CRISPR-TRIAGE: Evaluation of Parameter-Efficient Bi-Modal Off-Target Prioritization Under Budget-Constrained Screening
Year: 2026
Type: software
Abstract: CRISPR-TRIAGE is an open-science computational biology research archive evaluating parameter-efficient off-target prioritization under budget-constrained experimental validation budgets (K <= 50). This package contains the complete trained model checkpoints, multi-seed training pipelines (5 random seeds), biophysical feature extractors, pre-registered baseline models (LightGBM GBDT, Pairwise 1D-CNN, Logistic Regression, crispAI architecture), and evaluation results across two external cellular benchmarks (Listgarten 2018 GUIDE-seq and K562 cellular compilation). Key Scientific Findings & Methodological Disclosure: Algorithmic Parity with Gradient Boosted Trees: On the deduplicated independent cellular benchmark (K562 retrospective multi-assay compilation from DeepCRISPR / Yang et al. 2023; 11 guides, 16,172 unique candidate pairs, 110 confirmed positive guide-target sequence pairs without genomic coordinates), standard gradient boosting (LightGBM GBDT) trained on 15 simple biophysical mismatch features achieves Macro Recall@20 = 0.6350 vs 0.6416 for CRISPR-TRIAGE (margin +0.0066 / +0.66%; paired Wilcoxon W = 18.0, p = 1.0000; 95% bootstrap CI [-0.0979, +0.0900]). LightGBM runs 2.39x faster with 0 neural parameters. While lack of statistically significant superiority does not prove biological equivalence or disprove neural-network utility, the empirical evidence failed to reject the null hypothesis of no difference vs LightGBM under available statistical power. Multi-Seed Volatility in CRISPR Modeling: Retraining across 5 independent random seeds reveals substantial Macro Recall@20 fluctuations (+-6.5% standard deviation on identical data), demonstrating that single-seed reporting in the CRISPR machine-learning literature frequently reflects random seed-selection noise rather than genuine algorithmic superiority. Dataset Deduplication Importance: Forensic auditing documented the impact of 1,613 duplicate candidate rows (and 2 duplicate confirmed positive sites) in the K562 compilation, showing how uncleaned duplicates distort evaluation metrics. Two-Tier Reproducibility Scope: Tier 1 (Saved-Model Inference & K562 Cellular Benchmark Reproduction): 100% self-contained in this archive. All 5 trained seed weights, baselines, scalers, predictions, metric summaries, and the K562 cellular cohort execute fully offline without internet access. Tier 2 (Full Retraining on Primary Assay Datasets): Full retraining from raw sequence reads and complete Listgarten raw dataset reproduction requires separately downloading the 4 un-subsampled raw assay datasets (>20MB each), registered with authoritative accessions and cryptographic SHA256 checksums in DATA_MANIFEST.csv. Licensing, Provenance & Third-Party Rights: CRISPR-TRIAGE code, models, and documentation are licensed under the permissive MIT License. The crispAI architecture code (src/crispai_official/) is under MIT License (Copyright 2023 Furkan Ozden; reference: Özden & Minary, NAR 2024). Pretrained third-party weights for crispAI are omitted from the archive to ensure compliance with conflicting upstream documentation, with archive extraction utilities provided for the authors' official Zenodo records (13335960 / 12609337). Bundled K562 benchmark datasets are distributed under CC BY 4.0 (Yang et al. CSBJ 2023; Chuai et al. Genome Biol 2018). Raw CFD scoring lookup matrices (Doench et al. 2016 Supplementary Table 19) are strictly omitted from the release archive for third-party rights compliance; tested user-side acquisition and verification utilities are provided via scripts/download_cfd_matrices.py. See the root LICENSE file for full details.
Author: Kadayil, Rashid
Institution: Pondicherry University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225628275
