Models for editing the genome.
XS1 Biosciences researches AI for CRISPR genome editing: models that help researchers choose guides, anticipate their off-target risk and repair outcomes, and study AI-designed editors — each prediction a hypothesis until it is checked against experimental data.
Targeting a CRISPR edit, then reading it out
Scanning: the complex checks the DNA for a protospacer beside a PAM.
Illustrative schematic of CRISPR targeting. An RNA-guided complex scans DNA and docks where an NGG PAM sits beside a 20-nucleotide protospacer; the strands open into an R-loop as the guide’s spacer base-pairs with the target strand, and a blunt double-strand break is made about three base pairs from the PAM. Two data overlays follow: a distribution of predicted editing outcomes and a ranked list of candidate off-target sites with specificity scores. All values are generated for illustration, not model output.
Predicted editing outcomes
IllustrativeOff-target candidates
Illustrative01What XS1 Biosciences builds
XS1 builds the models that turn a target into a tested edit.
These are XS1 Biosciences' own research programs for genome editing. Each predicts something a researcher can check against the lab, and guide RNA design is where it starts.
Guide RNA design
Generating and ranking candidate guides for a target: sgRNAs for Cas9, crRNAs for Cas12a, and pegRNAs for prime editing, with their primer binding site and RT-template choices.
- PAM-aware across nucleases — NGG, TTTV, relaxed variants
- Scored for predicted on-target activity and specificity
Off-target prediction
Searching the genome for near-match sites — allowing mismatches and bulges — and scoring the risk at each, aggregated into a single specificity score for the guide.
- Genome-wide candidate search
- One calibrated specificity score per guide
Editing-outcome prediction
Predicting what the cell ends up with: repair-product (indel) distributions after a cut, base-editor bystander edits inside the editing window, and prime-editing efficiency and purity.
- Indel distributions and bystander edits
- Prime-edit efficiency and product purity
Editor design
Generative protein models that propose candidate Cas variants and compact editors, put forward as hypotheses for laboratory validation rather than asserted as results.
- Candidate nucleases and compact editors
- Proposed for the lab, not a finished claim
Screen design and analysis
Designing pooled guide libraries — guides per gene with matched controls — and analyzing the readout, from hit calling to single-cell Perturb-seq.
- Pooled libraries with controls
- Hit calling and Perturb-seq analysis
Research agents tie it together
Research agents chain these models end to end — design, predict, prioritize — and every prediction is checked against experimental data before it informs an experiment.
- Models composed into one workflow
- Checked against measurement, never a standalone verdict
02Guide design, shown
Pick a nuclease; watch which sites qualify.
Guide RNA design is XS1's core contribution to editing. Choose SpCas9, Cas12a or prime editing and the PAM rule changes, so a different set of protospacers qualifies — each scored for activity and specificity. The PAM rules and cut positions are accurate; the sequence is fictional and the scores are illustrative.
Designing a guide for a target
SpCas9: an NGG PAM sits 3' of a 20-nt protospacer; the blunt cut falls ~3 bp inside it.
Candidate guides · 3
On-target and specificity, 0–1
Switching nuclease changes which sites qualify, because the PAM rule changes. Scores here are illustrative; XS1 models rank real candidates and are checked against experimental data.
03How CRISPR works, in brief
A programmable nuclease, guided by RNA to a chosen site.
Cas9 or Cas12a is directed to a genomic site by a short guide RNA whose spacer base-pairs with the DNA target — but only where a PAM motif sits beside it. Wild-type Cas9 makes a double-strand break; the cell repairs it by error-prone end joining, which knocks genes out, or by homology-directed repair against a template, which writes precise changes. Base and prime editors reach smaller, defined edits without a full break, and dCas9 can repress or activate genes without cutting at all.
Cas9 / Cas12a
RNA-guided nucleases that cut DNA at a chosen site. Cas9 reads an NGG PAM; Cas12a reads a T-rich PAM and leaves staggered ends.
Guide RNA · spacer
The RNA that programs the nuclease. Its ~20-nt spacer base-pairs with the genomic protospacer and sets specificity.
PAM
A short motif (NGG for SpCas9) next to the protospacer that the enzyme must recognize to cut. It is absent from the guide, so it prevents self-targeting.
Double-strand break · NHEJ / HDR
Wild-type Cas9 makes a double-strand break. End joining (NHEJ) leaves small indels that knock a gene out; homology-directed repair (HDR) writes precise changes from a template.
Base & prime editing
Editors that avoid a full break. Base editors convert C→T or A→G with a nickase and deaminase; prime editors write small defined edits with a nickase, reverse transcriptase and a pegRNA.
CRISPRi / CRISPRa
Catalytically dead Cas9 (dCas9) fused to a repressor or activator to silence or boost a gene without cutting DNA.
Pooled screens · Perturb-seq
A library of guides perturbs many genes across many cells at once. Read out by single-cell RNA-seq, each cell links its perturbation to its transcriptome.
On-/off-target · outcome
Whether a guide cuts efficiently at its site, whether it also cuts lookalike sites, and which repair products a cut produces. Each is a sequence-to-outcome prediction.
04The modeling problems
Three questions decide whether an edit works.
Each is a sequence-to-outcome prediction with large experimental datasets behind it. XS1 Biosciences frames CRISPR as exactly this kind of research: turning a target locus and a candidate guide into calibrated predictions a researcher can act on, and comparing those predictions against measured data.
On-target activity
Will a chosen guide cut efficiently at its intended site? A score that ranks candidate guides for a locus.
Off-target risk
Will the same guide also cut sites that differ by a few mismatches? A specificity signal across the genome.
Editing outcome
What repair products will the cut most likely produce? For base and prime edits, the efficiency and purity of the intended change.
05Built on published work
A decade of methods to build on.
The field has strong published baselines. XS1 Biosciences treats them as the starting point and attributes them to their source. Capabilities such as AI-designed editors are the work of the groups named here, not XS1 results.
- On-target activityRule Set 2 / Azimuth and Rule Set 3 (Broad Institute)
- Off-target & unified designDeepCRISPR (Genome Biology, 2018) and Elevation (Microsoft Research)
- Editing outcomesinDelphi (Nature, 2018), with related models such as FORECasT and Lindel
- Prime editingPRIDICT and ePRIDICT (University of Zurich)
- AI-designed editorsOpenCRISPR-1 (Profluent) and Evo (Arc Institute)
- Screens at scaleGenome-scale Perturb-seq (Cell, 2022)
06Validation and limits
Every prediction is a hypothesis until it is measured.
We describe CRISPR models alongside the experimental data they are checked against — deep screens of guide activity, off-target assays and repair-outcome libraries — and report where a model is calibrated and where it is not. Predictions support research decisions about which guides and experiments to run first.
Checked against measurement
Guide-activity screens, off-target assays and repair-outcome libraries.
A signal for follow-up
Scores rank which guides and edits to test first, never a guarantee an edit will or will not occur.
Research use only
Cell- and sequence-level research. Not clinical, diagnostic or therapeutic, and no decisions about individuals.
Work with XS1 Biosciences
Bring XS1 Biosciences a genome-editing question.
A locus to design guides for, a screen to analyze, an editing outcome to predict: describe it in general terms and we will follow up.
Please do not send unpublished data, sequences or other confidential material through this website.