Exploring Representation-Aware Model Selection for Out-of-Distribution Gene Perturbation Prediction

Prediction of transcriptional responses to genetic perturbations is difficult when the target perturbation is absent from model training. A further challenge is that different predictors can behave differently on different unseen perturbations. This study therefore examines out-ofdistribution (OOD) perturbation prediction as a model-selection problem rather than assuming that one predictor is uniformly optimal. Using an Adamson Perturb-seq dataset containing 65,899 cells, 5,060 genes, and 82 perturbation conditions, we analyze 21 perturbations held out from training. Four candidate predictor configurations are evaluated: GO-only MLP, GenePT-only MLP, GenePT + GO concatenation MLP, and Gated Attention Fusion. For each perturbation, the observed candidate-model performance is used to quantify the opportunity for perturbationspecific selection. The best fixed strategy has mean R 2 =0.399650, while the oracle upper bound is 0.466234, giving an available selection gap of 0.066584. Importantly, the perturbation-level retrospective adaptive values reported in the study average to R 2 =0.421228, corresponding to a 0.021578 absolute and 5.40% observed relative improvement over the fixed strategy. Embedding Coverage is also examined as a compact representation-availability descriptor. In this exploratory analysis, it shows negative Spearman associations with OOD difficulty ( ρ =−0.5472, p = 0.0103) and with the observed adaptive improvement ( ρ = −0.4741, p = 0.0299), with both p-values treated as nominal exploratory results, whereas the paired Wilcoxon comparison between adaptive and fixed outcomes is not significant ( p = 0.0638). These results provide exploratory evidence that representation coverage can be useful for characterizing OOD behavior, but they do not establish universal superiority of adaptive selection. The study is deliberately positioned as a focused single-dataset analysis and identifies the validation required for broader claims.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11768707/v1
Primary Topic
Gene Regulatory Network Analysis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Exploring Representation-Aware Model Selection for Out-of-Distribution Gene Perturbation Prediction

Aman Tarikere
1 links
Research Square
Gene Regulatory Network Analysis
preprint

Exploring Representation-Aware Model Selection for Out-of-Distribution Gene Perturbation Prediction

Aman Tarikere
preprint en

Abstract

Prediction of transcriptional responses to genetic perturbations is difficult when the target perturbation is absent from model training. A further challenge is that different predictors can behave differently on different unseen perturbations. This study therefore examines out-ofdistribution (OOD) perturbation prediction as a model-selection problem rather than assuming that one predictor is uniformly optimal. Using an Adamson Perturb-seq dataset containing 65,899 cells, 5,060 genes, and 82 perturbation conditions, we analyze 21 perturbations held out from training. Four candidate predictor configurations are evaluated: GO-only MLP, GenePT-only MLP, GenePT + GO concatenation MLP, and Gated Attention Fusion. For each perturbation, the observed candidate-model performance is used to quantify the opportunity for perturbationspecific selection. The best fixed strategy has mean R 2 =0.399650, while the oracle upper bound is 0.466234, giving an available selection gap of 0.066584. Importantly, the perturbation-level retrospective adaptive values reported in the study average to R 2 =0.421228, corresponding to a 0.021578 absolute and 5.40% observed relative improvement over the fixed strategy. Embedding Coverage is also examined as a compact representation-availability descriptor. In this exploratory analysis, it shows negative Spearman associations with OOD difficulty ( ρ =−0.5472, p = 0.0103) and with the observed adaptive improvement ( ρ = −0.4741, p = 0.0299), with both p-values treated as nominal exploratory results, whereas the paired Wilcoxon comparison between adaptive and fixed outcomes is not significant ( p = 0.0638). These results provide exploratory evidence that representation coverage can be useful for characterizing OOD behavior, but they do not establish universal superiority of adaptive selection. The study is deliberately positioned as a focused single-dataset analysis and identifies the validation required for broader claims.

Research Square
University of Mysore (IN)
Gene Regulatory Network Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon