I fed my longevity stack proteins to BioReason-Pro — here’s what the AI biologist found
Arc Institute just released BioReason-Pro — the first multimodal reasoning model for protein function prediction. It doesn’t just label proteins; it reasons through them step by step, like a biologist colleague explaining their logic. Human experts preferred its annotations over curated UniProt entries 79% of the time. It’s free, open-source, and runs in your browser at app.bioreason.net.
I decided to test it on something personal: five key proteins from my longevity protocol — NAMPT, SIRT1, AMPK α2, AMPK β2, and mTOR. I gave the model only raw amino acid sequences and “Homo sapiens.” No hints, no context. Here’s what it found.
NAMPT — the bottleneck of your NAD+ pathway
Protein: Nicotinamide phosphoribosyltransferase, 491 aa (UniProt: P43490) Supplement: NMN
The model correctly identified the core function: phosphoribosyl transfer from PRPP to nicotinamide, producing NMN. Standard stuff. But then it predicted cytokine activity and receptor ligand activity — from sequence alone.
This matters because NAMPT exists in two forms: intracellular (enzyme, makes NMN) and secreted (eNAMPT/visfatin — acts as a cytokine and adipokine). The model caught this duality without any hint. Secreted eNAMPT declines with age and correlates with remaining lifespan in mice.
It also predicted nuclear localization (nuclear body, nucleoplasm) — consistent with NAMPT’s role in regulating NAD±dependent nuclear processes like PARP1 activity and sirtuin function.
Takeaway: We take NMN as a substrate, but NAMPT expression itself is the rate-limiting bottleneck. If NAMPT is downregulated, excess nicotinamide won’t help.
SIRT1 — the master regulator fed by NAD+
Protein: NAD-dependent deacetylase sirtuin-1, 747 aa (UniProt: Q96EB6) Supplement: Resveratrol (putative activator)
This is where the model really shined. The reasoning trace was massive, and the predicted biological processes read like a longevity textbook:
- Circadian rhythm + circadian regulation of gene expression — the NAMPT→NAD+→SIRT1→BMAL1 feedback loop
- Negative regulation of TOR signaling — SIRT1 activates TSC2, which inhibits mTOR. Same target as rapamycin, different path
- Positive regulation of autophagy/macroautophagy — SIRT1 deacetylates ATG5/7, triggering cellular cleanup
- Negative regulation of NF-κB — anti-inflammatory effect, directly relevant to inflammaging
- Positive regulation of gluconeogenesis + insulin receptor signaling — intersects with metformin via the AMPK→SIRT1 axis
- Negative regulation of cellular senescence — self-explanatory for longevity
For localization, it predicted PML bodies (linked to senescence and DNA repair) and rDNA heterochromatin — whose stability degrades with age, and SIRT1 helps maintain it. All from sequence.
AMPK — the energy switch
Proteins: Catalytic α2 subunit, 552 aa (UniProt: P54646) + Regulatory β2 subunit, 272 aa (UniProt: Q9Y478) Drug: Metformin
I ran both subunits separately.
For β2, the model correctly identified it as a non-enzymatic scaffold — not a catalyst, but the organizer. It found the glycogen-binding domain (positions AMPK near carbohydrate stores) and the ASC assembly domain (nucleates the heterotrimer). The model called it “a mechanistic bridge from carbohydrate status to kinase-driven responses” — an elegant description of how AMPK actually works.
For α2 (the catalytic subunit), the model decomposed four functional blocks: kinase domain → autoinhibitor → adenylate sensor → membrane targeting. Key predicted processes:
- Positive regulation of autophagy/macroautophagy — AMPK phosphorylates ULK1, same endpoint as SIRT1 but via a different mechanism
- Response to starvation / glucose starvation — the caloric restriction mimetic signal
- Lipid droplet organization — AMPK regulates lipolysis through ACC and HMGCR phosphorylation
mTOR — where rapamycin binds
Protein: Serine/threonine kinase mTOR, 2,549 aa (UniProt: P42345) Drug: Rapamycin
Here we hit a limitation. At 2,549 amino acids (above the ~2,000 training cutoff), the model produced only a domain map without a full reasoning trace. But even that was valuable:
- FKBP12-rapamycin binding domain at positions 2015–2113 — this is exactly where rapamycin (complexed with FKBP12) physically docks and inhibits mTORC1
- Catalytic PI3K-like domain at positions 2153–2484 — just 40 residues downstream
- HEAT/Armadillo repeats spanning the N-terminal half — the scaffold that mediates interactions with RAPTOR (mTORC1) and RICTOR (mTORC2)
The convergence: one system, four entry points
When you lay out results from all five runs, a single architecture emerges:
NAMPT → NMN → NAD+ → SIRT1 ──→ TSC2↑ ──→ mTOR↓
│
AMPK (α2+β2) ← metformin ──→ TSC2↑ ──→ mTOR↓
│
Rapamycin+FKBP12 ──→ [domain 2015-2113] ──→ mTOR↓
│
S6K1↓, 4E-BP1↓
│
protein synthesis↓, autophagy↑
Three different inputs — NMN/resveratrol, metformin, rapamycin — converge on the same outputs: mTOR↓ and autophagy↑. The model confirmed this independently for each protein, knowing nothing about my stack. It analyzed raw sequences and arrived at the same conclusions that biologists assembled over decades of experiments.
The whole experiment took about 20 minutes on a free web tool. A few years ago, this level of analysis would have required a bioinformatics team.
Links:
- Web interface: app.bioreason.net
- Paper: bioRxiv 2026.03.19
- Code & weights: github.com/bowang-lab/BioReason-Pro
- How to run it: paste an amino acid sequence from UniProt + organism name → select RL Model → submit
Has anyone else tried running their supplement targets through BioReason-Pro? Curious what others find.