← all projects

Biotech & Health Research 2025

CRISPR Biobuild Pipeline

CRISPR off-target prediction with deep learning

Closed source

The problem

Predicting where a CRISPR/Cas9 guide RNA will cut off-target is a sequence modeling problem, but a prediction without an explanation isn't useful to a wet-lab scientist deciding whether to trust it.

The approach

Built a benchmarking pipeline that compares transformer-based sequence models (CRISPRformer-inspired) against classical ML baselines (gradient boosting, SVMs) on real and synthetic off-target datasets, paired with an interpretation suite that surfaces attention visualization, SHAP attribution, and sequence-motif analysis for every prediction.

PythonPyTorchTransformersCRISPR/Cas9SHAPBioinformatics

This project benchmarks deep learning against classical ML for CRISPR/Cas9 off-target prediction, and builds the interpretability tooling needed to make either kind of prediction actionable in a wet lab.

the prediction problem

A CRISPR guide RNA can bind and cut at unintended genomic sites that resemble its target sequence closely enough. Predicting these off-target sites accurately - before running an experiment - is a sequence classification problem with real consequences for editing safety.

models compared

  • Transformer sequence models - CRISPRformer-inspired architectures that attend over the guide and target sequence to score mismatch tolerance.
  • Classical ML baselines - gradient boosting and SVMs over hand-engineered sequence and mismatch features, the prior standard.
guide_seq, target_seq = "GACGT...", "GACGA..."
score_transformer = crisprformer.predict(guide_seq, target_seq)
score_baseline = gbm_baseline.predict(featurize(guide_seq, target_seq))

interpretation suite

Every prediction ships with tooling to explain it, not just score it:

  • Attention visualization - which positions in the guide/target pair the transformer weighted most heavily.
  • SHAP attribution - per-feature contribution for both transformer and classical models, on a common footing.
  • Motif analysis - recurring sequence patterns associated with high off-target risk.

benchmarks

Models are evaluated on both real off-target datasets and synthetic benchmarks, the latter useful for probing edge cases (rare mismatch patterns, extreme GC content) that are underrepresented in real data.

status

Research project: the benchmarking pipeline and interpretation suite are functional, with ongoing work comparing architectures and validating interpretability outputs against known off-target biology.

What I took away

  • Transformer sequence models beat classical baselines on off-target prediction, but the gap is largest on sequences with subtle mismatches classical features don't encode well.
  • Interpretability isn't optional in this domain - attention maps and SHAP attribution are what let a biologist decide whether to trust a prediction.
  • Synthetic benchmarks are useful for stress-testing edge cases, but real off-target datasets are what actually calibrate model confidence.

0 comments

0/1000

Loading…