Skip to content

Latest commit

 

History

History
58 lines (45 loc) · 2.98 KB

File metadata and controls

58 lines (45 loc) · 2.98 KB

A reliability map of AI enhancer-design methods

A finding produced with CisFalcon from the committed data, no new model and no new wet-lab.

What is ours and what is not. The specificity-failure labels are Gosai/Tewhey 2024 wet-lab MPRA measurements. The contribution here is two things a scattered dataset does not give you on its own: a systematic per-method and per-target-cell reliability map, and a demonstration that a single frozen cross-lab verifier predicts where those failures cluster (pooled cross-lab AUROC 0.80, orientation-verified). Reproduce with python findings_failure_modes.py; figure in docs/failure_modes.png.

Finding 1: design methods vary about 30x in specificity-failure rate

AI design method specificity-failure rate CisFalcon AUROC
Hamiltonian MC 22.3% 0.922
Simulated Annealing (replicate) 16.9% 0.695
Simulated Annealing 2.5% 0.776
AdaLead 2.3% 0.613
FastSeqProp 0.7% 0.625

Scope: these are the five main generator codes, covering 92,537 of the 93,435 designs. The benchmark carries three more, the low-n unconstrained variants al_uc, fsp_uc and sa_uc (298 / 300 / 300 designs, measured failure 1.01% / 0.00% / 0.00%), excluded here as underpowered. The 30x is the spread over the five main codes (22.28 / 0.68 = 32.7x); across all eight it is unbounded, since two of the small codes record zero failures.

A design's odds of firing in the wrong cell depend heavily on how it was generated, from under 1% to over 22%. A lab picking a generator is also picking a specificity-failure rate, and today that choice is usually made blind.

Finding 2: the disease-relevant targets are the hardest

Target cell specificity-failure rate
SK-N-SH (brain) 10.1%
HepG2 (liver) 8.2%
K562 (blood) 0.6%

Designs for brain and liver cells, the two most gene-therapy-relevant targets here, fail about 14 to 17 times more often than blood-cell designs (10.06 / 0.59 = 17.0x for brain, 8.20 / 0.59 = 13.9x for liver, from data/gosai_designed/designed_scored.csv). The worst single combination is Hamiltonian-MC-designed brain enhancers, where roughly 1 in 3 fails (32.7%). This is exactly where a pre-synthesis verifier earns its cost: the hardest, most clinically important designs are the ones most likely to be quietly broken.

Finding 3: the verifier is strongest where the failures are

CisFalcon's per-method discrimination is highest (AUROC 0.922) on Hamiltonian MC, the method that fails most (22.3%), and it stays useful across every generator. A verifier that were only good on the safe methods would not matter; this one is sharpest on the failure-prone one.

Why it matters

The field has many enhancer generators and, before this, no systematic account of which ones fail, on which targets, and by how much. This map lets a lab anticipate specificity risk from the design method and the target cell before spending a dollar, and it is the reliability layer a frozen-model verifier makes possible.