Skip to content

CisFalcon

AI now designs synthetic enhancers meant to switch a gene on in one cell type and stay silent everywhere else.

Code Video Cerebral Valley

Core trick

Scores designed enhancer sequences with a frozen measured-activity model, then applies motif edits and re-scores to flag off-target cell-type activity.

Tags

sciencehealth

Description

AI now designs synthetic enhancers meant to switch a gene on in one cell type and stay silent everywhere else. When one fires in the wrong cell, the consequence can be clinical: in a real enhancer-AAV gene therapy for Dravet syndrome, driving the gene in all neurons instead of only interneurons raised mortality in mice, while the cell-type-specific version was safe and corrected seizures (Mich et al., Sci Transl Med 2025). CisFalcon catches that failure from sequence, before a lab spends six weeks and a synthesis run. Paste a designed enhancer into the live tool (https://cisfalcon-lifesci.fly.dev/) and it scores cell-type specificity against a frozen external measured-activity model, then applies the prescribed motif edit and re-scores from failing to passing (a proposed fix, not wet-lab validated). On 93,435 independent designs from a different lab it separates failures at AUROC 0.80; ranking a batch safest-first cuts wasted syntheses about 70%. We also tested the harder within-tissue question directly, on real in-vivo cortical enhancers with an independent model (AUROC 0.71). It flags risk; it does not design DNA.
belumume/cisfalcon★ 20 forksPythonMITCreated Jul 10, 2026 · last push Aug 12, 2026

README excerpt

# CisFalcon

**Live tool: https://cisfalcon-lifesci.fly.dev/**

![The CisFalcon Overview tab: catch the AI-designed enhancers that will fail before the lab synthesizes them. A frozen measured-activity model gives cross-lab AUROC 0.80 against 0.50 chance on 93,435 held-out designs. Ranking safest-first and synthesizing the safer half cuts failures 41% conditioned within one target cell and one generator; pooled across a mixed batch it reads 70% (6.29% to 1.91%), but a sequence-free rule using only the stratum base rates scores about 90% pooled, so the pooled number is not evidence of per-design skill.](docs/hero_overview.png)

*A real Gosai/Tewhey design that CisFalcon flags as off-target, diagnoses, and closes the loop on: disrupt the K562 driver motifs (GATA1, TAL1, GATA2), install the HepG2 grammar (HNF4A, HNF1A, CEBPA), and the same external model moves the predicted specificity gap from failing to passing. This is an in-silico consistency check on one design, not wet-lab validation, captured live from the tool.*

## In 60 seconds

**The problem.** AI now designs synthetic enhancers: short DNA meant to switch a gene on in one cell type and stay silent everywhere else, the targeting step of gene therapy. In the one published library where this has been measured end to end (93,435 Gosai/Tewhey designs, scored cross-lineage in vitro), about **1 in 16** of those designs is secretly broken and fires in the wrong cell; in-vivo cortical screens report far higher failure rates sti

Team