Inspired by the human cognitive strategy of STD used to solve complex problems and its previous successes in complex tasks20, 21, 29, we introduce a conceptually novel supervised learning paradigm (Fig. 1a). This paradigm involves partitioning a complex problem into discrete subtasks based on inherent logical structures, where such logical structures clearly exists in single-cell genetic perturbation prediction. Within this context, the problem can be divided into three subtasks following a inherent logical structure: (1) identifying DEG postperturbations, (2) determining the regulatory directions of these genes, and (3) predicting the magnitudes of gene expression changes (Fig. 1a). Therefore, we propose STAMP, an innovative universal framework designed to predict the outcomes of genetic perturbations, including both single genetic perturbations and multiple genetic perturbations within cell line or across cell lines. Central to the functionality of STAMP is its compatibility as a “plug-in” strategy that can be coupled with either a pretrained gene embedding matrix or a dynamically learnable gene embedding matrix, enabling its adaptability to various scenarios. The outputs of STAMP are the outcomes of the three subtasks (Fig. 1b).

Specifically, STAMP is constructed on the basis of gene embeddings that capture the intrinsic biological nature of genetic interactions (Fig. 1b). STAMP strategically divides this perturbation prediction problem into three subtasks, each of which leverages these gene embeddings to unravel different aspects of genetic perturbations. In the first subtask, STAMP utilizes the gene embeddings of perturbed genes to predict DEG postperturbations. This process effectively constructs a mapping from the gene embedding space to the corresponding DEG postperturbation space (Fig. 1b). Given that the majority of genes remain unaffected by perturbations in most cases, the changes of gene expression vectors postperturbations characteristically exhibit high-dimensional sparsity. Therefore, the DEGs identified in the first subtask can be considered as a perturbation-specific dimension reduction process, helping to improve the signal-to-noise ratio in data. The second subtask further exploits this special property and employs the gene embeddings to predict the regulatory directions of the DEG postperturbations (Fig. 1b). In this subtask, STAMP extends the mapping from the gene embedding space to a new dimension, focusing on the regulatory direction of the postperturbation space. This step is pivotal for delineating the regulatory trajectories of DEGs, whether they are upregulated or downregulated following perturbation. The third subtask involves the use of gene embeddings to predict the magnitudes of gene expression changes (Fig. 1b). This subtask aids in quantifying the extents of expression alterations, providing a quantitative picture of the impacts of genetic perturbations. During the model optimization process, STAMP harnesses the power of multitask supervised learning30 in combination with intermediate supervised learning techniques to jointly solve the three subtasks via gradient descent by considering two key optimization steps: 1) jointly learning the first and second subtasks; and 2) setting the second subtask as the intermediate supervisory signal for the third subtask. To this end, this three-subtask structure not only enhances the prediction accuracy of the model but also establishes a more rigorous and generalizable evaluation system for assessing methodologies within the domain of genetic perturbation prediction. a. Inspired by the STD technique used by humans to solve for complex problems and the pioneering success achieved in tasks such as solving the visual sudoku problem, unraveling

For evaluating the proposed approach in the context of single genetic perturbations, a model was constructed on the basis of gene perturbations that were excluded from the training dataset, ensuring that the predictive capabilities of this model were tested on previously unseen genes. This comparison utilized preprocessed data from four distinct datasets: the RPE1_essential dataset31 with 1694 perturbations, the K562_essential dataset31 with 1741 perturbations, the TFAtlas dataset32 with 1183 perturbations and the K562_GW dataset31 with 7131 perturbations (Supplementary Notes 1-4, Supplementary Table 1). Each model was trained separately on each dataset via fivefold cross-validation (Supplementary Note 6). STAMP was designed to adapt both well-pretrained gene embeddings and learnable gene embeddings. Specifically, we utilized prominent well-pretrained gene embeddings from foundational models such as scBERT, Geneformer and scGPT (these models are denoted as “scBERT+STAMP”, “Geneformer+STAMP”, and “scGPT+STAMP”, respectively) and dynamic learnable gene embeddings in the GEARS with a network structure (this model is denoted as “GEARS+STAMP”) as the inputs of STAMP (Supplementary Note 5). For the GEARS, we extended the model by integrating STAMP into its outcome prediction component. This integration step was designed to leverage the advanced predictive abilities of STAMP within the existing GEARS framework (Supplementary Note 5). Furthermore, to establish a comprehensive and rigorous evaluation, we included random gene embeddings (this model is denoted as “Random+STAMP”), the original GEARS (this model is denoted as “GEARS”), and the CPA as baseline models for comparison purposes. Notably, the CPA was originally designed for predicting the outcomes of compositional perturbations, of which single perturbations were observed during training. We modified this model by substituting its learnable gene embedding component with various well-pretrained gene embeddings (these models denoted are as “scBERT+CPA”, “Geneformer+CPA”, and “scGPT+CPA”; see Supplementary Note 5). This adjustment enabled the CPA to predict outcomes for previously unseen perturbations, thus facilitating fair comparisons.

The ground truths of the DEGs, their regulatory directions and the fold changes observed after perturbations in the three subtasks were identified by a standard statistical analysis33, 34 conducted on each dataset (Fig. 2a, Supplementary Note 4). As the identification of DEG postperturbations in the first subtask and the determination of the regulatory directions of these DEGs in the second subtask are both binary classification tasks, we evaluated the performance of the rested models by calculating the area under the receiver operating characteristic curve (ROC-AUC) and the area under the precision-recall curve (PR-AUC) for both the first subtask and the second subtask. In the third subtask, we measured the Pearson correlation coefficient35 (PCC) and mean squared error (MSE)36. Furthermore, considering that the majority of genes exhibit small variations between their unperturbed and perturbed states, our evaluations in the second and third subtasks were further confined to more challenging settings. Specifically, we focused on the evaluation of the top 20 most differentially upregulated genes and the top 20 most differentially downregulated genes (top 40 DEGs; Supplementary Note 4). Overall, the test results indicate that when coupled with the STAMP strategy, models such as the GEARS or gene embeddings from large foundation models showed remarkable improvements across the three subtasks and testing scenarios involving the top 40 DEGs (Figs. 2b-d, Extended Data Fig. 2, Supplementary Note 7, Supplementary Tables 3-10, Supplementary Figs. 1-3). Additionally, the poor performance of the CPA with pretrained gene embeddings (“scBERT+CPA”, “Geneformer+CPA” and “scGPT+CPA”) indicated that the STAMP strategy is necessary for achieving improved genetic perturbation prediction. Moreover, benchmark tests conducted on different gene embeddings within perturbation prediction tasks revealed that the “GEARS+STAMP” and “scGPT+STAMP” models consistently outperformed other large foundational models across all testing scenarios. In addition, when comparing the time costs of the methods, “scGPT+STAMP” had a faster computation speed than “GEARS+STAMP” (Supplementary Table 2). Collectively, these findings highlight that both well-defined gene embeddings and the implementation of the STAMP strategy are essential prerequisites for accurately predicting single genetic perturbation outcomes. a. The ground truths of three subtasks were identified through a standard single-cell statistical analysis. b. The ROC-AUC and PR-AUC values of different models in subtask-1. c. The ROC-AUC and PR-AUC values of different models in subtask-2. e. The Pearson correlations and MSEs of different models in subtask-3. The data are presented as their mean values, and the error bars show the 95% confidence intervals (CIs) of the mean estimates.

STAMP has a remarkable ability to predict the outcomes of multiple genetic perturbations. This is achieved by transforming the embeddings of multiple genes into inputs for the predictive model. However, it should be noted that in the available datasets for multiple genetic perturbation scenarios, two-gene perturbations are the predominant components. Therefore, to address the challenge of predicting multiple genetic perturbations, the model evaluation was centered around two-gene perturbation scenarios. This prediction scenario could be categorized into three distinct cases, each defined by the extent of the model’s exposure to the genes during training (Fig. 3a). The first case encompassed instances where the model had previously encountered each of the two genes within the training data (zero of two unseen). In the second case, the model was exposed to only one of the two genes in the perturbation pair during the training phase (one of two unseen). The third case was the most challenging scenario, where neither of the two genes in the combination were presented in the training data (two of two unseen). To conduct this comprehensive comparison, we utilized preprocessed data from three distinct datasets: the PRJNA551220 dataset5 with 226 perturbations, the Perturb-CITE-seq dataset37 with 1360 perturbations and the PRJNA787633 dataset38 with 1573 perturbations. Each model was trained separately on each dataset via fivefold cross-validation (Supplementary Note 1, Supplementary Table 1).

For each dataset, the ground truths of the three subtasks were carefully determined by conducting a standard statistical analysis for each specific prediction case (Supplementary Note 4). Subsequently, each model underwent a comprehensive evaluation across the three subtasks for every prediction case. Additionally, a focused evaluation was conducted on the second and third subtasks, particularly for the top 40 DEGs identified in each prediction case. This multilevel assessment strategy ensured a thorough and nuanced evaluation of the predictive capabilities of the models. As a result, consistent with the findings of challenge 1, the models or gene embeddings derived from large foundational models and coupled with the STAMP strategy exhibited substantial improvements across all testing scenarios (Figs. 3b-d, Supplementary information). Among these, the “GEARS+STAMP” and “scGPT+STAMP” models displayed the best performance, yielding better performance than those of the other foundational models (Supplementary Tables 11-28, Supplementary Figs. 4-5). In addition, when comparing the time costs of the models, “scGPT+STAMP” had a faster computational speed than “GEARS+STAMP” (Supplementary Table 2). Collectively, these findings again highlight that coupling a model with the STAMP strategy is an essential prerequisite for accurately predicting multiple genetic perturbation outcomes. a. Illustration of the three testing scenarios for evaluating the two-gene perturbation prediction capabilities of models. X and Y represent different single genetic perturbations.

A significant challenge in the domain of genetic perturbation prediction across cell lines is the variability in the genetic interaction patterns across different cell lines (Fig. 4a). This heterogeneity often results in models trained on one cell line dataset exhibiting poor generalizability to other cell line datasets. Few-shot learning and zero-shot learning represent prominent methods in domain adaptation areas19, 39, 40. However, such interaction pattern diversity in different cell lines complicates the feasibility of performing zero-shot prediction39 across cell lines. Further compounding this issue is the limited availability of diverse cell line datasets. This scarcity hinders the effective application of advanced techniques such as meta-learning40, 41, which are currently infeasible due to dataset constraints. Therefore, the key question here is how quickly a predictive model can adapt to a novel cell line. To address this difficult challenge, we conducted an in-depth evaluation of the ability of STAMP to predict the outcomes of genetic perturbations in a novel cell line dataset for which very few samples (few-shot setting), including zero samples (zero-shot setting), were available for training (Supplementary Note 9). In this evaluation, we utilized the data from 8 distinct datasets and preprocessed data, including 3 CRISPRi datasets and 5 CRISPRa datasets. The 3 CRISPRi datasets consisted of the K562_GW dataset31, RPE1_essential dataset31 and Perturb-CITE-seq dataset37. The 5 CRISPRa datasets consisted of the TFatlas dataset, the PRJNA551220 dataset5, the PRJNA787633 dataset38, the PRJNA641125a dataset42 and the PRJNA609688 dataset43 (Supplementary Note 8).

For each novel cell line, two test scenarios are applicable, i.e. few-shot learning (few perturbation samples available) and zero-shot learning (no perturbation samples available). For few-shot learning, we adopted two distinct strategies to evaluate the performance of STAMP: (1) Training STAMP from scratch with random initialization, using a small number of available perturbations for the novel cell lines. Evaluating the first strategy is the most straightforward way to ascertain the ability of a model to learn and predict perturbation outcomes from a constrained dataset, which is a common scenario in practical applications. (2) Fine-tuning STAMP, initially pretrained on a dataset from a different cell line, on the same small set of perturbations available for the novel cell lines. For zero-shot learning, only the second strategy is applicable while the zero-shot prediction is conducted by directly predicting outcomes for novel cell lines using pretrained STAMP without the need for any finetuning. Noted that the evaluations of the second strategy in two test scenarios were constrained by the use of cell lines with the same CRISPR type. During this process, one of the cell line datasets (the source dataset) was used to pretrain the model and generalize it to predict the novel cell line datasets (the target datasets). To this end, STAMP was trained in two phases: on all perturbations from the source cell line dataset (pretraining phase) and then on the small number of perturbations available from one of the target cell line datasets (fine-tuning phase) or zero number of perturbations available for zero-shot learning without fine-tuning. Note that in all the tests involving the second strategy, for simplicity, we use the name of the dataset to represent the model (for example, “PRJNA609688 pretrain” denotes the STAMP model pretrained on the PRJNA609688 dataset and fine-tuned with a small set of samples from the target dataset). As “scGPT+STAMP” showed superior performance and fast computational speeds in our previous study, we used it as the representative STAMP model in this challenging evaluation (Supplementary Table 2). Furthermore, to establish a comprehensive comparison, we included “GEARS” and “scGPT+CPA” as baseline models for comparison purposes. Each evaluation was conducted via fivefold cross-validation, and the model evaluation was conducted across the three subtasks. Moreover, a focused evaluation was also conducted on the second and third subtasks, particularly for the top 40 DEGs.

In the evaluations of the first strategy, the performance of “GEARS” and “scGPT+CPA” improved slowly as samples from the new cell line dataset were added to the training set. In contrast, “scGPT+STAMP” improved rapidly, with a large increase in performance after examining only an additional 10-20 samples in most cases (Figs. 4b-d, Extended Data Fig. 4, Supplementary Figs. 6-12). In the evaluations of the second strategy, we evaluated the performance of “scGPT+STAMP” when it was pretrained on different source datasets and subsequently applied to a target dataset (Figs. 4b-d, Extended Data Fig. 4, Supplementary Figs. 6-12). Interestingly, although the pretrained “scGPT+STAMP” model sometimes exhibited superior performance in the zero-shot setting39 (sample number =0) compared to that of training “scGPT+STAMP” from scratch, the latter often outperformed the former in the few-shot setting44 (sample number > 0; Supplementary Figs. 6-12). This discrepancy could be attributed to the gene interaction pattern differences across different cell lines. The presence of consistent gene interaction patterns between certain cell lines contributed to the enhanced performance of the pretrained STAMP model in the zero-shot setting, surpassing that of the randomly initialized models. However, differences in the interaction patterns of certain gene pairs could lead to a nonbenign parameter initialization process, limiting the improvement exhibited by the model in the few-shot setting. Notably, this phenomenon diminished when the model was both trained and evaluated within the same cell line (Figs. 4b-d, Extended Data Fig. 4), where the PRJNA609688 dataset and the PRJNA551220 dataset represented the same cell line from distinct sequencing experiments. When conducting the evaluations on the PRJNA609688 dataset and the PRJNA551220 dataset, STAMP was pretrained on one of these two datasets, and it consistently achieved robust performance in both zero-shot and few-shot settings (Figs. 4b-d, Extended Data Fig. 4). Further exploration into the correlation between the zero-shot performance of pretrained models and cell line similarity revealed a notable trend: models pretrained on cell line datasets with higher degrees of similarity tend to exhibit enhanced performance in the zero-shot setting (Supplementary Fig. 13). Collectively, these findings highlight the ability of STAMP to adapt across cell lines and its robustness to experimental batch effects45.

To further demonstrate that STAMP can quickly adapt to a novel cell line, we applied “scGPT+STAMP” to the scCRISPR-seq dataset46, for which only 24 perturbations were available (Supplementary Note 8). On this dataset46, Papalexi et al. applied ECCITE-seq47 to explore the molecular networks that regulate PD-L1 expression and reported that IFNGR1, IFNGR2, IRF1, JAK2, and STAT1 are positive regulators of PD-L1, while BRD4 and CUL3 are negative regulators of PD-L146 (Supplementary Note 8). Therefore, in this scenario we examined whether STAMP could accurately predict the effects of the 7 perturbations on PD-L1 expressions with a small set of perturbations. As the first strategy (Training STAMP from scratch) exhibited superior performance in few-shot setting and it is a more straightforward way in practical scenarios, it was adopted in this dataset. STAMP without training (denoted as “Random” in this scenario) and “GEARS” were used as baseline models for comparison purposes. As the Pearson correlation coefficient and MSE cannot be calculated with only a single gene expression value, the performances of the baseline “scGPT+STAMP”, “GEARS”, and “Random” models were examined only in the first and second subtasks. In the first subtask, if the predicted score of PD-L1 for a certain gene perturbation was above the 80th percentile, then PD-L1 was classified as a differentially expressed gene for that perturbation (using the 85th, 90th, 95th percentiles, we arrived at similar conclusions; Supplementary Fig. 13). In the second subtask, the 50th percentile was used as the threshold for defining the direction of regulation.

As the results showed, in the first subtask, “scGPT+STAMP” successfully predicted that 5 out of these 7 genetic perturbations had significant effects on PD-L1 expression, outperforming the “GEARS” and “Random” baseline models (Fig. 4e). Furthermore, “scGPT+STAMP” achieved the best performance in the second subtask (Fig. 4e). In addition, we evaluated the performance of STAMP in a more challenging scenario where both the first and second subtasks for perturbation prediction needed to be correct. “scGPT+STAMP” successfully predicted 4 (IFNGR1, IFNGR2, IRF1, and JAK2) out of the 7 perturbations, which was twice the success rate of “GEARS” (Fig. 4e). In subsequent analyses, we explored the mechanisms underlying the downregulation of PD-L1 expression. For each of these 4 successfully predicted genetic perturbations, all DEGs in the first subtask identified by “scGPT+STAMP” were subjected to a gene ontology (GO) analysis48 (Supplementary Note 10). Our findings revealed that these perturbations led to the inhibition of immune-related pathways (Fig. 4f and Supplementary Fig. 14), resulting in a decrease in PD-L1 expression. This outcome aligned with the findings from prior research46, 49, 50, further validating the predictive accuracy of our model. In summary, the above results demonstrate that STAMP can quickly adapt to a novel cell line, which is essential in the perturbation prediction field because perturbation sequencing technology is still in its early stage and cost-ineffective, and many cell lines have limited perturbation data. was fine-tuned with various numbers of samples from the target datasets (0, 2, 4, 6, 8, 10, 15, 20, 25, 30, and 40 samples). The plot shows the average performance as a function of the number of samples utilized from the target cell line dataset. e. The accuracies achieved by different models in three testing scenarios, i.e., the first subtask, the second subtask, and the testing scenario where both the first subtask and the second subtask needed to be correct. f.

In scenarios involving two-gene perturbations, a simple way to predict the outcome of a combinatorial perturbation might involve summing the individual effects of perturbing each gene separately. However, this approach is primarily applicable when the gene interactions (GI) are additive, where they are referred to as additive interactions5 (Supplementary Note 11). Biological reality is often complex, with genes known to engage in various forms of nonadditive interactions following perturbation. These interactions, which have been defined in previous studies5, include subtypes such as synergy, suppression, neomorphism, redundancy and epistasis10 (Fig. 5b, Supplementary Note 12). Genetic interaction scores (GI scores), such as magnitudes, model fitting scores, the equality of contributions and similarity scores, have been defined to identify these different genetic interaction subtypes5, 10 (Supplementary information). Previous work10 has advocated for the use of a single GI score threshold to categorize different GI interaction subtypes (Supplementary Note 11). In this study, we evaluated the performance of “scGPT+STAMP” and “GEARS” in terms of identifying new genetic interactions. To estimate the upper performance limit, we also calculated the GI scores based on the ground-truth gene expression values10 (Ground truth in Fig.5) and collected the GI scores calculated in the original study5 for all combinatorial genes (Ori_study in Fig.5, Supplementary Note 12). We also utilized Precision@10 to evaluate the performance of the models (Supplementary Note 13). This metric calculates the proportion of predicted combinations within the top ten that truly exhibit a specific genetic interaction subtype, as reported by the original study. Compared to “GEARS”, “scGPT+STAMP” improved the Precision@10 metrics obtained for four out of five GI subtypes, and the improvements exceeded 50% for neomorphism, redundancy and epistasis (Fig. 5a).

However, all GI scores from different models and sources showed that the tested models struggled to identify suppression and redundancy subtypes (Fig. 5a). Notably, only the GI scores from the original study had a Precision@10 of 0.2 for suppression, and the GI scores from the original study had a Precision@10 of 0.4 for redundancy (Fig. 5a). This indicated that relying on a single GI score is insufficient for interpreting all GI interaction subtypes and highlighted the importance of designing a more suitable GI score threshold system to accurately identify GI subtypes, as a single GI score threshold falls short of capturing the complexity inherent in these GIs (Fig. 5a, Supplementary Notes 12-13).

To develop a more accurate GI score threshold system for identifying GI subtypes, we constructed a decision tree51 model utilizing GI scores calculated from ground-truth gene expression values (Supplementary Notes 12-13). This approach was chosen due to the superior explanatory power of decision trees, facilitating a clearer understanding of the classification criteria and underlying patterns in the data (Fig. 5c). Based on the decision tree model, we observed that the upper accuracy limit of identifying different GI subtypes based on the ground-truth gene expression values exceeded 0.75 across all nonadditive GI subtypes. Notably, this accuracy exceeded 0.85 for four out of the five nonadditive subtypes (Fig. 5d). When “scGPT+STAMP” was compared with “GEARS” and the “Random” baseline, we also incorporated additive interactions into this evaluation system. STAMP outperformed the other models for five out of the six GI subtypes, while STAMP exhibited comparable accuracy to that of “GEARS” for the epistasis subtype (Fig. 5d). Collectively, our findings indicate that the newly designed GI score threshold system offers a more accurate evaluation of the effectiveness of different models in terms of identifying GI subtypes and that STAMP exhibits a superior ability to accurately discern these GIs, underscoring its potential role as a robust model in this domain. a. Precision@10 scores obtained when predicting genetic interactions from 125 two-gene perturbations with GI scores obtained by STAMP (“scGPT+STAMP”), GEARS (“GEARS”), the ground-truth gene expression vector (“Ground truth”) and the original study (“Ori_study”). b. Definitions of genetic interaction subtypes. c. The decision tree of the GI score threshold system was constructed from the ground-truth gene expression vector, and each GI score was normalized to a z score. d. Comparison among the accuracies of different models in terms of predicting genetic interactions with the decision tree constructed in c.

a. The ROC-AUC and PR-AUC values produced by different models in subtask-2. b. The

Extended Data Fig. 4: Challenge 3. Genetic perturbation outcomes predicted across cell lines in the PRJNA551220 dataset (considering the top 40 DEGs). a. The ROC-AUC and PR-AUC values produced by different models on pretrained cell line datasets in subtask-2. b. The Pearson correlations and MSEs yielded by different models on pretrained cell line datasets in subtask-3. Each model was fine-tuned with various numbers of samples from the target datasets (0, 2, 4, 6, 8, 10, 15, 20, 25, 30, and 40 samples). The plot shows the average performance as a function of the number of samples acquired from the target cell line dataset.

Tang, F. et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nature methods 6, 377-382 (2009). (2022).

Bock, C. et al. High-content CRISPR screening. Nature Reviews Methods Primers 2, 8 Dixit, A. et al. Perturb-Seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens. cell 167, 1853-1866. e1817 (2016).

Cheng, J. et al. Massively Parallel CRISPR‐Based Genetic Perturbation Screening at Single‐Cell Resolution. Advanced Science 10, 2204484 (2023).

Norman, T.M. et al. Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science 365, 786-793 (2019).

Kamimoto, K. et al. Dissecting cell identity via network inference and in silico gene 10.

Qiu, X. et al. Mapping transcriptomic vector fields of single cells. Cell 185, 690-711. e645 (2022).

Adamson, B. et al. A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response. Cell 167, 1867-1882. e1821 (2016).

Jaitin, D.A. et al. Dissecting immune circuits by linking CRISPR-pooled screens with single-cell RNA-seq. Cell 167, 1883-1896. e1815 (2016).

Roohani, Y., Huang, K. & Leskovec, J. Predicting transcriptional outcomes of novel multigene perturbations with gears. Nature Biotechnology, 1-9 (2023).

Lotfollahi, M. et al. Predicting cellular responses to complex perturbations in high‐ 28.

Piękos, P., Malinowski, M. & Michalewski, H. in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers) 383-394 (2021). Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022).

Zhang, Y. & Yang, Q. A survey on multi-task learning. IEEE Transactions on Knowledge and Data Engineering 34, 5586-5609 (2021).

Replogle, J.M. et al. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell 185, 2559-2575. e2528 (2022).

Joung, J. et al. A transcription factor atlas of directed differentiation. Cell 186, 209-229. e226 (2023).

Wolf, F.A., Angerer, P. & Theis, F.J. SCANPY: large-scale single-cell gene expression data analysis. Genome biology 19, 1-5 (2018).

Squair, J.W. et al. Confronting false discoveries in single-cell differential expression. Nature communications 12, 5692 (2021).

Cohen, I. et al. Pearson correlation coefficient. Noise reduction in speech processing, 1-4 (2009).

Prasad, N.N. & Rao, J.N. The estimation of the mean squared error of small-area estimators. Journal of the American statistical association, 163-171 (1990).

Frangieh, C.J. et al. Multimodal pooled Perturb-CITE-seq screens in patient models define mechanisms of cancer immune evasion. Nature genetics 53, 332-341 (2021). Schmidt, R. et al. CRISPR activation and interference screens decode stimulation responses in primary human T cells. Science 375, eabj4008 (2022).

Wang, W., Zheng, V.W., Yu, H. & Miao, C. A survey of zero-shot learning: Settings, methods, and applications. ACM Transactions on Intelligent Systems and Technology (TIST) 10, 1-37 (2019).

Gao, Y. et al. Pan-Peptide Meta Learning for T-cell receptor–antigen binding recognition. Nature Machine Intelligence, 1-14 (2023).

Hospedales, T., Antoniou, A., Micaelli, P. & Storkey, A. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence 44, 5149-5169 (2021).

Tian, R. et al. Genome-wide CRISPRi/a screens in human neurons link lysosomal failure to ferroptosis. Nature neuroscience 24, 1020-1034 (2021).

Replogle, J.M. et al. Combinatorial single-cell CRISPR screens by direct guide RNA capture and targeted sequencing. Nature biotechnology 38, 954-961 (2020).

Wang, Y., Yao, Q., Kwok, J.T. & Ni, L.M. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur) 53, 1-34 (2020).

Tran, H.T.N. et al. A benchmark of batch-effect correction methods for single-cell RNA sequencing data. Genome biology 21, 1-32 (2020).

Papalexi, E. et al. Characterizing the molecular regulation of inhibitory immune checkpoints with multimodal single-cell screens. Nature genetics 53, 322-331 (2021). Mimitou, E.P. et al. Multiplexed detection of proteins, transcriptomes, clonotypes and CRISPR perturbations in single cells. Nature methods 16, 409-412 (2019).

Young, M.D., Wakefield, M.J., Smyth, G.K. & Oshlack, A. Gene ontology analysis for

Moon, J.W. et al. IFNγ induces PD-L1 overexpression by JAK2/STAT1/IRF-1 signaling

Garcia-Diaz, A. et al. Interferon receptor signaling pathways regulating PD-L1 and

De Ville, B. Decision trees. Wiley Interdisciplinary Reviews: Computational Statistics 5, 448-455 (2013).

Song, B. et al. Decoding Heterogenous Single-cell Perturbation Responses. bioRxiv,