Pith. sign in

REVIEW 5 major objections 5 minor 35 references

Platform for Representation and Integration of multimodal Molecular Embeddings

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims nine gene embeddings from different modalities carry barely overlapping molecular signals; an autoencoder merging them (PRISME) matches the best single embedding and wins on missing-value imputation.

desk verdict A useful integration study with a plausible central claim, but the evidence is underpowered: no baselines, no significance tests, and the common-gene subset that anchors both the SVCCA and benchmark analyses is never characterized. read the letter →

arxiv 2507.07367 v1 pith:RW7LI3BH submitted 2025-07-10 q-bio.BM cs.LG

Erika Yilin Zheng , Yu Yan , Baradwaj Simha Sankar , Ethan Ji , Steven Swee , Irsyad Adam , Ding Wang , Alexander Russell Pelletier
show 3 more authors
Alex Bui Wei Wang Peipei Ping
This is my paper · ORCID
classification q-bio.BMcs.LG
keywords molecularembeddingsgeneembeddingintegrationautoencoderSVCCAmultimodalrepresentationlearningmissingvalueimputationbenchmarktasks
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

Existing molecular embeddings are each built from one data type — gene co-expression, single-cell transcriptomes, protein sequences, literature text, or knowledge graphs — and prior benchmarks have found that no single embedding dominates across tasks. This paper attempts to establish that these embeddings capture largely non-overlapping biological signals, making integration worthwhile. The authors build an adjusted SVCCA that subtracts the mean correlation from a shuffled-gene null distribution, and after this adjustment most embedding pairs show near-zero shared signal. They then train PRISME, an autoencoder that concatenates nine published gene embeddings into a single 512-dimensional latent vector, and benchmark it on nine prediction tasks. PRISME performs consistently on all tasks, reaches the top accuracy and AUC on gene-gene interaction and protein-protein interaction, and shows a clear advantage when genes missing from a task are imputed through the autoencoder.

What carries the argument

The load-bearing objects are two. First, the adjusted SVCCA workflow: it applies singular vector canonical correlation analysis to aligned gene embedding matrices, then builds an empirical null distribution by shuffling the gene order of full matrices one hundred times and recomputing SVCCA; subtracting the mean background correlation yields the reported similarity, the instrument by which the paper establishes that distinct modalities carry non-overlapping signals. Second, the PRISME autoencoder: a two-layer encoder with Leaky ReLU activation (negative slope 0.01) projecting the concatenated nine 512-dimensional embeddings first to 1024 and then to 512 dimensions, a single linear decoder, and a mean-squared-error loss with per-feature weights that make each input embedding contribute in proportion to the total dimension divided by that embedding's dimension. The 512-dimensional encoder output is the unified multimodal embedding that is benchmarked.

What would settle it

Count the genes with embeddings in all nine sources and compare their properties (number of publications, tissue expression breadth, disease associations) with the full protein-coding gene set. If the intersection is small or enriched for well-characterized genes, the near-zero adjusted SVCCA similarities and the imputation gains reported for PRISME would be expected to shrink or disappear on genes that lack coverage in one or more sources.

Watch

Extended reading notes

Core claim

The paper's central discovery is that representations of the same genes from different sources are complementary rather than redundant. Using the adjusted SVCCA, the raw canonical correlation for each embedding pair is corrected by subtracting the background correlation obtained from one hundred shuffled-gene runs; most pairs then fall near zero, and the only substantial remaining similarity (adjusted SVCCA 0.27) is between GenePT and BioLinkBERT, which were trained on the same text summaries. The paper interprets this as evidence that each modality contributes distinct molecular signals that no single embedding contains. Based on this, PRISME is an autoencoder that projects the concatenation of nine 512-dimensional embeddings into a 512-dimensional latent space, trained with a dimension-weighted reconstruction loss. In the benchmark evaluation, the integrated representation achieves the highest accuracy and AUC on gene-gene interaction prediction (accuracy 0.77, AUC 0.85) and protein-protein interaction prediction (accuracy 0.76, AUC 0.83), and in the missing-value imputation experiments the paper reports that it outperformed the individual embedding methods, producing a slight accuracy gain and a substantial AUC gain on the post-translational modification and pathology prognostics tasks.

Load-bearing premise

The argument depends on the genes used to measure complementarity and rank benchmarks being representative of the full human gene space: the analyses retain only genes or gene pairs that have embeddings in all nine sources, and the paper does not report how many genes survive that intersection or whether that subset is biased toward well-studied genes.

Editorial extensions

If this is right

  • No single data modality or embedding scheme is sufficient for general biomedical prediction; an integrated representation built from existing embeddings is the safer default.
  • A unified 512-dimensional representation can be used across diverse downstream tasks, removing the need to select or retrain an embedding per task.
  • When genes are missing from a task's input, PRISME's autoencoder can impute a representation from the remaining modalities, preserving or improving downstream accuracy and AUC.
  • The integration workflow is modular: new embedding methods and modalities can be appended as additional input channels without redesigning the autoencoder.

Reading between the lines

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

  • The complementarity measurements and benchmarks were computed on the intersection of all nine embedding sources, so the paper's headline result is only established for genes that happen to be covered everywhere; a direct test on genes absent from one or more sources is still missing.
  • Because the loss weights each embedding by total-dimension ratio, the autoencoder implicitly assumes every modality contributes equally informative signal; ablating these weights or learning them from data could show whether the integration is genuinely exploiting complementarity or simply averaging.
  • The imputation experiment trains each embedding's autoencoder separately for ten epochs and then imputes; applying PRISME to genes entirely missing from some embedding sources (rather than missing only from a task) would be a stronger test of the claimed imputation advantage.
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

5 major / 5 minor

Summary. The manuscript collects nine published gene/molecular embeddings spanning omics, literature/text, and knowledge-graph modalities, harmonizes them to 512-dimensional vectors, and proposes an adjusted SVCCA procedure that subtracts a shuffled-data background correlation to quantify representational overlap. Based on low adjusted SVCCA values among most embedding pairs, the authors conclude that existing embeddings capture largely non-overlapping molecular signals. They then introduce PRISME, an autoencoder that encodes the concatenated embeddings into a 512-dimensional latent space, and evaluate it on nine downstream prediction tasks, including missing-value imputation experiments. The paper claims that PRISME performs consistently on par with or better than individual embeddings and outperforms them in imputation scenarios, supporting the value of multimodal integration.

Significance. If the central claims are substantiated, the work would provide a useful public resource and a credible argument that no single existing gene embedding is sufficient for general biomedical prediction. The manuscript has clear strengths: the code is publicly available, the collection spans nine heterogeneous embedding sources, the evaluation covers nine benchmark tasks, and the adjusted SVCCA idea is a reasonable attempt to control for chance correlations. However, the evidence presented in the current version is under-supported: the common-gene subset is uncharacterized, the 512-dimensional projection is unspecified, the benchmark results lack error bars or significance tests, the imputation protocol is under-specified, and no concatenation or linear baseline is included. These gaps are load-bearing for the paper's two headline conclusions.

major comments (5)
  1. [§3.1] The statement that 'All embeddings were formatted as 512-dimensional vectors' is not backed by any description of the projection method. The source dimensions in Table 1 range from 200 (Gene2Vec) to 1024 (ProtTrans), so a projection must be applied, yet the text never states whether this is PCA, a learned linear map, a per-embedding autoencoder, or something else. This affects both the PRISME input and the SVCCA comparisons, and without this detail the results are not reproducible and the comparison may be biased by the projection choice.
  2. [Figure 3 / §4.2] The adjusted SVCCA procedure is not fully specified and the reported values lack quantification of uncertainty. The paper does not state how many SVD components are retained, whether the reported 'SVCCA' value is the mean canonical correlation, the first canonical correlation, or the sum, and no standard deviations or significance thresholds are given for the adjusted scores. The claim that embeddings capture 'largely non-overlapping molecular signals' rests entirely on these point estimates. In particular, the reported perfect SVCCA similarity of 1.0 between the two Know2BIO embeddings needs explanation, since CCA on finite data with SVD truncation can yield near-1 correlations even for unrelated high-dimensional subspaces; the adjusted value alone does not resolve this.
  3. [Figure 4 / §4.4] The benchmarking results are presented as point estimates without error bars, repeated-run variance, or significance tests, so the claim that 'PRISME demonstrated consistent performance' is not statistically supported. More importantly, the comparison set is incomplete: there is no baseline consisting of simple concatenation of all nine embeddings, nor a concatenation followed by a linear layer, which is the natural control for an autoencoder integration. Without such a baseline, the observed performance could be driven by the extra capacity or the concatenation itself rather than by multimodal compression.
  4. [§4.4.1 and §7.2] The missing-value imputation protocol is too under-specified to verify the paper's headline imputation claim. It is unclear which genes are treated as missing, which embedding sources are missing for those genes, how a missing vector is represented and initialized, and how the autoencoder is trained to impute it. For individual embeddings, the sentence 'Each embedding was trained in an autoencoder for 10 epochs' is ambiguous, since a single-embedding autoencoder cannot by itself produce a vector for a gene absent from that embedding. The description in Section 7.2 also does not clarify whether the benchmark tasks in the imputation experiments use the same common-gene intersection as the main benchmarks.
  5. [§7.2 / common gene set] The paper never reports the number of genes or gene pairs that remain after requiring presence in all nine embedding sources, nor does it assess whether this common subset is representative of the full human gene space. The gene-gene interaction task explicitly 'retained only gene pairs present in all embedding sources,' and the SVCCA analyses and other benchmarks implicitly use the same intersection for fair comparison. If this common set is small or enriched for well-studied genes, both the measured complementarity and the benchmark ranking of PRISME may fail to generalize to genes missing from one or more sources—which is exactly the scenario the imputation experiments are meant to address.
minor comments (5)
  1. [Eq. (1)] The definition of the weight W_j is incomplete; the text describes weighting by 'the ratio of the total concatenated dimension to the dimensionality of that particular set,' but this should be stated formally in the equation or immediately after it, since the indexing over features within a set is not defined.
  2. [§3.3] The sentence 'The output layer from the encoder can be extracted to be used as the multimodal molecular embedding' is vague; please specify whether the extracted representation is the 512-dimensional output of the second linear layer or some other layer.
  3. [Figure 4 caption] The caption contains a grammatical error ('The graphs shows') and would benefit from stating what error bars or confidence intervals, if any, are shown; currently none are visible.
  4. [References] Reference [2] is formatted as a long consortium author list and is distracting; it should be abbreviated to a standard consortium citation.
  5. [§4.4.1] The phrase 'missing gene values' conflates missing genes with missing embedding vectors; please clarify whether the imputation is performed for genes that have no embedding in a particular source or for genes entirely absent from the embedding vocabulary.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PRISME's derivation and benchmarks are self-contained, and the central claims rest on independent SVCCA measurements and external downstream tasks rather than on re-fitted definitions or self-citation chains.

full rationale

The paper's central derivation chain is not circular. The adjusted SVCCA score is defined as the original SVCCA correlation minus a mean background correlation obtained by shuffling gene matrices over 100 iterations; this is a statistical calibration step, not a definition that presupposes the conclusion of non-overlap. PRISME is an autoencoder trained with the weighted MSE reconstruction loss in Eq. 1, reconstructing the input embeddings; no benchmark label or downstream outcome is used in its training, so evaluating PRISME on the nine downstream tasks is an independent test of the learned representation. The imputation experiments similarly train autoencoders on the input embeddings and then measure benchmark performance, which does not amount to fitting the benchmark labels through the integration procedure. The benchmark tasks include some that individual embeddings were originally designed for (e.g., Geneformer for dosage sensitivity), so PRISME may inherit their biases, but that is a contamination / generalization concern, not circularity: the autoencoder was not optimized on those tasks. The self-citation to Know2BIO [28] by the same authors is a data source included among the nine embeddings, but neither the non-overlap claim nor the PRISME advantage depends on this citation as an authority; the claims are supported by the paper's own SVCCA computations and by externally cited benchmarks [16, 34]. The paper's reliance on a common-gene subset for fair comparison (Section 7.2) is a limitation in external validity rather than a definitional reduction. Thus no step of the derivation is equivalent to its own input by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the choice of nine public embeddings, the unspecified projection to 512 dimensions, the permutation null for SVCCA, and the autoencoder architecture. None of these are derived; they are selection decisions. The paper also assumes that the common-gene subset preserves the full complementarity signal. No new physical entities are introduced.

free parameters (5)
  • SVCCA components retained
    The number of singular vectors/CCA components kept is not reported; SVCCA values and the adjusted complementarity scores depend on this choice.
  • shuffle iterations = 100
    The empirical null distribution uses 100 shuffles (Section 3.2, Figure 1); the stability of the subtracted background depends on this number.
  • autoencoder architecture dimensions = 1024, 512
    Encoder hidden layer 1024 and latent dimension 512 are hand-chosen without reported tuning (Section 3.3), and the latent size fixes the output representation.
  • dimension-based loss weights = D/d_j
    Loss weighting in Eq. 1 scales each embedding set by the ratio of total dimension to its own dimension; this weighting scheme is a modeling choice that affects reconstruction balance.
  • embedding projection to 512
    All embeddings are formatted as 512-dimensional vectors (Section 3.1) but the projection method is unspecified, and it is an implicit free choice.
assumptions (5)
  • standard math SVCCA and CCA linear-algebra definitions from Raghu et al. and Hardoon et al.
    The method relies on standard subspace alignment and canonical correlation theory without modification.
  • domain assumption HGNC identifier standardization preserves gene-level correspondence
    Section 3.1 maps identifiers via MyGene.info, Ensembl, UniProt; in ambiguous cases the first listed gene name is kept (Supp Method 1), which assumes one-to-one mapping is adequate.
  • domain assumption Full-matrix shuffling produces a valid null for representational overlap
    Section 3.2 treats shuffled correlations as chance background; this assumes breaking the gene alignment removes all meaningful signal while preserving marginal structure.
  • domain assumption Benchmark tasks from prior studies are valid measures of biological signal
    Nine tasks are adapted from Geneformer, Gene2Vec, DeepLoc, Kan-Tor et al.; the paper assumes these labels are reliable and not leaking via training data of component embeddings.
  • ad hoc to paper Reconstruction error is a sufficient objective for preserving task-relevant signal
    PRISME trains with weighted MSE (Eq. 1), and the paper assumes the resulting latent code retains predictive information without a direct check such as probing each modality's contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Platform for Representation and Integration of multimodal Molecular Embeddings." pith.science (2026). https://pith.science/paper/RW7LI3BH

@misc{pith2026250707367,
  author       = {Pith},
  title        = {Pith review of: Platform for Representation and Integration of multimodal Molecular Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RW7LI3BH}},
  note         = {Machine review of arXiv:2507.07367}
}
read the original abstract

Existing machine learning methods for molecular (e.g., gene) embeddings are restricted to specific tasks or data modalities, limiting their effectiveness within narrow domains. As a result, they fail to capture the full breadth of gene functions and interactions across diverse biological contexts. In this study, we have systematically evaluated knowledge representations of biomolecules across multiple dimensions representing a task-agnostic manner spanning three major data sources, including omics experimental data, literature-derived text data, and knowledge graph-based representations. To distinguish between meaningful biological signals from chance correlations, we devised an adjusted variant of Singular Vector Canonical Correlation Analysis (SVCCA) that quantifies signal redundancy and complementarity across different data modalities and sources. These analyses reveal that existing embeddings capture largely non-overlapping molecular signals, highlighting the value of embedding integration. Building on this insight, we propose Platform for Representation and Integration of multimodal Molecular Embeddings (PRISME), a machine learning based workflow using an autoencoder to integrate these heterogeneous embeddings into a unified multimodal representation. We validated this approach across various benchmark tasks, where PRISME demonstrated consistent performance, and outperformed individual embedding methods in missing value imputations. This new framework supports comprehensive modeling of biomolecules, advancing the development of robust, broadly applicable multimodal embeddings optimized for downstream biomedical machine learning applications.

Figures

Figures reproduced from arXiv: 2507.07367 by the authors.

Figure 2
Figure 2. Architecture of PRISME, which com￾prises an encoder and a decoder. The encoder comprises two hidden linear layers; the decoder is a single linear layer that reconstructs the input embedding from the encoded representation. The output of the encoder serves as the unified multi￾modal molecular embedding. The model is trained with the objective to minimize the mean squared error (MSE) between the reconstructed output a… view at source ↗
Figure 3
Figure 3. SVCCA (left) and adjusted SVCCA (right) similarity plots between molecular embed￾dings. 4.3 Benchmarking Tasks To comprehensively evaluate how well molecular embed￾dings capture biomedical signals, we curated a total of nine benchmark downstream prediction tasks, includ￾ing gene dosage sensitivity [23], gene–gene interaction [23], Gene Ontology (GO) [5], protein–protein interac￾tion (PPI) [20], protein subcellular l… view at source ↗
Figure 4
Figure 4. PRISME demonstrates robust performance across nine molecular property prediction bench￾marks, with notable improvements in missing value imputation tasks. The graphs shows performance comparison of PRISME versus individual embedding methods on nine benchmarking tasks. PRISME achieves highest performance on gene-gene interaction (Acc.=0.77, AUC=0.85) and PPI prediction (Acc.=0.76, AUC=0.83). specific task. This is po… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [1]

    Uniprot: the universal protein knowledgebase in 2025.Nucleic Acids Research, 53(D1):D609–D617, 2025

    work page 2025

  2. [2]

    US DOE Joint Genome Institute: Hawkins Trevor 4 Branscomb Elbert 4 Predki Paul 4 Richardson Paul 4 Wenning Sarah 4 Slezak Tom 4 Doggett Nor- man 4 Cheng Jan-Fang 4 Olsen Anne 4 Lucas Su- san 4 Elkin Christopher 4 Uberbacher Edward 4 Frazier Marvin 4, RIKEN Genomic Sciences Cen- ter: Sakaki Yoshiyuki 9 Fujiyama Asao 9 Hattori Masahira 9 Yada Tetsushi 9 Toy...

    work page 2001

  3. [3]

    Local genetic effects on gene expression across 44 human tissues.BioRxiv, page 074450, 2016

    Francois Aguet, Andrew A Brown, Stephane E Cas- tel, Joe R Davis, Pejman Mohammadi, Ayellet V Segre, Zachary Zappala, Nathan S Abell, Laure Fre- sard, Eric R Gamazon, et al. Local genetic effects on gene expression across 44 human tissues.BioRxiv, page 074450, 2016

    work page 2016

  4. [4]

    Deeploc: prediction of protein subcellu- lar localization using deep learning.Bioinformatics, 33(21):3387–3395, 2017

    José Juan Almagro Armenteros, Casper Kaae Søn- derby, Søren Kaae Sønderby, Henrik Nielsen, and Ole Winther. Deeploc: prediction of protein subcellu- lar localization using deep learning.Bioinformatics, 33(21):3387–3395, 2017

    work page 2017

  5. [5]

    Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

    Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

    work page 2000

  6. [6]

    Multi-relational poincaré graph embeddings.Advances in neural information processing systems, 32, 2019

    Ivana Balazevic, Carl Allen, and Timothy Hospedales. Multi-relational poincaré graph embeddings.Advances in neural information processing systems, 32, 2019

    work page 2019

  7. [7]

    Translating embeddings for modeling multi- relational data.Advances in neural information processing systems, 26, 2013

    Antoine Bordes, Nicolas Usunier, Alberto Garcia- Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi- relational data.Advances in neural information processing systems, 26, 2013

    work page 2013

  8. [8]

    Eval- uation of input data modality choices on functional gene embeddings.NAR Genomics and Bioinformat- ics, 5(4):lqad095, 2023

    Felix Brechtmann, Thibault Bechtler, Shubhankar Londhe, ChristianMertes, andJulienGagneur. Eval- uation of input data modality choices on functional gene embeddings.NAR Genomics and Bioinformat- ics, 5(4):lqad095, 2023

    work page 2023

Show all 35 references
  1. [9]

    Genept: a simple but effective foundation model for genes and cells built from chatgpt.bioRxiv, pages 2023–10, 2024

    Yiqun Chen and James Zou. Genept: a simple but effective foundation model for genes and cells built from chatgpt.bioRxiv, pages 2023–10, 2024

    2023

  2. [10]

    scgpt: toward building a foundation model for single- cell multi-omics using generative ai.Nature Methods, 21(8):1470–1480, 2024

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. scgpt: toward building a foundation model for single- cell multi-omics using generative ai.Nature Methods, 21(8):1470–1480, 2024

    2024

  3. [11]

    Gene2vec: distributed representation of genes based on co-expression.BMC genomics, 20:7–15, 2019

    Jingcheng Du, Peilin Jia, Yulin Dai, Cui Tao, Zhong- ming Zhao, and Degui Zhi. Gene2vec: distributed representation of genes based on co-expression.BMC genomics, 20:7–15, 2019

    2019

  4. [12]

    Gene expression omnibus: Ncbi gene expression and hybridization array data repository.Nucleic acids research, 30(1):207–210, 2002

    Ron Edgar, Michael Domrachev, and Alex E Lash. Gene expression omnibus: Ncbi gene expression and hybridization array data repository.Nucleic acids research, 30(1):207–210, 2002

    2002

  5. [13]

    Prottrans: Toward understanding the language of life through self-supervised learning

    Ahmed Elnaggar, Michael Heinzinger, Christian Dal- lago, Ghalia Rehawi, Yu Wang, Llion Jones, Tom Gibbs, Tamas Feher, Christoph Angerer, Martin Steinegger, et al. Prottrans: Toward understanding the language of life through self-supervised learning. IEEE transactions on patter...

    2021

  6. [14]

    struc2vec: Learning node representations from structural identity.CoRR, 2017

    Daniel R Figueiredo, Leonardo Filipe Rodrigues Ribeiro, and Pedro HP Saverese. struc2vec: Learning node representations from structural identity.CoRR, 2017

    2017

  7. [15]

    Canonical correlation analysis: An overview with application to learning methods.Neu- ral computation, 16(12):2639–2664, 2004

    David R Hardoon, Sandor Szedmak, and John Shawe-Taylor. Canonical correlation analysis: An overview with application to learning methods.Neu- ral computation, 16(12):2639–2664, 2004

    2004

  8. [16]

    Does your model understand genes? a benchmark of gene prop- erties for biological and text models.arXiv preprint arXiv:2412.04075, 2024

    Yoav Kan-Tor, Michael Morris Danziger, Eden Zo- har, Matan Ninio, and Yishai Shimoni. Does your model understand genes? a benchmark of gene prop- erties for biological and text models.arXiv preprint arXiv:2412.04075, 2024

    2024 arXiv

  9. [17]

    The hugo gene nomenclature committee (hgnc).Human genet- ics, 109(6), 2001

    Sue Povey, Ruth Lovering, Elspeth Bruford, Mathew Wright, Michael Lush, and Hester Wain. The hugo gene nomenclature committee (hgnc).Human genet- ics, 109(6), 2001

    2001

  10. [18]

    Svcca: Singular vector canon- ical correlation analysis for deep learning dynamics 6 and interpretability.Advances in neural information processing systems, 30, 2017

    Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canon- ical correlation analysis for deep learning dynamics 6 and interpretability.Advances in neural information processing systems, 30, 2017

    2017

  11. [19]

    String v10: protein–protein interaction networks, integrated over the tree of life.Nucleic acids research, 43(D1):D447– D452, 2015

    Damian Szklarczyk, Andrea Franceschini, Stefan Wyder, Kristoffer Forslund, Davide Heller, Jaime Huerta-Cepas, Milan Simonovic, Alexander Roth, Alberto Santos, Kalliopi P Tsafou, et al. String v10: protein–protein interaction networks, integrated over the tree of life.Nucleic a...

    2015

  12. [20]

    Damian Szklarczyk, Annika L Gable, David Lyon, Alexander Junge, Stefan Wyder, Jaime Huerta- Cepas, Milan Simonovic, Nadezhda T Doncheva, John H Morris, Peer Bork, et al. String v11: protein–protein association networks with increased coverage, supporting functional discovery i...

    2019

  13. [21]

    Large language models streamline automated machine learning for clinical studies.Nature Communications, 15(1):1603, 2024

    Soroosh Tayebi Arasteh, Tianyu Han, Mahshad Lotfinia, Christiane Kuhl, Jakob Nikolas Kather, Daniel Truhn, and Sven Nebelung. Large language models streamline automated machine learning for clinical studies.Nature Communications, 15(1):1603, 2024

    2024

  14. [22]

    The human protein atlas,

    The Human Protein Atlas. The human protein atlas,

  15. [23]

    Transfer learning enables predic- tions in network biology.Nature, 618(7965):616–624, 2023

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. Transfer learning enables predic- tions in network biology.Nature, 618(7965):616–624, 2023

    2023

  16. [24]

    Defining a can- cer dependency map.Cell, 170(3):564–576, 2017

    Aviad Tsherniak, Francisca Vazquez, Phil G Mont- gomery, Barbara A Weir, Gregory Kryukov, Glenn S Cowley, Stanley Gill, William F Harrington, Sasha Pantel, John M Krill-Burger, et al. Defining a can- cer dependency map.Cell, 170(3):564–576, 2017

    2017

  17. [25]

    Tissue-based map of the human pro- teome.Science, 347(6220):1260419, 2015

    Mathias Uhlén, Linn Fagerberg, Björn M Hallström, Cecilia Lindskog, Per Oksvold, Adil Mardinoglu, Åsa Sivertsson, Caroline Kampf, Evelina Sjöstedt, Anna Asplund, et al. Tissue-based map of the human pro- teome.Science, 347(6220):1260419, 2015

    2015

  18. [26]

    Biogps and mygene

    Chunlei Wu, Ian MacLeod, and Andrew I Su. Biogps and mygene. info: organizing online, gene-centric information.Nucleic acids research, 41(D1):D561– D565, 2013

    2013

  19. [27]

    Prgefne: predicting disease-related genesbyfastnetworkembedding.Methods, 192:3–12, 2021

    Ju Xiang, Ning-Rui Zhang, Jia-Shuai Zhang, Xiao-Yi Lv, and Min Li. Prgefne: predicting disease-related genesbyfastnetworkembedding.Methods, 192:3–12, 2021

    2021

  20. [28]

    Know2bio: A comprehensive dual-view benchmark for evolving biomedical knowledge graphs.arXiv preprint arXiv:2310.03221, 2023

    Yijia Xiao, Dylan Steinecke, Alexander Russell Pel- letier, Yushi Bai, Peipei Ping, and Wei Wang. Know2bio: A comprehensive dual-view benchmark for evolving biomedical knowledge graphs.arXiv preprint arXiv:2310.03221, 2023

    2023 arXiv

  21. [29]

    Seq2seq fingerprint: An unsupervised deep molecular embedding for drug discovery

    Zheng Xu, Sheng Wang, Feiyun Zhu, and Junzhou Huang. Seq2seq fingerprint: An unsupervised deep molecular embedding for drug discovery. InPro- ceedings of the 8th ACM international conference on bioinformatics, computational biology, and health in- formatics, pages 285–294, 2017

    2017

  22. [30]

    Linkbert: Pretraining language models with docu- ment links.arXiv preprint arXiv:2203.15827, 2022

    MichihiroYasunaga, JureLeskovec, andPercyLiang. Linkbert: Pretraining language models with docu- ment links.arXiv preprint arXiv:2203.15827, 2022

    2022 arXiv

  23. [31]

    The ensembl rest api: Ensembl data for any language.Bioinformatics, 31(1):143–145, 2015

    Andrew Yates, Kathryn Beal, Stephen Keenan, William McLaren, Miguel Pignatelli, Graham RS Ritchie, Magali Ruffier, Kieron Taylor, Alessandro Vullo, and Paul Flicek. The ensembl rest api: Ensembl data for any language.Bioinformatics, 31(1):143–145, 2015

    2015

  24. [32]

    A comprehensive capability analysis of gpt-3 and gpt-3.5 series models, 2023

    Junjie Ye, Xuanting Chen, Nuo Xu, Can Zu, Zekai Shao, Shichun Liu, Yuhan Cui, Zeyang Zhou, Chao Gong, Yang Shen, Jie Zhou, Siming Chen, Tao Gui, Qi Zhang, and Xuanjing Huang. A comprehensive capability analysis of gpt-3 and gpt-3.5 series models, 2023

    2023

  25. [33]

    Graph embedding on biomedical networks: methods, applications and evaluations.Bioinformat- ics, 36(4):1241–1251, 2020

    Xiang Yue, Zhen Wang, Jingong Huang, Srinivasan Parthasarathy, Soheil Moosavinasab, Yungui Huang, Simon M Lin, Wen Zhang, Ping Zhang, and Huan Sun. Graph embedding on biomedical networks: methods, applications and evaluations.Bioinformat- ics, 36(4):1241–1251, 2020

    2020

  26. [34]

    geneformer- 12L-30M

    Jeffrey Zhong, Lechuan Li, Ruth Dannenfelser, and Vicky Yao. Benchmarking gene embeddings from sequence, expression, network, and text models for functional prediction tasks.bioRxiv, pages 2025–01, 2025. 7 Supplementary Methods 7.1 Supplementary Method 1 - Molecular Embedding ...

    2025

  27. [2024]

    Accessed: 2025-07-07

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.