Kairos

The AI-scientist agent on the Evolrix AI biomanufacturing platform. Live

September 2, 2026 · v0.19.6

Kairos is the AI-scientist agent on the Evolrix AI biomanufacturing platform. Give it a protein; it returns a panel of synthesis-ready coding-sequence candidates — each with a transparent scorecard and risk flags, not one opaque "optimized" sequence. Every candidate provably encodes the exact protein, and every score is a deterministic computation you can re-run over the printed sequence.

Scope — Dry-lab design ranking: synthesis safety, host resemblance, 5′ translation initiation, and protein developability. Not an expression or titer prediction. Real expression awaits wet-lab data.

How it scores

Four explainable axes — never one fragile score

Each candidate is scored on four independent axes, every one a deterministic, re-runnable computation with its provenance shown.

01

Manufacturability

GC window, homopolymers, tandem repeats, restriction sites, forbidden motifs.

02

Host-likeness

A genome-derived K. phaffii codon model (codon usage + %MinMax rhythm + codon-pair bias) that deliberately rejects naive max-CAI over-optimization.

03

5′ translation initiation

Start-proximal mRNA accessibility — an open start means better initiation.

04

Protein developability

Length, MW, pI, cysteines, N-glycosylation motifs, hydrophobicity.

EvoCodon

A self-developed, genome-learned codon model

Candidates are proposed by transparent host heuristics and EvoCodon — a self-developed, genome-learned codon model trained on public K. phaffii genomes that reproduces host-specific codon rhythm beyond frequency-based methods. The learned candidate competes on the same scorecard as every other — it does not auto-win.

Secretion design

For a secreted protein, Kairos selects a source-traceable secretion leader — e.g. the S. cerevisiae α-mating-factor prepro (UniProt P01149) or the SUC2 invertase signal (UniProt P00724), each carrying its UniProt record and supporting literature — fuses it N-terminally, and designs the coding sequence of the whole construct. Correct by construction: the design encodes exactly signal + target. Real secretion efficiency is a wet-lab readout (see scope below).

Beyond the construct

Manufacturability, and hypotheses that get verified

Since v0.17.0 every dossier carries a downstream-process panel: fermentation, DSP sequence, step-wise recovery (multiplicative), and directional cost vs the benchmark, in four parity tiers — every number cited. Absolute titer, recovery, scale-up, and cleaning stay WARN: wet-lab required. Design is now two-dimensional — candidates rank on host-fidelity × total recovery, so a high-fidelity intracellular design can lose to a slightly lower-fidelity secreted one when the latter is far cheaper to purify.

Since v0.19.0 Kairos works hypothesis-first: it proposes explicit hypotheses about your question and verifies each against the retrieved evidence. Every dossier carries hypothesis-verification cards — supported, contested, unverified, insufficient evidence — with a hard sparse-evidence rule: fewer than two sources means honestly “insufficient”.

Featured run

A Kairos design run on Cytochrome c

UniProt Q6Q4H8 — a native Komagataella phaffii protein. Twelve reproducible candidates, ranked across all four axes, with the recommended one in the top Pareto tier.

Pichia CDS design report design_50a60d8960 komagataella_phaffii 110 aa · 12 candidates
Dry-lab design ranking — synthesis safety · host resemblance · 5′ initiation · protein developability. Not an expression or titer prediction.

Cited evidence retrieved live from the Kairos knowledge layer — every entry a real, openable source.

PMID:30587177

AOX1 promoter regulation

Methanol-regulated PAOX1 (Mxr1 positive-feedback circuit) — grounds the expression strategy.

PMID:37888283

Nitrogen source & MUT pathway

Ammonium sulfate raises MUT / PEX gene expression — tune nitrogen to up-regulate the pathway.

PMID:38650288

Fed-batch strategy

Glycerol for biomass, then methanol induction — the standard Pichia fermentation rhythm.

PMID:31590267

Proteolysis risk

Proteolytic degradation of recombinant protein (HSA) in P. pastoris — watch degradation; add inhibitors if needed.

Recommended

method: uniform top Pareto tier flag: extreme_pi
Candidate panel — top of the Pareto ranking
RankMethodPareto tierFlags
1uniformtop1
2host_weightedtop1
3gc_balancedtop1
4gc_balancedtop2
5host_weightedtop3
Sequence validity strong · clean translates to the exact target protein; in-frame; no premature stop
Manufacturability strong · clean GC window, homopolymers, tandem repeats, restriction sites, forbidden motifs
Host-likeness moderate · clean genome-derived codon usage + %MinMax rhythm + codon-pair bias (rejects max-CAI degenerates)
Protein developability strong · 1 flag length, MW, pI, cysteines, N-glycosylation sequons, hydrophobicity
Every candidate is synonymous — encodes the exact input protein (verified by translation). Ranking is multi-objective (Pareto) with per-criterion explanations: multiple candidates, never one opaque "best". The extreme_pi flag on the winner is an intrinsic property of the target protein, not a design defect. Sequence withheld in the public report; available to invited collaborators. Engine: Kairos v0.1.0.

Review layer

Every claim, audited — before you read it

Since v0.5.0, every Kairos dossier passes through a deterministic review layer before it reaches you. The reviewer checks three things mechanically: every cited PMID or UniProt ID is genuinely present in the retrieved evidence (no fabricated citations); every number quoted in the reasoning matches the design panel’s output (no number drift); and every conclusion carries a backing source (no unsupported assertions). Since v0.12.0 it adds one more deterministic check: whether the cited source actually supports the claim — same organism, same target protein, same expression-system aspect, and whether the wording is entailed, partially supported, or contradicted. Support problems are always WARN; FAIL stays reserved for the one unambiguous case — a citation that does not exist. Each claim receives a verdict — PASS, WARN, or FAIL — with its evidence trace, and the audit rides along as one panel of the Kairos workbench: evidence, design, structure, manufacturability, audit.

The reviewer was stress-tested adversarially before launch — a fabricated PMID:99999999 was flagged, an inflated fidelity score (0.95 vs the panel’s 0.82) was flagged FAIL, and an ungrounded “will definitely express at 5 g/L” claim was flagged as unsupported. On honest model output it does not false-positive.

Honesty boundary: the review layer verifies traceability — citations present, numbers consistent, claims grounded. It does not judge scientific correctness, and it never predicts expression or titer. How the review layer works →

FAQ

Can it design secreted proteins?

Yes. Kairos selects a source-traceable secretion signal peptide, fuses it, and designs the full construct's CDS. It chooses the leader and guarantees a valid construct; real secretion efficiency is wet-lab, not predicted.

References

Sources behind the design

The host heuristics, secretion leaders, and host-likeness model draw on public, traceable literature and reference databases. Each is a real, citable source.

  1. UniProt P01149 — S. cerevisiae α-mating-factor prepro. UniProt:P01149
  2. UniProt P00724 — SUC2 invertase signal. UniProt:P00724
  3. Merkaš M, et al. The MFα signal sequence in yeast-based protein secretion: challenges and innovations. Appl Microbiol Biotechnol, 2025. PMID:40471355
  4. Aw R, et al. A systematic analysis of the expression of the anti-HIV VRC01 antibody in Pichia pastoris through signal peptide optimization. Protein Expr Purif, 2018. PMID:29601964
  5. Ahmad M, et al. Chang CH, et al. Enhancing the efficiency of the Pichia pastoris AOX1 promoter via the synthetic positive feedback circuit of transcription factor Mxr1. BMC Biotechnol, 2018. PMID:30587177
  6. Zha J, et al. Advances in Metabolic Engineering of Pichia pastoris Strains as Powerful Cell Factories. J Fungi (Basel), 2023. PMID:37888283
  7. Maity N, et al. Statistically Designed Medium Reveals Interactions between Metabolism and Genetic Information Processing for Production of Stable Human Serum Albumin in Pichia pastoris. Biomolecules, 2019. PMID:31590267

Ask Kairos a question

Kairos is in research preview, Pichia-scoped. We are onboarding industrial R&D teams who need transparent, re-runnable design ranking — not guesses.

Enter ↵ to ask Kairos · Shift+Enter for newline

✓ Opening Kairos with your question…

Or request access by email