Search bioRxiv⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.11.04.764088

Plate design confounds donor variation and model evaluation in single-cell perturbation screens

Ghajari, G.·Thirunarayan, K.·Amsaad, F.
View original source ↗Export metadata ↓

Abstract

Single-cell screens of immune cells from many donors are used to study why individuals respond differently and to benchmark response models. Each donor-perturbation combination typically occupies one well, controls are grouped, and donors share one layout, so technical effects can look biological. Using a 90-cytokine screen, two compound screens, four cohorts and simulations, we show that well effects explain most replicate-control variation in abundant cell types, controls carry a shared offset, and identical control wells differ by hundreds of differentially expressed genes at a standard threshold, about as many as two receptor-negative proteins do. Negative controls, design-aware estimators and perturbation-specific evaluation change the interpretation of four recent analyses. A deep-learning model that scores below a perturbation-agnostic predictor under standard evaluation outperforms a simple baseline on perturbation-specific differences in two of three cell types (significantly in one with stricter controls), but predicts less well the component shared with inert wells.

Explore related subjects

bioinformaticsPreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ghajari, G., Thirunarayan, K., Amsaad, F.. 2026-11-06. Plate design confounds donor variation and model evaluation in single-cell perturbation screens. https://doi.org/10.64898/2026.11.04.764088

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Contrastive Learning Improves Generalization on M2OR Olfactory Binding Dataset

Olfaction-based biosensors stand to have profound impacts in fields such as medical diagnostics, national security, and drug discovery. A promising mode of vapor phase biodetection mimics volatile organic compound-olfactory receptor (VOC-OR) interactions with target-specific protein or peptide probes; however, lack of experimental coverage remains a core bottleneck. Current computational approaches aim to improve prediction of VOC-OR binding to reduce experimental burden, and current machine learning approaches fail to demonstrate a level of generalization that addresses current engineering challenges. Here, using the Molecular to Olfactory Dataset (M2OR) and Insect Olfactory Binding Protein Database (iOBPdb) datasets, data partitioning methods and their performance are evaluated on a neural network and contrastive learning framework on such VOC-OR binding datasets, varying: model training, architecture, and structure of the input data distribution. It was found that holding out certain chemical or proteomic motifs from the training data diminished model performance with a two layer neural network model. The model was able to generalize to state of the art performance (0.88 AUCROC) on every data split by applying a contrastive learning objective defined by shared molecular constituents and a stochastic sampler. This suggests the contrastive objective provides a strong inductive bias for neural networks to predict novel VOC-OR pairs. These findings highlight that model performance on pairwise biological interaction datasets can be inflated through identity leakage from constituent overlap, and emphasize the need for more robust evaluation protocols. Furthermore, we show contrastive regularization mitigates dataset degeneracy through enforcement of compatibility in the models latent geometry that may extend to other binder target-datasets where identity leakage confounds evaluation.

bioinformatics↗
bioRxiv2026

LinearRaccess: linear-time approximation of RNA accessibility without base-pair span limits

Motivation: RNA accessibility is the probability that every nucleotide in a contiguous interval is simultaneously unpaired; it is widely used to predict binding sites for RNA-binding proteins and other RNA-RNA interactions, often in the form of the opening energy, an equivalent measure on a logarithmic scale. Existing tools restrict the maximum base-pair span, discarding long-range base pairs that matter biologically. Others sample structures, introducing statistical variance, or compute accessibility exactly without a span limit, at cubic time cost. No existing method computes the accessibility of windows longer than one nucleotide deterministically in linear time while preserving long-range base pairs. We present LinearRaccess, which computes accessibility directly in linear time via beam-pruned inside-outside dynamic programming with a difference-array aggregation step, without restricting base-pair span. Results: For a sequence of length N and a fixed beam width b, LinearRaccess runs in O(N) time and O(Nb) memory. Against exact accessibility on 12,872 bpRNA-1m(90) sequences of 100-1,000 nt, its median opening-energy error is below 0.001 kcal/mol (0.14 kcal/mol for 700-1,000 nt), and it distinguishes windows of very small accessibility, including 35-nt windows, below the resolution of sampling-based estimates. LinearRaccess analyzes the full 29,903-nt SARS-CoV-2 genome in 419 seconds. Conclusion: LinearRaccess is the first method to compute the accessibility of windows longer than one nucleotide deterministically (without sampling) in linear time without restricting base-pair span. Availability: https://github.com/TakumiOtagaki/LinearRaccess, archived at Zenodo (https://doi.org/10.5281/zenodo.23598502).

bioinformatics↗
bioRxiv2026

Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA

DNA has two strands, and for most regulatory label prediction tasks a sequence and its reverse complement should get same answer. Yet DNA language models are tokenized, trained and tested on one strand at a time. We show that this makes them inconsistent. BPE and $k$-mer tokenizers split 92--100\% of transcription-factor motifs differently on the two strands, and NT-v2, DNABERT-2 and HyenaDNA change 25-31\% of their predictions when the test sequences are reverse-complemented. We introduce {\model}, a compact DNA language model with a learned tokenizer that guarantees mirror-symmetric token boundaries across reverse-complement sequences, regardless of its learned parameters. We prove that this symmetry holds through the whole model. As a result, the same model can give identical answers on both strands when the label does not depend on the strand, and can still tell the strands apart when it does, which no strand-blind model can learn. Counting only predictions that are correct on both strands, \model\ more than doubles the MCC of NT-v2, DNABERT-2 and HyenaDNA. With 2.3M parameters, it is more accurate than the Caduceus models trained on the same data, and not significantly different from NT-v2 (98M parameters) on Genomic Benchmarks or from DNABERT-2 (117M) on the Nucleotide Transformer tasks. It also stays accurate on inputs sixteen times longer than those it was trained on. The same model trained with reverse-complement augmentation instead still changes about one prediction in ten with the strand, unless its outputs are averaged over both strands. Without any annotation, the tokenizer learns to keep the splice-donor signal \texttt{AGGT} in a single token. Code: https://github.com/NIRJHOR-DATTA/MirrorDNA.

bioinformatics↗