bioX Home EH14 X BGC-1 EH18 Designer API News

Models

Scores and specifications. Every figure is read from the measurement reports when this page is served.

EH14 X

parameters
257,298,562
window
512bases per read
vocabulary
6one token per base
bases read
4500 Mseawater and sediment
compute
—
TaskOursSimplest method that works
Real sequence or PCR artefact 0.9735UCHIME 0.9442
Marine barcode novelty 0.79694-mer counting 0.8227

All 27 measurements for EH14 X →16 ahead, 11 behind

BGC-1internal · no API · weights not distributed

parameters
321,017,986bgc-330M
window
2048a cluster is ~20,000 bases
tokens
5.52 G8 cards, 0.69 G each
GPU hours
180.8rented RTX 4090s
final loss
1.0876perplexity 2.967
families held out
1,534split by family
Compound class from a clusterCorrect
Guessing the largest class0.2281
Nearest known neighbour0.4625
EH14 X, 512-base window0.5031
BGC-1, same 512-base window0.6937
BGC-1, its own 2,048-base window0.7406
Continuing a clusterDistance to real cluster DNA
BGC-11.6283
Order-3 Markov chain, same prompt4.94

All 3 measurements for BGC-1 →

EH18 Designerinternal · no API · weights not distributed

parameters
85.2 Meh18-85M-A_oud.pt
gate
67.7%unselected fresh draws
floor
62.6%paste-together, real leader and real core
ceiling
82.7%real held out precursors
NeuRiPP
74.2%floor 71.4%
bar set in advance
65.2%to claim it beats paste-together
tested in a lab
0of 978 candidates
Written precursor against the enzyme requirementsFull gateNeuRiPP only
Random, matched for composition20.2%22.4%
Paste-together, leader i with core j — the floor62.6%71.4%
EH18 Designer, unselected fresh draws67.7%74.2%
Real held-out precursors — the ceiling82.7%94.7%

The floor is paste-together, not random: it has a real leader and a real core, and only the pairing is broken. That is the honest thing to beat. On NeuRiPP the model is above it and that figure does not depend on length. On the full gate it is undecided: the bar set in advance is 65.2% and the two length regimes differ, design median 72 residues against 63. An earlier version of this page compared against the random floor and so showed the model four times above it.

Novelty, alignment against all 3,008 training precursors, on all 600 designsIdentity
median43.0%
highest55.4%
above the 80% line0
SizeGate, five fresh draws
85 M parameters55.2%
393 M, best step54.5%
393 M, last step50.0%

These three figures come from an older instrument: five fresh draws each, not the 990 held-out pairs used above. They are comparable to each other and not to the 67.7% at the top of this section. The reading is that 393 M is not better than 85 M, 0.7 points apart, which is a draw.

Cyanobactin coresMeet the requirements
EH18 Designer0.25%
Random from the same twelve residues — floor0.45%
CandidatesCount
written by EH18 Designer750
enumerated or found without a model5,955
total6,705
tested in a lab0

All measurements for EH18 Designer →

Access

API key
self-service
weights, on-premise, higher limit
by application
every answer carries
release and weights hash