Lab Newsletter — August 24, 2026: Designing the Edit

CRISPR made the cut easy. What stayed hard was everything around it — the two questions that decide whether an edit is a tool or a gamble. Will this guide actually cut? And after the enzyme snips the DNA and the cell scrambles to repair the break, what sequence will you be left with? For years the second question was treated as unanswerable: double-strand-break repair was assumed to be stochastic, a mess of random insertions and deletions you could only sequence after the fact. Both questions turned out to be prediction problems — and machine learning, given enough measured edits to learn from, answered them.

🎯 Will the edit work?

The first question is guide activity, and it fell to data. Early on, Doench and colleagues (Nature Biotechnology, 2016; senior author David Root — with Microsoft Research co-authors already in the byline) profiled “the off-target activity of thousands of sgRNAs” and built design rules “to maximize on-target activity and minimize off-target effects.” Deep learning then sharpened it. DeepSpCas9 (Kim et al., Science Advances, 2019; senior author Hyongbum Henry Kim) trained on SpCas9-induced indel frequencies at “12,832 target sequences” and “showed high generalization performance” on datasets it had never seen — the trait that makes a predictor actually usable. Its sibling DeepCpf1 (Kim et al., Nature Biotechnology, 2018) learned from “15,000 target sequences” with a convolutional neural network, and — tellingly — got better once it “incorporated chromatin accessibility information”: the genome’s packaging, not just its letters, shapes where an enzyme can work.

🧬 What will the edit produce?

Then the harder, more surprising result: the aftermath of the cut is predictable. inDelphi (Shen et al., Nature, 2018; senior author Richard Sherwood) trained on a library of “2,000 Cas9 guide RNAs” and showed template-free editing “is predictable and capable of precise repair to a predicted genotype,” forecasting deletions and single-base insertions “with high accuracy (r = 0.87)” across five cell lines. Its headline is startling: “5–11% of Cas9 guide RNAs” are precise-50 — a single repair genotype makes up more than half of all products — and the team corrected “195 human disease-relevant alleles,” including patient cells for Hermansky–Pudlak syndrome and Menkes disease, using nothing but the cell’s own repair. FORECasT (Allen et al., Nature Biotechnology, 2019; senior author Leopold Parts) nailed the point at scale — “>10⁹ mutational outcomes” from “>40,000 guide RNAs” — confirming outcomes “are not random, but depend on DNA sequence.” The idea then jumped to the newer editors. BE-Hive (Arbab et al., Cell, 2020; senior author David Liu) predicts base-editing “genotypic outcomes (R ≈ 0.9) and efficiency (R ≈ 0.7)” and corrected “3,388 disease-associated SNVs with ≥90% precision.” And PRIDICT (Mathis et al., Nature Biotechnology, 2023; senior author Gerald Schwank), an “attention-based bidirectional recurrent neural network” trained on “92,423 pegRNAs,” predicts prime-editing efficiency (Spearman R = 0.85) and tells you which guides to bother with — high-scoring pegRNAs edited “12-fold” better in vitro and “tenfold” better in the liver in vivo.

🧭 The honest frontier — and why it’s our kind of problem

Two caveats keep this grounded, and both are the lab’s native tongue. First, the predictions are still narrow and context-bound. A recent review is candid that among the many CRISPR-design tools, “assessment of their application scenarios and performance benchmarks are limited,” and the newest deep-learning predictors “have not been systematically evaluated” (Konstantakos et al., Nucleic Acids Research, 2022). And the outcomes themselves shift with the cell: FORECasT found each guide has an “individual cell-line-dependent bias,” so a model trained in one line is a hypothesis, not a guarantee, in another. Second, a predicted genotype is not a measured one. Every anchor here earns its keep by validating — inDelphi’s 195 alleles, BE-Hive’s 3,388 SNVs — the same prove-it discipline this digest keeps returning to: the model proposes, sequencing disposes.

Here’s why it lands for us. Reading a genome to predict what a natural mutation does was last week’s story; this is the inverse — predicting how to write the correction. It is the design half of a perturbation: a virtual cell or human-cell simulator tries to predict a cell’s response to a change, and these models design the change itself. Put them in a loop and you have the design brain of a self-driving lab — propose an edit, make it, measure it, learn, repeat — the exact engine our autonomous research and imaging-farm work is built around. And these predictors are precisely the kind of model that should live open and callable on shared infrastructure — the same BioImage Model Zoo and ImJoy ethos we build for imaging — so an edit-design claim can be reproduced and checked, not just trusted. Predict the change, keep the machine honest against the sequencer, and close the loop — that’s the whole job, written one base pair at a time.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.