Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Deep Active Learning based Experimental Design to Uncover Synergistic Genetic Interactions for Host Targeted Therapeutics

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An ensemble deep active-learning framework uncovers 92% of the top 400 HIV gene pairs after observing less than 6.3% of the pairwise knockdown matrix.

desk verdict First double-knockdown active learning at 356×356 scale; plausible 92% coverage on one dataset, but bilinear assumption and missing artifacts bound the claim. read the letter →

arxiv 2502.01012 v1 pith:W3K2WIWK submitted 2025-02-03 cs.LG q-bio.QMstat.ME

Haonan Zhu , Mary Silva , Jose Cadena , Braden Soper , Micha{l} Lisicki , Braian Peetoom , Sergio E. Baranzini , Shivshankar Sundaram
show 2 more authors
Priyadip Ray Jeff Drocco
This is my paper · ORCID
classification cs.LGq-bio.QMstat.ME MSC 62K0568T07
keywords DeepActiveLearningSequentialExperimentalDesignCombinationTherapiesGene-GeneInteractionsKnockdownExperimentsRepresentationEnsembleMethodGraph
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that double-gene knockdown experiments, which search a combinatorial space far too large for brute-force testing, can be guided by a deep active-learning loop with biological priors from a knowledge graph. The framework embeds genes from a biomedical knowledge graph (SPOKE), predicts the viral-load effect of any unobserved gene pair with a bilinear scoring model, and uses an ensemble of 20 models to decide which 400 pairs to test in each round. On the measured 356-by-356 matrix of HIV double-knockdown viral loads, the paper reports that this procedure uncovers 92% of the 400 most effective gene pairs after observing less than 6.3% of the matrix. If this holds, combination-therapy screens could prioritize the most promising host targets without brute-force testing of all roughly 127,000 pairs. The work matters because existing active-learning methods for genetic screens have been demonstrated only on single knockdowns or small double-knockdown datasets.

What carries the argument

The load-bearing object is the bilinear regression model $\hat{y}_{ij} = \mathrm{Softplus}(x_i^\top A x_j) + b$, where $x_i$ and $x_j$ are gene embeddings produced by a Relational Graph Convolutional Network (R-GCN) trained on a subgraph of the SPOKE biomedical knowledge graph. The bilinear form is what lets the model rank unobserved gene pairs from their embeddings alone, and the Softplus output plus bias keeps predictions on the positive viral-load scale. Before the active-learning loop starts, the embeddings are initialized by self-supervised link prediction with the DistMulti scoring model, so early rounds begin from biologically meaningful geometry. In the loop, an ensemble of $M=20$ independently initialized models is trained on the observed pairs with a Huber loss, and the spread of their predictions defines the uncertainty that acquisition strategies consume: the winning strategy picks pairs with the lowest 10% quantile of ensemble predictions, while the maximum-variance strategy picks pairs the ensemble disagrees about most.

What would settle it

Run the identical active-learning loop on a second double-knockdown dataset whose interaction surface is known to be high-rank or otherwise poorly matched to the bilinear form, for example a synthetic landscape generated by nonlinear pairwise rules the model cannot express, and check whether Coverage@400 still approaches 92% at roughly 6.3% observation; the central claim fails if coverage collapses on that landscape. A cheaper in-dataset check is to remove the low-rank constraint, replacing the shared embedding space with per-pair parameters, and see whether the reported coverage advantage of the constrained model disappears.

Watch

Extended reading notes

Core claim

The central claim is that a unified deep active learning framework can sequentially design double-knockdown experiments at the scale of a 356-by-356 interaction matrix. Gene embeddings are first learned by a Relational Graph Convolutional Network on a subgraph of the SPOKE knowledge graph, then fine-tuned on observed viral-load values; the regression head predicts the viral load of a double knockdown as $\hat{y}_{ij} = \mathrm{Softplus}(x_i^\top A x_j) + b$, a bilinear form over the two genes' embeddings. An ensemble of $M=20$ such models supplies uncertainty estimates, and the acquisition rule that minimizes the 10% quantile of ensemble predictions ('optimism') achieves 92% Coverage@400, meaning 92% of the top-400 gene pairs are found after under 6.3% of the matrix is observed. The ablation results support three subsidiary claims: the ensemble's uncertainty information, rather than the base model's greedy picks, drives long-run coverage; the knowledge-graph initialization accelerates early discovery but its advantage fades as data accumulate; and fine-tuning the embeddings on experimental data is essential. Pathway enrichment of the selected pairs shows both acquisition strategies concentrate on translation-related biological processes, which the paper reads as biologically coherent with HIV Tat's role in RNA polymerase II recruitment.

Load-bearing premise

The framework assumes the viral-load effect of any double knockdown is well captured by a low-rank bilinear score built from knowledge-graph gene embeddings; if real genetic interactions deviate from that shape, the ranking of unobserved pairs degrades and the 92% coverage result may not transfer to other datasets.

Editorial extensions

If this is right

  • On this dataset, a screening campaign for combination-therapy targets would not need to test all pairs: 92% of the top-400 pairs are recovered at a budget under 6.3% of the matrix, corresponding to the paper's protocol of 17 rounds of 400 observations.
  • Ensemble uncertainty, not just point predictions, is what buys long-run coverage: the greedy base model lags the optimism strategy, while maximum variance produces the lowest prediction error on unseen pairs even though it covers fewer top pairs.
  • The knowledge-graph prior matters most at the start: the ablation with random initialization shows SPOKE embeddings accelerate early discovery, but the benefit diminishes as observed data accumulate, so the graph and the data complement each other.
  • The selected gene pairs are interpretable: genes picked across rounds are enriched for translation and transcription-elongation processes, consistent with HIV Tat enhancing RNA polymerase II activity, which supports the biological plausibility of the recommendations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because nothing in the machinery is HIV-specific, the same loop should transfer to other host-pathogen interaction screens or drug-combination assays, provided those landscapes are also approximately low-rank in gene-embedding space; that proviso is the untested premise that decides whether the 92% result generalizes.
  • The bilinear scoring rule captures pairwise effects only; moving to triple knockdowns or higher-order combinations would require a tensor analogue of $x_i^\top A x_j$, although the active-learning loop and ensemble uncertainty machinery would carry over.
  • The coverage metric treats the measured top-400 pairs as ground truth; re-running the benchmark against a noisy surrogate of the viral-load matrix would quantify how much of the reported 92% survives realistic assay error, a robustness check the paper does not perform.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a deep active learning (DeepAL) framework for selecting pairwise gene-knockdown experiments to identify combinations that reduce HIV viral load. Gene embeddings are obtained from an R-GCN trained on the SPOKE knowledge graph, and a bilinear regression head (Eq. 5) predicts viral load for a pair. An ensemble of M=20 models provides uncertainty estimates, and several acquisition functions (greedy, BADGE, maximum variance, optimism at the 10% quantile) are compared. On a 356×356 HIV double-knockdown dataset, the optimism strategy achieves 92% Coverage@400 after observing less than 6.3% of the matrix, with ablations showing that ensemble uncertainty, SPOKE-initialized embeddings, and fine-tuning contribute to performance. The paper also includes a pathway enrichment analysis of selected pairs.

Significance. If confirmed, the result is of practical interest for combinatorial experimental design in host-targeted therapies, where brute-force screening of pairwise knockdowns is infeasible. The manuscript includes a thoughtful ablation suite (random initialization, frozen features, unconstrained features) that isolates the contribution of the knowledge graph and fine-tuning, and the use of a heterogeneous knowledge graph (SPOKE) is a sensible design choice. The evaluation is, however, based on a single real dataset, and the headline coverage figure is reported without any measure of uncertainty, so the strength of the evidence for the central claim is currently limited.

major comments (4)
  1. [Section 4, Fig. 2(a)] The headline claim that the ensemble DeepAL framework uncovers 92% of the top 400 gene-pairs after observing less than 6.3% of the matrix is stated as a point estimate with no measure of variability across the 20 replicates described in the figure captions. Without error bars, a standard deviation, or a confidence interval for the final-round coverage, the reader cannot assess whether the difference between the optimism strategy and the greedy or random baselines is statistically reliable. Please report the mean and variance (or confidence interval) over replicates for the terminal coverage, and consider showing error bars or shaded regions in Figs. 2 and 3.
  2. [Section 3.2.3, Eq. (5)] The regression model assumes the double-knockdown viral load is, up to a Softplus transform, a bilinear form in the SPOKE-derived gene embeddings (with a shared symmetric matrix A). The entire active-learning loop inherits this low-rank restriction, but the paper never tests whether the success of the framework depends on this assumption. The UF-Ensemble ablation removes the graph prior but retains the same bilinear head, so it cannot detect a misspecification of the head itself. To support the generalization of the central claim, add a comparison with a strictly more expressive predictor (e.g., an MLP over the embeddings, or a higher-rank factorization) in the same active-learning loop, or a synthetic experiment with known non-bilinear interaction structure. Without such a check, the 92% coverage result may be a property of this particular dataset's approximate low-rank landscape rather than of the proposed framework's general capacity.
  3. [Title, Abstract, Section 3.2.3] The word 'synergistic' appears in the title and throughout, but Eq. (5) models the joint effect of a double knockdown directly; it does not include or compare against single-gene effects, so it cannot distinguish synergy from mere strong additivity. A synergistic interaction is defined by a deviation from the additive expectation of single perturbations. As written, the paper demonstrates efficient discovery of highly effective double-knockdown pairs, but not that these pairs are synergistic. Please either incorporate an additive baseline into the model (e.g., predict y_ij and compare to f(x_i)+f(x_j)) or revise the claims and title to say 'potent pairwise knockdowns' or 'effective pairwise interventions'.
  4. [Reproducibility (entire manuscript)] The manuscript does not provide code, data, or a full training configuration (e.g., learning rates, optimizer, epochs, random seeds, hardware). The central result is a single-number claim about algorithm performance in a simulation; without the implementation details or a release, the result cannot be independently verified. Provide a reproducibility statement with code and data availability, or at minimum complete hyperparameter and training details and the variance over replicates.
minor comments (6)
  1. [Section 4.1] The relationship between the stated 17 rounds × 400 pairs = 6,800 observations and 'less than 6.3% of the entire matrix' (6,800 / 126,736 ≈ 5.36%) should be clarified; the percentage appears to have been rounded or computed differently.
  2. [Algorithm 1, Table 1] Provide concrete values for the regularization parameters (dropout rate, weight decay), optimizer, learning rate, number of epochs, and how ensemble diversity is maintained across the 20 seeds.
  3. [Table 2] The 'Badge' acquisition is described as arg max over ||x_i ⊙ x_j||, which does not correspond to the gradient-based BADGE method of Ash et al. (ref. [2]); clarify whether this is an approximation or rename the strategy.
  4. [Section 4.4] State the significance threshold and multiple-testing correction used for the GOATOOLS enrichment analysis; otherwise the claim that terms are 'overrepresented' is not quantitatively defined.
  5. [Figures 2 and 3] Add error bars or shaded intervals to the plots, since they show averages over 20 replicates without any dispersion information.
  6. [Throughout] Typographical errors should be fixed: 'Abalation' in Fig. 3; 'Mean Absolut Errors' in the y-axis labels; 'summarized over20 replicates' missing a space.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DeepAL result is a supervised active-learning evaluation against a fixed ground-truth matrix, not a reduction of the claim to its inputs.

full rationale

The paper's central claim, that the ensemble DeepAL framework uncovers 92% of the top 400 gene pairs after observing less than 6.3% of the matrix, is an empirical active-learning result evaluated against a fixed external ground truth. The bilinear predictor in Eq. 5, y_hat_ij = Softplus(x_i^T A x_j) + b, is trained only on observed pairs via the Huber loss in Algorithm 1 line 7, and the acquisition functions in Table 2 select batches from model predictions. Coverage@400 is then computed against the pre-existing top-400 viral-load pairs from the Gordon et al. dataset, not against the model's own scores, so nothing in the equations forces the discovered set to coincide with the ground-truth top 400. If the low-rank bilinear assumption or the SPOKE-derived embeddings failed, the coverage would drop; the result is therefore not equivalent to the model's training objective by construction. The SPOKE knowledge graph is an external prior rather than an output of this paper, and no load-bearing step relies on a self-citation chain: the model architecture is justified by citations to standard external methods (R-GCN, DistMulti, deep ensembles), not by prior work of these authors. The mild concern that the 92% result may not transfer to other interaction maps because of the bilinear assumption is a generalizability and correctness question, not a circularity. No fitted parameter is renamed as a prediction, no quantity is defined in terms of the target claim, and no uniqueness theorem is imported from the authors' own prior work. Therefore the derivation is self-contained with respect to circularity.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

The central claim depends on the bilinear low-rank model, the external knowledge graph prior, the fixed HIV dataset, and several hand-chosen hyperparameters (d=50, L=3, M=20, batch 400, quantile 0.10). No new biological or mathematical entities are introduced.

free parameters (9)
  • Embedding dimension d = 50
    Set by hand in Table 1; controls the rank of the bilinear model.
  • Hidden dimension d_h = 64
    Set by hand in Table 1 for the R-GCN.
  • Number of R-GCN layers L = 3
    Chosen to balance expressiveness and over-smoothing (footnote 3).
  • Ensemble size M = 20
    Used for uncertainty quantification; no ablation across M is shown.
  • Batch size N = 400
    Chosen to balance computational and experimental capabilities; also used in the 17-round evaluation.
  • Rounds T = 17
    Evaluation budget; 17 rounds at 400 pairs yields 6800 observations.
  • Random walk length s = 5
    Subgraph sampling in Section 3.1.
  • Random walks per node w = 5
    Subgraph sampling in Section 3.1.
  • Optimism quantile = 0.10
    Acquisition parameter selected because it achieved the best coverage in the comparison; this is a form of tuning on the benchmark.
assumptions (6)
  • domain assumption The HIV genetic interaction matrix from Gordon et al. [3] is an accurate ground truth for viral load after double knockdown.
    Introduced in Section 2.2; the entire simulation evaluates against this fixed matrix, so errors in the source data propagate to the reported coverage.
  • domain assumption The viral-load landscape can be approximated by the bilinear form y_ij = Softplus(x_i^T A x_j) + b with shared gene embeddings.
    Equation (5) in Section 3.2.3; this low-rank assumption determines which unobserved pairs can be ranked correctly and is not validated against alternative interaction models.
  • domain assumption SPOKE graph node and edge structure carries information relevant to HIV-related gene function.
    Section 3.1; the framework relies on random-walk subgraphs from SPOKE to initialize embeddings, and the ablation shows this helps, but the graph itself is an external resource whose relevance is assumed.
  • domain assumption Active learning simulation on a complete observed matrix faithfully represents the real experimental design problem.
    Section 4; the paper re-labels entries as observed/unobserved, so it assumes the cost and noise structure of real experiments are captured by this simulation.
  • ad hoc to paper The 10% optimism quantile acquisition is a reasonable exploration-exploitation heuristic.
    Section 3.4 and 4.2; chosen because it performed best in the benchmark, not derived from a regret bound or decision-theoretic principle.
  • standard math Standard R-GCN, DistMulti, and ensemble methods behave as described in their source references.
    Used throughout Section 3; the paper relies on existing results for graph neural networks and deep ensembles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Active Learning based Experimental Design to Uncover Synergistic Genetic Interactions for Host Targeted Therapeutics." pith.science (2026). https://pith.science/paper/W3K2WIWK

@misc{pith2026250201012,
  author       = {Pith},
  title        = {Pith review of: Deep Active Learning based Experimental Design to Uncover Synergistic Genetic Interactions for Host Targeted Therapeutics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3K2WIWK}},
  note         = {Machine review of arXiv:2502.01012}
}
read the original abstract

Recent technological advances have introduced new high-throughput methods for studying host-virus interactions, but testing synergistic interactions between host gene pairs during infection remains relatively slow and labor intensive. Identification of multiple gene knockdowns that effectively inhibit viral replication requires a search over the combinatorial space of all possible target gene pairs and is infeasible via brute-force experiments. Although active learning methods for sequential experimental design have shown promise, existing approaches have generally been restricted to single-gene knockdowns or small-scale double knockdown datasets. In this study, we present an integrated Deep Active Learning (DeepAL) framework that incorporates information from a biological knowledge graph (SPOKE, the Scalable Precision Medicine Open Knowledge Engine) to efficiently search the configuration space of a large dataset of all pairwise knockdowns of 356 human genes in HIV infection. Through graph representation learning, the framework is able to generate task-specific representations of genes while also balancing the exploration-exploitation trade-off to pinpoint highly effective double-knockdown pairs. We additionally present an ensemble method for uncertainty quantification and an interpretation of the gene pairs selected by our algorithm via pathway analysis. To our knowledge, this is the first work to show promising results on double-gene knockdown experimental data of appreciable scale (356 by 356 matrix).

Figures

Figures reproduced from arXiv: 2502.01012 by the authors.

Figure 1
Figure 1. Flowchart of the proposed deep active learning framework. of gene-pairs to uncover. The overall framework is summa￾rized in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison among the different acquisition strategies, and results are summarized over 20 replicates. Optimism with 10% quantiles achieves the best performance in terms of coverage at terminal phase, and uncovers 92% of the top 400 gene-pairs while only < 6.3% of the entire matrix is observed. Maximum variance strategy performs the best in learning a generalizable model that predicts viral-replicates on unseen data … view at source ↗
Figure 3
Figure 3. Abalation studies to validate our proposed approach, and results are summarized over 20 replicates. The comparison between the base model and ensemble approaches shows that ensemble method is able to provides a meaningful uncertainty quantification that benefits coverage in the long-term; The comparison between DeepAL-Ensemble versus UF-Ensemble shows that the graph representation of the models from SPOKE offers a m… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Pathway enrichment analysis on the gene-pairs selected by DeepAL. The left panel shows the biological process terms selected optimizing by maximum variance, and the right panel shows the terms enriched when optimizing by maximum optimism. The x-axis represents the roun…
Figure 1
Figure 1. Figure 1: Comparison among the different acquisition strategies, and results are summarized over 20 replicates. Optimism with 10% quantiles achieves the best performance in terms of coverage at terminal phase. Maximum variance strategy performs the best in learning a generalizab…
Figure 2
Figure 2. Figure 2: Abalation studies to validate our proposed approach, and results are summarized over 20 replicates. Results for Batch Size=800 In this subsection, we provide additional results where we run all algorithms for 9 rounds (including the initial round of random selection), …
Figure 3
Figure 3. Figure 3: Comparison among the different acquisition strategies, and results are summarized over 20 replicates. Optimism with 10% quantiles achieves the best performance in terms of coverage at terminal phase. Maximum variance strategy performs the best in learning a generalizab…
Figure 4
Figure 4. Figure 4: Abalation studies to validate our proposed approach, and results are summarized over 20 replicates [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 25 canonical work pages

  1. [1]

    A review of uncertainty quantification in deep learning: Techniques, applications and challenges

    Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Abbas Khosravi, U Rajendra Acharya, et al. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information fusion, 76:243–297, 2021

    2021

  2. [2]

    Deep batch active learning by diverse, uncertain gradient lower bounds.arXiv preprint arXiv:1906.03671, 2019

    Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds.arXiv preprint arXiv:1906.03671, 2019

    arXiv 1906

  3. [3]

    A quantitative genetic interaction map of hiv infection

    David E Gordon, Ariane Watson, Assen Roguev, Simin Zheng, Gwendolyn M Jang, Joshua Kane, Jiewei Xu, Jef- frey Z Guo, Erica Stevenson, Danielle L Swaney, et al. A quantitative genetic interaction map of hiv infection. Molecular cell, 78(2):197–209, 2020

    work page 2020

  4. [4]

    Crispr-gpt: An llm agent for automated design of gene-editing experiments

    Kaixuan Huang, Yuanhao Qu, Henry Cousins, William A Johnson, Di Yin, Mihir Shah, Denny Zhou, Russ Altman, Mengdi Wang, and Le Cong. Crispr-gpt: An llm agent for automated design of gene-editing experiments. arXiv preprint arXiv:2404.18021, 2024

    arXiv 2024

  5. [5]

    Sequential optimal experimental design of perturbation screens guided by multi-modal priors

    Kexin Huang, Romain Lopez, Jan-Christian Hütter, Taka- masa Kudo, Antonio Rios, and Aviv Regev. Sequential optimal experimental design of perturbation screens guided by multi-modal priors. InInternational Conference on Re- search in Computational Molecular Biology, pages 17–37. Springer, 2024

    work page 2024

  6. [6]

    Robust estimation of a location param- eter

    Peter J Huber. Robust estimation of a location param- eter. In Breakthroughs in statistics: Methodology and distribution, pages 492–518. Springer, 1992

    work page 1992

  7. [7]

    Automated discov- ery of pairwise interactions from unstructured data.arXiv preprint arXiv:2409.07594, 2024

    Moksh Jain, Ali Denton, Shawn Whitfield, Aniket Didolkar, Berton Earnshaw, Jason Hartford, et al. Automated discov- ery of pairwise interactions from unstructured data.arXiv preprint arXiv:2409.07594, 2024

    arXiv 2024

  8. [8]

    Active learning to discover pairwise genetic interac- tions via representation learning

    Moksh Jain, Alisandra Kaye Denton, Shawn T Whitfield, Aniket Rajiv Didolkar, Berton Earnshaw, and Jason Hart- ford. Active learning to discover pairwise genetic interac- tions via representation learning. InICLR 2024 Workshop on Machine Learning for Genomics Explorations

    work page 2024

Show all 37 references
  1. [9]

    Bilinear bandits with low-rank structure

    Kwang-Sung Jun, Rebecca Willett, Stephen Wright, and Robert Nowak. Bilinear bandits with low-rank structure. In International Conference on Machine Learning, pages 3163–3172. PMLR, 2019

    2019

  2. [10]

    Goatools: A python library for gene ontology analyses

    Drew V Klopfenstein, Luke Zhang, Brent S Pedersen, et al. Goatools: A python library for gene ontology analyses. Scientific Reports, 8(1):10872, 2018

    2018

  3. [11]

    Simple and scalable predictive uncertainty es- timation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty es- timation using deep ensembles. Advances in neural information processing systems, 30, 2017

    2017

  4. [12]

    Discobax dis- covery of optimal intervention sets in genomic experiment design

    Clare Lyle, Arash Mehrjou, Pascal Notin, Andrew Jesson, Stefan Bauer, Yarin Gal, and Patrick Schwab. Discobax dis- covery of optimal intervention sets in genomic experiment design. InInternational Conference on Machine Learning, pages 23170–23189. PMLR, 2023

    2023

  5. [13]

    Genedisco: A benchmark for experimental design in drug discovery

    Arash Mehrjou, Ashkan Soleymani, Andrew Jesson, Pas- cal Notin, Yarin Gal, Stefan Bauer, and Patrick Schwab. Genedisco: A benchmark for experimental design in drug discovery. arXiv preprint arXiv:2110.11875, 2021

    2021 arXiv

  6. [14]

    Neural collapse with unconstrained features

    Dustin G Mixon, Hans Parshall, and Jianzong Pi. Neural collapse with unconstrained features. Sampling Theory, Signal Processing, and Data Analysis, 20(2):11, 2022

    2022

  7. [15]

    The scalable precision medicine open knowledge engine (spoke): a massive knowledge graph of biomedical information

    John H Morris, Karthik Soman, Rabia E Akbas, Xiaoyuan Zhou, Brett Smith, Elaine C Meng, Conrad C Huang, Gabriel Cerono, Gundolf Schenk, Angela Rizk-Jackson, et al. The scalable precision medicine open knowledge engine (spoke): a massive knowledge graph of biomedical informatio...

    2023

  8. [16]

    Exploring genetic interaction mani- folds constructed from rich single-cell phenotypes.Science, 365(6455):786–793, 2019

    Thomas M Norman, Max A Horlbeck, Joseph M Replogle, Alex Y Ge, Albert Xu, Marco Jost, Luke A Gilbert, and Jonathan S Weissman. Exploring genetic interaction mani- folds constructed from rich single-cell phenotypes.Science, 365(6455):786–793, 2019

    2019

  9. [17]

    Neural design for genetic perturbation ex- periments

    Aldo Pacchiano, Drausin Wulsin, Robert A Barton, and Luis Voloch. Neural design for genetic perturbation ex- periments. In The Eleventh International Conference on Learning Representations

  10. [18]

    Deep- walk: Online learning of social representations

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deep- walk: Online learning of social representations. InProceed- ings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701–710, 2014

    2014

  11. [19]

    Optimal design of experiments

    Friedrich Pukelsheim. Optimal design of experiments. SIAM, 2006

    2006

  12. [20]

    A crispr toolbox to study virus–host in- teractions

    Andreas S Puschnik, Karim Majzoub, Yaw Shin Ooi, and Jan E Carette. A crispr toolbox to study virus–host in- teractions. Nature Reviews Microbiology, 15(6):351–364, 2017

    2017

  13. [21]

    Hiv tat controls rna polymerase ii and the epigenetic landscape to transcrip- tionally reprogram target immune cells

    Jonathan E Reeder, Youn-Tae Kwak, Ryan P McNamara, Christian V Forst, and Iván D’Orso. Hiv tat controls rna polymerase ii and the epigenetic landscape to transcrip- tionally reprogram target immune cells. Elife, 4:e08955, 2015

    2015

  14. [22]

    Mapping information-rich genotype- phenotype landscapes with genome-scale perturb-seq.Cell, 185(14):2559–2575, 2022

    Joseph M Replogle, Reuben A Saunders, Angela N Pog- son, Jeffrey A Hussmann, Alexander Lenail, Alina Guna, Lauren Mascibroda, Eric J Wagner, Karen Adelman, Gila Lithwick-Yanai, et al. Mapping information-rich genotype- phenotype landscapes with genome-scale perturb-seq.Cell, 1...

    2022

  15. [23]

    Antiviral ther- apy

    Douglas D Richman and Neal Nathanson. Antiviral ther- apy. In Viral pathogenesis, pages 271–287. Elsevier, 2016. 8 Haonan Zhu et al

    2016

  16. [24]

    Best arm identification in graphical bi- linear bandits

    Geovani Rizk, Albert Thomas, Igor Colin, Rida Laraki, and Yann Chevaleyre. Best arm identification in graphical bi- linear bandits. In International Conference on Machine Learning, pages 9010–9019. PMLR, 2021

    2021

  17. [25]

    Biodiscoveryagent: An ai agent for designing genetic perturbation experiments

    Yusuf Roohani, Andrew Lee, Qian Huang, Jian Vora, Zachary Steinhart, Kexin Huang, Alexander Marson, Percy Liang, and Jure Leskovec. Biodiscoveryagent: An ai agent for designing genetic perturbation experiments. arXiv preprint arXiv:2405.17631, 2024

    2024 arXiv

  18. [26]

    A survey on oversmoothing in graph neural networks

    T Konstantin Rusch, Michael M Bronstein, and Siddhartha Mishra. A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993, 2023

    2023 arXiv

  19. [27]

    Modeling re- lational data with graph convolutional networks

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. Modeling re- lational data with graph convolutional networks. InThe se- mantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedi...

    2018

  20. [28]

    Dropout: a sim- ple way to prevent neural networks from overfitting.The journal of machine learning research, 15(1):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a sim- ple way to prevent neural networks from overfitting.The journal of machine learning research, 15(1):1929–1958, 2014

    1929

  21. [29]

    A review of active learning approaches to experimental design for uncover- ing biological networks

    Yuriy Sverchkov and Mark Craven. A review of active learning approaches to experimental design for uncover- ing biological networks. PLoS computational biology, 13(6):e1005466, 2017

    2017

  22. [30]

    Hiv-1 antiretrovi- ral resistance: scientific principles and clinical applications

    Michele W Tang and Robert W Shafer. Hiv-1 antiretrovi- ral resistance: scientific principles and clinical applications. Drugs, 72:e1–e25, 2012

    2012

  23. [31]

    Complex embeddings for simple link prediction

    Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex embeddings for simple link prediction. InInternational conference on machine learning, pages 2071–2080. PMLR, 2016

    2016

  24. [32]

    Halluci- nation is inevitable: An innate limitation of large language models

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Halluci- nation is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024

    2024 arXiv

  25. [33]

    Embedding entities and relations for learn- ing and inference in knowledge bases

    Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learn- ing and inference in knowledge bases. arXiv preprint arXiv:1412.6575, 2014

    2014 arXiv

  26. [34]

    Towards deeper graph neural net- works with differentiable group normalization

    Kaixiong Zhou, Xiao Huang, Yuening Li, Daochen Zha, Rui Chen, and Xia Hu. Towards deeper graph neural net- works with differentiable group normalization. Advances in neural information processing systems, 33:4917–4928, 2020

    2020

  27. [35]

    A geometric analysis of neural collapse with unconstrained features.arXiv preprint arXiv:2105.02375, 2021

    Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features.arXiv preprint arXiv:2105.02375, 2021. Haonan Zhu received the Ph.D. de- gree in electrical and computer engineer- ing fro...

    2021 arXiv

  28. [2022]

    Sergio E Baranzinireceived his Ph.D

    He is currently a Postdoctoral Researcher at Baranzini Lab in UCSF. Sergio E Baranzinireceived his Ph.D. degree in molecular human genetics from the University of Buenos Aires in 1997. He is currently a distinguished Profes- sor of Neurology at the University of California San...

    1997 arXiv

  29. [2023]

    Mary Silvareceived an M.S

    He is currently a Postdoctoral Re- searcher with the Lawrence Livermore National Laboratory. Mary Silvareceived an M.S. in Statis- tics from University of California, Santa Cruz. She is currently a Data Scientist in the Biomolecular Design and Devel- opment Group at Lawrence L...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.