprising the mutant amino acid is relative to wild-type. Full definition — no structure required, no experimental data needed at inference time.

The published ProteinGym A standardized benchmark suite for protein variant effect predictors, covering 217 deep mutational scanning assays across diverse protein families. Full definition zero-shot substitution benchmark places ESM-2 650M at a mean Spearman rho A rank correlation coefficient (−1 to +1) that measures whether two variables agree in order, not magnitude. The primary metric for variant effect benchmarks. Full definition of 0.414 across 217 substitution DMS A lab technique that measures the functional effect of every possible single amino acid substitution across a protein. The gold standard for variant effect data. Full definition assays, ranking 45 of 97 models on the live leaderboard CSV (accessed 2026-05-08). That is the external published number.

Our internal validation — run on a 5-protein subset and self-reported — produced a median Spearman rho of 0.515, approximately 24% above the published ESM-2 650M baseline (internal benchmark — see validation details). The five proteins and their individual results:

The BRCA1 entry is the full ProteinGym BRCA1_HUMAN_Findlay_2018 dataset. Domain-level recomputations from the same Findlay 2018 saturation genome editing data show the BRCT domain alone at rho 0.534 (internal benchmark) and the RING domain at 0.409 (internal benchmark) — these are narrower scopes of the same dataset, not a separate benchmark. Sample sizes: BRCT N=1,262; RING N=575.

For methods, protein-specific results, and the full variant tables, see the validation details page or the NeuroAutomata product page.

Calmodulin is a calcium-binding protein whose functional effect depends on binding-affinity changes at protein-protein interfaces rather than on structural stability or folding. ESM-2 A protein language model by Meta AI trained on 250 million protein sequences. Predicts how amino acid mutations affect protein function from sequence alone — no structure required. Full definition scores evolutionary plausibility — patterns learned from sequence conservation across millions of proteins — not binding energy. On the calmodulin DMS A lab technique that measures the functional effect of every possible single amino acid substitution across a protein. The gold standard for variant effect data. Full definition dataset, ESM-2 produced a Spearman rho A rank correlation coefficient (−1 to +1) that measures whether two variables agree in order, not magnitude. The primary metric for variant effect benchmarks. Full definition of 0.212 against experimental data (internal benchmark, see validation details).

Protein-protein binding contexts are a known weak spot for this approach. Mutations that affect an interaction interface without destabilizing the fold are outside what evolutionary sequence conservation captures well. If your protein's function is primarily binding-affinity-driven, interpret ESM-2 scores cautiously and pair with structural or experimental data.

The HPA multi-agent system translates natural language queries into structured queries against the Human Protein Atlas JSON API. Over 18 months of development, the architecture evolved from a naive RAG Retrieval-Augmented Generation — a technique where an AI retrieves relevant documents from a database before generating a response, grounding answers in real data to reduce hallucinations. Full definition baseline through iterative revision to a verification-first multi-agent A network of specialized, autonomous AI agents that collaborate and divide complex tasks among themselves to solve problems too difficult for a single model. Full definition pipeline. The journey post documents the architecture decisions, the cross-validation approach used to confirm query results against HPA's own API responses, and what failed along the way.

Read: Building an AI Multi-Agent System for Human Protein Atlas Data — 18-month journey

The companion methodology post covers reproducibility protocols, agent role specifications, and the cross-validation procedure used to confirm query results against Human Protein Atlas API responses. It is intended for readers who want to understand the technical decisions behind the validation approach rather than just the results.

Read: Validation methodology — HPA multi-agent system

Validation tells you whether the system works. Verification tells you whether the claims we publish about it are accurate. These are not the same question.

Veritas is the AI agent responsible for verification at Axon Agentic. It operates separately from the agents that produce content — Amara (marketing), Astro (engineering), and Kiran (product). Veritas cannot edit content produced by other agents. Its findings cannot be overridden by other AI agents; any exception requires Jonathan Agoot's explicit written approval, recorded with a reason.

What Veritas does and does not cover is documented on the methodology page. The hub points there rather than reproducing the implementation here.

Every factual claim on Axon Agentic's public-facing pages is recorded in a structured YAML manifest. Each entry carries the claim text, source citation, evidence strength, and Veritas verification status. The files are served as plain text and are designed to be readable by both humans and AI language model Large Language Model — an AI model trained on vast text that can understand and generate human language. Examples: GPT-5, Claude, Gemini. Full definition systems that process structured data.

The claims manifest for this page and all published Axon Agentic content: axonagentic.ai/research/claims.yaml

Each substantive public-facing post has a paired Veritas verification report — a fact-check of the quantitative claims and source citations in that piece. Reports list each claim, the cited source, the verification result, and any corrections applied before publication.

Available reports are listed at /verification.

Verification audits the science. Disclosure audits the operation — who is on the team, what they can and cannot do, and what data practices cover the product.

Axon Agentic's team is six AI agents and one human. The human is Jonathan Agoot, who founded the company and is responsible for everything it publishes. No agent can approve its own factual claims or route around the Veritas verification step. No agent has access to customer data, personal information, or financial systems.

The AI Agent Staff page documents each agent's role, what it does and does not access, and the approval requirements for each category of output.

NeuroAutomata's privacy policy documents the data practices for the protein analysis platform. This is a platform-specific policy scoped to neuroautomata.axonagentic.ai — it is not an Axon Agentic site-wide policy.

Read: NeuroAutomata privacy policy

For NeuroAutomata:

Axon Agentic (2026). NeuroAutomata: ESM-2 protein analysis platform, on the latest production release. https://neuroautomata.axonagentic.ai

For the HPA multi-agent system:

Axon Agentic (2025). HPA Multi-Agent System: Verification-first natural language queries for Human Protein Atlas data. https://axonagentic.ai/blog/ai-natural-language-human-protein-atlas-18-month-journey

No DOI is assigned to this work at this time. No CITATION.cff is available (repositories are private).

If you are building AI systems for scientific research and want to discuss architecture or the verification approach, get in touch.

If you want to run your protein sequence through NeuroAutomata — score variants, view the mutation landscape, or explore ESM-2 predictions — the platform is available at neuroautomata.axonagentic.ai .