aikixp-client 1.0.0

Official Python client for the Aiki-XP bacterial protein-expression prediction API (Aikium, Inc.).

aikixp-client

Official Python client for the Aiki-XP bacterial protein-expression prediction API.

Aiki-XP is a leakage-controlled multimodal model trained on 492,026 genes from 385 bacterial species; it predicts within-species relative protein expression (per-species z-score). See the paper and main repository for background.

Install

pip install aikixp-client

Python ≥ 3.9. The only dependency is requests.

Quick start

from aikixp_client import Client

client = Client()

# --- Tier A: protein only, no host required (fastest) --------------
r = client.predict_tier_a("MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGGPGLGLNHVGQIIR")
print(r["predictions"][0]["predicted_expression"])  # e.g. -0.083

# --- Tier D: full 9-modality, A100 (GPU) --------------------------
r = client.predict_tier_d(
    protein="MKT...",
    cds="ATG...TAA",
    host="NC_000913.3",
    mode="native",
)
print(r["predictions"][0]["predicted_expression"])
print(r["modalities_filled"])  # ['evo2_7b_full_operon_pca4096', 'hyenadna_dna_cds', ...]

# --- Look up a gene in the 244K held-out CV corpus ----------------
g = client.lookup_gene("Escherichia_coli_K12|NP_417556.2")
row = g["results"][0]
print(row["tier_d"], row["true_expression"], row["cv_fold"])

# --- Reproduce the paper's numbers on a fresh random sample -------
sample = client.sample_lookup(n=5000, seed=42)

# --- Compare all 5 tiers on the same sequence, in parallel --------
results = client.compare_all_tiers(
    protein="MKT...",
    cds="ATG...TAA",
    host="NC_000913.3",
    mode="native",
)
for tier in ("A", "B", "B+", "C", "D"):
    print(tier, results[tier]["predictions"][0]["predicted_expression"])

Auto-fill a CDS from a protein + host

r = client.cds_for_protein(
    protein="MAKPIL...",
    host="NC_000913.3",
)
print(r["cds"], r["source"])  # "ATG...", "native" or "codon_optimized"

Host catalog

hosts = client.hosts()            # 1,831 bacterial reference genomes
print(f"{len(hosts)} hosts cached")
e_coli = [h for h in hosts if h["acc"] == "NC_000913.3"][0]
print(e_coli["name"], e_coli["n_test"])

Error handling

All HTTP errors raise :class:AikixpError with the status code and response body.

from aikixp_client import Client, AikixpError

try:
    client.predict_tier_d(...)
except AikixpError as e:
    if e.status in (502, 504):
        # Cold-start timeout; retry after a short delay
        ...
    else:
        raise

Tiers at a glance

Tier	Modalities	Hardware	Cold	Warm	ρ_non-conserved
A	ESM-C + ProtT5 (protein only)	CPU	~33 s	~15 s	0.518
B	+ HyenaDNA + classical	A100-40 GB	~44 s	~6 s	0.531
B+	+ Evo-2 7B init-window + RNA init	A100-40 GB	~79 s	~10 s	0.543
C	+ Evo-2 7B full operon	A100-40 GB	~8 s	~7 s	0.575
D	+ Bacformer-large + operon structure	A100-80 GB	~6 s	~5 s	0.590

Gotchas

• The server scales to zero after ~5 minutes of idle — first calls after a quiet period cold-start a fresh GPU container. Default timeouts account for this (120 s for CPU endpoints, 1200 s for GPU endpoints).
• Tagged sequences (His6, HiBit, FLAG, etc.) are out-of-distribution. Strip tags first using aikixp.sequence_normalization from the main repo.
• Native mode requires the user-supplied CDS to exact-substring-match the host genome; if it doesn't, the server silently falls back to heterologous mode with the lacZ anchor. The response's predictions[0]["operon_source"] field tells you which path was taken.

License

Code: Apache 2.0. Aiki-XP model weights + data: CC-BY 4.0 (via Zenodo).

Citation

@article{tien2026aikixp,
  title = {Aiki-XP: leakage-controlled multimodal prediction of within-species
           relative protein expression at pan-bacterial scale},
  author = {Tien, Hudson and Sharma Meda, Radheesh and Shastry, Shankar and
            Mysore, Venkatesh},
  journal = {Nature Biotechnology},
  year = {2026},
  note = {Under review. Preprint on bioRxiv.}
}