Pith. sign in

REVIEW 2 major objections 4 minor 119 references

A 52.6B-token biology corpus more than doubles a fixed 16B model's score on a matched biological eval, with general language nearly intact.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 07:54 UTC pith:MIZPMXWI

load-bearing objection The corpus is a real resource worth building on, but the +0.114 corpus-attribution claim is weakened by an unmatched-compute ablation (54B vs ~106B tokens), so the paper needs revision rather than rejection. the 2 major comments →

arxiv 2607.08803 v2 pith:MIZPMXWI submitted 2026-07-09 q-bio.QM cs.AIcs.LG

TheBioCollection: Unified Pre-Training Scale LLM Corpus for Biology

Hyunjin Seo , Hyeon Hwang , Gyubok Lee , Jay Shin , Jimin Park , Taesoo Kim , Sanghoon Lee , Hongjoon Ahn
show 2 more authors
Sungjun Han Sangwon Jung
This is my paper
classification q-bio.QM cs.AIcs.LG
keywords TheBioCollectionBioLMpre-training corpusbiologyinstruction tuninggenomicssmall moleculesevaluation suite
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's claim is that carefully constructed data, not a new model architecture, can give a large language model broad biological ability. TheBioCollection turns scattered public databases—small molecules, proteins, genomic sequences, cells, and pathways—into a 52.6B-token corpus of structured text, enriches records with tool-computed biological properties, and includes instruction tasks such as protein-binding and DNA/RNA feature localization whose answers can be checked programmatically. Holding a single base LLM fixed, training on this corpus plus scientific and web text more than doubles its average score on a matched 18-task evaluation (0.223 to 0.499), with gains in every domain, while five general-language benchmarks fall by an average of only 0.9 points. In the controlled comparison, the corpus itself contributes +0.114 over annealing on the same scientific and web text alone. The same results show the corpus trailing text-only annealing on perturbation-response prediction (−0.126) and one protein-design task (−0.064), which the authors attribute to thin cell-domain coverage and to degenerate text-only outputs.

Core claim

On its own terms, the discovery is that a consolidated, enriched, instruction-bearing corpus transfers biological knowledge to a general LLM at pre-training scale. Each record pairs an entity's sequence or identifier with a natural-language description and computed properties; cross-domain records chain a protein to its pathway, its binders, and the cellular response they induce. The paired evaluation suite asks for recognition, generation, prediction, and two-hop reasoning across five domains. The authors report the average evaluation score rises from 0.223 to 0.499, and that the portion not explained by web-text annealing is +0.114.

What carries the argument

The central mechanism is a corpus-construction pipeline that renders biological databases in a shared language interface: tagged tokens keep SMILES, protein, DNA, and RNA sequences distinct from surrounding prose; standard computational tools add deterministic properties as explicit text; and source-derived instruction tasks provide supervision with programmatically checkable answers, such as recovering a masked binding site or locating a genomic feature in JSON. A matched evaluation suite then tests whether those representations transfer. The pipeline's work is to make structured biological knowledge available to a general language model as ordinary text, so that annealing can absorb it wit

Load-bearing premise

The central claim rests on the evaluation being a genuine, contamination-free measure of biological capability; because the evaluation is assembled from the same public benchmark families and source-derived task types used to build the corpus, the reported gains could overstate transfer if the overlap-removal checks missed near-duplicate examples.

What would settle it

Retrain the same base model on TheBioCollection with a stricter and independently implemented decontamination—removing any training record that shares a short sequence segment, entity identifier, or instruction–answer hash with the evaluation—and rerun TheBioCollection-Eval; if the +0.114 corpus effect over text-annealing shrinks materially, the headline gains were memorization rather than transferable biological knowledge.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A model trained with TheBioCollection reaches 0.499 overall on the matched suite versus 0.223 for the base, so biological capability can be added at pre-training scale without changing the architecture.
  • The controlled comparison isolates +0.114 to the corpus itself over scientific/web-text annealing, so the gains are not just a side effect of continued pretraining.
  • The five general-language benchmarks drop by an average of 0.9 points, suggesting biological and general ability need not trade off sharply.
  • Two-hop cross-domain questions improve from 0.313 to 0.507, indicating the corpus supports reasoning across molecule–protein–pathway boundaries, not just single-domain facts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same recipe of tagged tokens, tool-computed narratives, and verifiable instruction tasks could plausibly be carried to other structured scientific domains, such as materials chemistry or climate data, where heterogeneous databases are also locked in non-text formats.
  • Because the evaluation is drawn from the same task families as the training corpus, the +0.114 corpus effect should be re-checked on a pre-registered, independently curated suite to separate transferable knowledge from format memorization.
  • The reported weakness in perturbation-response prediction suggests cell-domain records are the current bottleneck; a direct extension would enrich single-cell and perturbation data and re-run the CRISPRi response task.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper introduces TheBioCollection, a 52.6B-token biological pre-training corpus built by consolidating public databases across small molecules, proteins, genomics, cells/pathways, enriching records with tool-computed feature narratives, and adding instruction tasks for protein binding and DNA/RNA localization. It also introduces TheBioCollection-Eval, a matched 1,650-example suite with 18 tasks, and trains the Gravity-16B-A3B base model on the corpus mixed with general scientific/web text. The headline result is that training on TheBioCollection more than doubles the overall biological score (0.223→0.499, Table 3), and a text-annealing-only control is used to attribute +0.114 to the corpus (Table 5), while general-language benchmarks drop by 0.9 points on average.

Significance. If the attribution is sound, this would be a useful contribution: a large, commercially usable biological corpus, a transparent construction pipeline, a matched evaluation suite, and per-task results. The ablation design (same base checkpoint and annealing recipe) is a good idea, and the decontamination details (entity separation, exact/subsequence/15-mer overlap) are reported explicitly. However, the central claim of corpus-specific benefit currently rests on an ablation with unequal token budgets and on an evaluation suite assembled from the same task families used to build the corpus; both need to be addressed before the corpus effect can be considered established.

major comments (2)
  1. [§3.4 / Table 5] The isolation claim is not yet established. Ours is trained on the 54B-token general text plus the 52.6B-token TheBioCollection, while Text-annealing only is trained on the general text alone; nothing in §3.1 or Table 5 states that the total number of tokens or steps was matched. The +0.114 gap therefore conflates the corpus's content with the effect of ~52.6B additional training tokens. Please report token/step budgets and add a matched-token control (e.g., text-annealing only with the general text replayed so total tokens equal those of Ours).
  2. [§2.4–2.5 / Tables 2, 11] The evaluation suite is partly circular. TheBioCollection-Eval is assembled from the same public instruction datasets (Mol-Instructions, SMolInstruct, etc.) and source-derived task families (protein binding, DNA/RNA localization) used to construct the corpus. The stated decontamination is entity-level (exact/subsequence/15-mer, sequence hash, coordinates), but it does not remove format/template leakage: the model may memorize the JSON schemas and instruction phrasings rather than acquire transferable biological knowledge. Please report results on independently constructed external benchmarks, or add a format-control ablation (e.g., train on corpus records with answers/targets randomized) to quantify how much of the gain is format familiarity.
minor comments (4)
  1. [§3.3] The overall score is reported as '0.223 to 0.449', but Table 3 and Table 5 give Ours = 0.499; please correct the typo.
  2. [Figure 1] 'Homo sapience' should be 'Homo sapiens'.
  3. [Appendix D] The per-task tables (Tables 7–10) report Base and Ours but not the text-annealing-only condition. Adding that column would let readers verify the nondegeneracy and perturbation-response claims in §3.4.
  4. [Availability] Please state explicitly where the corpus and evaluation suite will be released and under what licenses, since the paper claims commercial usability and open release.

Circularity Check

0 steps flagged

No demonstrated circularity; the main evaluation/training overlap is addressed by entity-level decontamination, and the ablation token-count mismatch is a validity concern rather than a circular step.

full rationale

The paper's claimed derivation chain is an empirical training/evaluation loop, not a formal derivation, and no step reduces to its inputs by construction. TheBioCollection-Eval is assembled from benchmarks and source-derived task families that also contribute to the corpus (Sections 2.4 and 2.5, Tables 2 and 11), but the paper explicitly performs entity-level decontamination: 'For any task whose format also appears in training, we separate train and test entities to prevent overlap' and removes exact, subsequence, or shared 15-mer overlap for binder targets. This makes the eval a held-out generalization test within the same task distribution rather than a tautological restatement of the training data; whether the decontamination is sufficient is a leakage/validity question, not circularity. The TA-only vs Ours ablation is not matched in total training tokens: Ours adds the 52.6B-token corpus 'on top of that same general text,' so the +0.114 attribution is confounded by data volume and compute. That is a real correctness risk, but it is not an instance of the enumerated circularity patterns. The only clear self-citation is a minor reference to the authors' earlier VibeProteinBench in a list of scattered resources, and the citation to Trillion Labs for the base model is an artifact reference rather than a load-bearing uniqueness claim. No circular step within the seven enumerated kinds is demonstrated.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim is empirical, not derivational, so the ledger records the hand-chosen thresholds and domain assumptions that the corpus/eval construction rests on. No invented physical entities are introduced. The most load-bearing choices are the contact cutoff and decontamination thresholds, since they define the training and evaluation labels.

free parameters (4)
  • Heavy-atom contact cutoff for binding interfaces = 5.0 Å
    Used to define binding-site residues in protein-binding instruction construction (Appendix C); determines the labels for all binding tasks and the eval.
  • Eval sampling size per subtask = 100 single-domain; 50 cross-domain
    Chosen by hand in §2.5; small samples make aggregated scores noisy and limit statistical resolution.
  • Training recipe hyperparameters = seq len 8,192; batch 8.4M tokens; LR 4e-4; weight decay 0.01
    Chosen by hand in §3.1; held across compared models, but the annealing mixture and corpus/text ratio are not fully quantified.
  • Decontamination thresholds = exact string; subsequence; shared 15-mer overlap
    Chosen thresholds in §2.5 determine whether train/eval leakage is removed; the central validity check depends on these choices.
axioms (3)
  • domain assumption Public bio databases and tool outputs (RDKit, DSSP, ESMFold, etc.) are correct enough to serve as training and evaluation labels.
    The entire corpus and eval treat database annotations and tool-computed features as ground truth; systematic errors in these sources would propagate into all measured gains.
  • domain assumption Base Gravity-16B-A3B's pretraining contains no biological corpus, so observed biological gains are attributable to TheBioCollection training.
    §3.1 states this as the reason for choosing the base model, but it is an external claim about the base checkpoint that the paper does not independently verify.
  • domain assumption The decontamination procedure fully prevents train/eval overlap, so gains on TheBioCollection-Eval are not memorization.
    §2.5 describes exact/subsequence/15-mer overlap removal and entity separation, but effectiveness is assumed rather than demonstrated by leakage probes.

pith-pipeline@v1.3.0-alltime-deepseek · 27586 in / 11623 out tokens · 110988 ms · 2026-08-02T07:54:02.315740+00:00 · methodology

0 comments
read the original abstract

The push toward large language models for biology (BioLM) has created a need for training corpora that can endow models with a genuine understanding of biology. However, existing biological resources, such as molecular databases, protein repositories, genomic annotations, single-cell atlases, and pathway databases, are scattered across heterogeneous formats and remain unorganized into a cohesive corpus for language model training. We present TheBioCollection, a 52.6B-token pre-training-scale corpus that converts these disparate resources into a unified, training-ready form spanning small molecules, proteins, genomic sequences, cells, and pathways. Beyond consolidating existing data, TheBioCollection enriches each record with tool-computed biological properties and introduces new instruction tasks for capabilities that current corpora barely cover. We pair the corpus with TheBioCollection-Eval, a matched suite probing recognition, generation, and prediction across molecular, protein, genomic, cellular, and cross-domain settings. Holding the base Gravity-16B-A3B architecture fixed, training on TheBioCollection more than doubles its overall score on TheBioCollection-Eval with gains in every domain, while leaving general linguistic ability nearly intact.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

119 extracted references · 13 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2412.07775 , year=

    Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets , author=. arXiv preprint arXiv:2412.07775 , year=

    Pith/arXiv arXiv

  2. [2]

    Graph diffusion transformers for multi-conditional molecular generation , author=

  3. [3]

    Nature Biotechnology , volume=

    Protein structure prediction with in-cell photo-crosslinking mass spectrometry and deep learning , author=. Nature Biotechnology , volume=

  4. [4]

    arXiv preprint arXiv:2605.10978 , year=

    VibeProteinBench: An Evaluation Benchmark for Language-interfaced Vibe Protein Design , author=. arXiv preprint arXiv:2605.10978 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2502.07527 , year=

    Nature language model: deciphering the language of nature for scientific discovery , author=. arXiv preprint arXiv:2502.07527 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2509.21320 , year=

    SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines , author=. arXiv preprint arXiv:2509.21320 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2503.06687 , year=

    UniGenX: a unified generative foundation model that couples sequence, structure and function to accelerate scientific design across proteins, molecules and materials , author=. arXiv preprint arXiv:2503.06687 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2504.06196 , year=

    Txgemma: Efficient and agentic llms for therapeutics , author=. arXiv preprint arXiv:2504.06196 , year=

    Pith/arXiv arXiv

  9. [9]

    bioRxiv , year=

    BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning , author=. bioRxiv , year=

  10. [10]

    arXiv preprint arXiv:2306.08018 , year=

    Mol-instructions: A large-scale biomolecular instruction dataset for large language models , author=. arXiv preprint arXiv:2306.08018 , year=

    Pith/arXiv arXiv

  11. [11]

    MolLangBench: A Comprehensive Benchmark for Language-Prompted Molecular Structure Recognition, Editing, and Generation , author=

  12. [12]

    Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute , author=

  13. [13]

    Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations , author=

  14. [14]

    Findings of the Association for Computational Linguistics , year=

    Biot5+: Towards generalized biological understanding with iupac integration and multi-task tuning , author=. Findings of the Association for Computational Linguistics , year=

  15. [15]

    arXiv preprint arXiv:2406.06316 , year=

    Tx-llm: A large language model for therapeutics , author=. arXiv preprint arXiv:2406.06316 , year=

    Pith/arXiv arXiv

  16. [16]

    Science , volume=

    Evolutionary-scale prediction of atomic-level protein structure with a language model , author=. Science , volume=

  17. [17]

    Nature Methods , volume=

    Nucleotide transformer: building and evaluating robust foundation models for human genomics , author=. Nature Methods , volume=

  18. [18]

    DNABERT-2: Efficient foundation model and benchmark for multi-species genomes , author=

  19. [19]

    Nature , volume=

    Transfer learning enables predictions in network biology , author=. Nature , volume=

  20. [20]

    Nature methods , volume=

    scGPT: toward building a foundation model for single-cell multi-omics using generative AI , author=. Nature methods , volume=

  21. [21]

    Science , volume=

    Sequence modeling and design from molecular to genome scale with Evo , author=. Science , volume=

  22. [22]

    Nature , volume=

    Genome modelling and design across all domains of life with Evo 2 , author=. Nature , volume=

  23. [23]

    Evaluating protein transfer learning with TAPE , author=

  24. [24]

    Peer: a comprehensive and multi-task benchmark for protein sequence understanding , author=

  25. [25]

    Proteingym: Large-scale benchmarks for protein fitness prediction and design , author=

  26. [26]

    arXiv preprint arXiv:2311.12570 , year=

    Bend: Benchmarking dna language models on biologically meaningful tasks , author=. arXiv preprint arXiv:2311.12570 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2102.09548 , year=

    Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development , author=. arXiv preprint arXiv:2102.09548 , year=

    Pith/arXiv arXiv

  28. [28]

    Nucleic Acids Research , volume =

    PubChem 2025 update , author =. Nucleic Acids Research , volume =

    2025

  29. [29]

    Nucleic Acids Research , volume =

    BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data , author =. Nucleic Acids Research , volume =

    2024

  30. [30]

    Nucleic Acids Research , volume =

    The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods , author =. Nucleic Acids Research , volume =

    2023

  31. [31]

    Nucleic Acids Research , volume =

    UniProt: the Universal Protein Knowledgebase in 2025 , author =. Nucleic Acids Research , volume =

    2025

  32. [32]

    Nucleic Acids Research , volume =

    The Protein Data Bank , author =. Nucleic Acids Research , volume =

  33. [33]

    Nature , volume =

    Expanded encyclopaedias of DNA elements in the human and mouse genomes , author =. Nature , volume =

  34. [34]

    Nucleic Acids Research , volume =

    GENCODE 2021 , author =. Nucleic Acids Research , volume =

    2021

  35. [35]

    Nucleic Acids Research , volume =

    RNAcentral 2021: secondary structure integration, improved sequence search and new member databases , author =. Nucleic Acids Research , volume =

    2021

  36. [36]

    Nucleic Acids Research , volume =

    Rfam 14: expanded coverage of metagenomic, viral and microRNA families , author =. Nucleic Acids Research , volume =

  37. [37]

    arXiv preprint arXiv:2402.09391 , year=

    Llasmol: Advancing large language models for chemistry with a large-scale, comprehensive, high-quality instruction tuning dataset , author=. arXiv preprint arXiv:2402.09391 , year=

    Pith/arXiv arXiv

  38. [38]

    Cell2Sentence: Teaching Large Language Models the Language of Biology , author =

  39. [39]

    Science , volume =

    The Tabula Sapiens: A multiple-organ, single-cell transcriptomic atlas of humans , author =. Science , volume =

  40. [40]

    Perturbench: Benchmarking machine learning models for cellular perturbation analysis , author=

  41. [41]

    Nature Methods , volume =

    Human BioMolecular Atlas Program (HuBMAP): 3D Human Reference Atlas construction and usage , author =. Nature Methods , volume =

  42. [42]

    Nature Methods , volume =

    Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations , author =. Nature Methods , volume =

  43. [43]

    Genetics , volume =

    The Gene Ontology knowledgebase in 2023 , author =. Genetics , volume =

    2023

  44. [44]

    Nucleic Acids Research , volume =

    The Reactome Pathway Knowledgebase 2024 , author =. Nucleic Acids Research , volume =

    2024

  45. [45]

    eLife , volume =

    Systematic integration of biomedical knowledge prioritizes drugs for repurposing , author =. eLife , volume =

  46. [46]

    2020 , howpublished =

    Drug Repurposing Knowledge Graph (DRKG) , author =. 2020 , howpublished =

    2020

  47. [47]

    Nucleic Acids Research , volume =

    The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest , author =. Nucleic Acids Research , volume =

    2023

  48. [48]

    Genome Research , volume =

    Benchmark and integration of resources for the estimation of human transcription factor activities , author =. Genome Research , volume =

  49. [49]

    Tissue-based map of the human proteome , author =

    Proteomics. Tissue-based map of the human proteome , author =. Science , volume =

  50. [50]

    2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , year =

    A Fine-tuning Dataset and Benchmark for Large Language Models for Protein Understanding , author =. 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , year =

    2024

  51. [51]

    bioRxiv , pages=

    BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning , author=. bioRxiv , pages=

  52. [52]

    Biopolymers , volume=

    Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features , author=. Biopolymers , volume=

  53. [53]

    Nucleic Acids Research , volume =

    AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage , author =. Nucleic Acids Research , volume =

    2025

  54. [54]

    Nucleic Acids Research , volume =

    DrugBank 6.0: the DrugBank Knowledgebase for 2024 , author =. Nucleic Acids Research , volume =

    2024

  55. [55]

    Cell , volume =

    A Next Generation Connectivity Map: L1000 Platform and the First 1,000,000 Profiles , author =. Cell , volume =

  56. [56]

    Nucleic Acids Research , volume =

  57. [57]

    BioRxiv , year=

    PLINDER: The protein-ligand interactions dataset and evaluation resource , author=. BioRxiv , year=

  58. [58]

    bioRxiv , year=

    PINDER: The protein interaction dataset and evaluation resource , author=. bioRxiv , year=

  59. [59]

    Learning to Design Protein-Protein Interactions with Enhanced Generalization , author =

  60. [60]

    Morehead, Alex and Chen, Chen and Sedova, Ada and Cheng, Jianlin , journal =

  61. [61]

    Zhu, Ning and Ming, Yanyu and Zhang, Chengyun and Sen, Cao and Li, Chongyang and Guo, Jingjing and Duan, Hongliang , journal =

  62. [62]

    Frontiers in Bioinformatics , volume =

    Propedia v2.3: A novel representation approach for the peptide-protein interaction database using graph-based structural signatures , author =. Frontiers in Bioinformatics , volume =

  63. [63]

    Cell , volume =

    Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq , author =. Cell , volume =

  64. [64]

    Science , volume =

    Exploring genetic interaction manifolds constructed from rich single-cell phenotypes , author =. Science , volume =

  65. [65]

    and others , journal =

    Ovek Baydar, Damla and Rauluseviciute, Ieva and Aronsen, Dina R. and others , journal =

  66. [66]

    Kozomara, Ana and Birgaoanu, Maria and Griffiths-Jones, Sam , journal =

  67. [67]

    and Lowe, Todd M

    Chan, Patricia P. and Lowe, Todd M. , journal =

  68. [68]

    The flan collection: Designing data and methods for effective instruction tuning , author=

  69. [69]

    bioRxiv , year=

    Protenix-v1: Toward high-accuracy open-source biomolecular structure prediction , author=. bioRxiv , year=

  70. [70]

    doi:10.5281/zenodo.18428170 , url =

    Landrum, Greg and Tosco, Paolo and Kelley, Brian and Rodriguez, Ricardo and Cosgrove, David and Vianello, Riccardo and Gedeck, Peter and others , year =. doi:10.5281/zenodo.18428170 , url =

    doi:10.5281/zenodo.18428170

  71. [71]

    National science review , volume=

    Foundation models in bioinformatics , author=. National science review , volume=. 2025 , publisher=

    2025

  72. [72]

    Nature Machine Intelligence , volume=

    Generalized biological foundation model with unified nucleic acid and protein language , author=. Nature Machine Intelligence , volume=. 2025 , publisher=

    2025

  73. [73]

    Nature Biotechnology , pages=

    Generalist biological artificial intelligence in modeling the language of life , author=. Nature Biotechnology , pages=. 2026 , publisher=

    2026

  74. [74]

    Nature Methods , volume=

    Language models for biological research: a primer , author=. Nature Methods , volume=. 2024 , publisher=

    2024

  75. [75]

    Communications Biology , volume=

    SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders , author=. Communications Biology , volume=. 2023 , publisher=

    2023

  76. [76]

    Nature methods , volume=

    Large-scale foundation model on single-cell transcriptomics , author=. Nature methods , volume=. 2024 , publisher=

    2024

  77. [77]

    Nature Biotechnology , volume=

    Predicting transcriptional outcomes of novel multigene perturbations with GEARS , author=. Nature Biotechnology , volume=. 2024 , publisher=

    2024

  78. [78]

    Scientific data , volume=

    Building a knowledge graph to enable precision medicine , author=. Scientific data , volume=. 2023 , publisher=

    2023

  79. [79]

    Nature methods , volume=

    OmniPath: guidelines and gateway for literature-curated signaling pathway resources , author=. Nature methods , volume=. 2016 , publisher=

    2016

  80. [80]

    Saprot: Protein language modeling with structure-aware vocabulary , author=

Showing first 80 references.