Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Find Central Dogma Again: Leveraging Multilingual Transfer in Large Language Models

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

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

Pith's one-line read Sentence-pair training lets GPT-2 rediscover the genetic code

desk verdict A creative cross-modal transfer experiment whose headline 81% is an artifact of best-of-100 seed selection and a test set that rewards shallow sequence statistics. read the letter →

arxiv 2502.06253 v3 pith:DZAJQS75 submitted 2025-02-10 q-bio.GN

Wang Liang This is my paper · ORCID
classification q-bio.GN
keywords centraldogmageneticcodemultilingualtransferzero-shotlearningDNA-proteinalignmentGPT-2PAWS-X
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

The paper sets out to show that a large language model can rediscover the central dogma of molecular biology—the rule that DNA sequences encode protein sequences—without ever being taught the rule. Its experiment trains a GPT-2 model from scratch on DNA, protein, and multilingual Wikipedia text, fine-tunes it on the PAWS-X sentence-paraphrase judgment task, and then tests it on a DNA-protein alignment classification task. The best fine-tuned model reaches 81% test accuracy on the random-negative version of that task, which the authors take as evidence that natural-language transfer produced a zero-shot grasp of genetic coding rules. The result is framed as a bridge between natural language and 'genetic language,' suggesting that structurally similar language tasks can transfer to biological sequence analysis.

What carries the argument

The mechanism is cross-modal language transfer mediated by a shared Byte Pair Encoding tokenizer and a structurally analogous fine-tuning task. PAWS-X, a cross-lingual sentence-paraphrase dataset, is chosen because deciding whether two sentences mean the same thing is formally similar to deciding whether a DNA sequence and a protein sequence encode the same information. The model learns the sentence1-to-sentence2 mapping during fine-tuning, and the paper argues this mapping transfers to the DNA-to-protein alignment task.

What would settle it

Build a test set where negative protein sequences are generated by shuffling or mutating codons of true positives so that overall amino-acid composition and sequence length are nearly unchanged; if the fine-tuned model still scores near 81%, it has not learned the codon mapping.

Watch

Extended reading notes

Core claim

The paper claims that a GPT-2 model pretrained on sequences from seven natural languages plus DNA and protein, then fine-tuned on Chinese-Korean paraphrase pairs from PAWS-X, can classify whether a DNA sequence and a protein sequence correspond, reaching 81% accuracy on a test set where negatives are random proteins of low similarity to the true translation. The authors interpret this as the model rediscovering the genetic code through language transfer, without prior knowledge of the central dogma. They further report that pretraining on more languages improves transfer, that fine-tuning on structurally similar natural-language tasks helps, and that the effect is highly sensitive to random seeds during both pretraining and fine-tuning.

Load-bearing premise

The claim rests on the assumption that the DNA-protein test set actually measures knowledge of the genetic code; since the test uses only cleanly matching positives and random, low-similarity proteins as negatives, a classifier could score highly by detecting simpler statistical differences.

Editorial extensions

If this is right

  • Multilingual pretraining improves zero-shot transfer to genomic tasks beyond English-only or biology-only pretraining.
  • Fine-tuning on a structurally similar natural-language dataset can substitute for supervised training on the genomic task itself.
  • Pre-trained models without fine-tuning already show above-random DNA-protein alignment accuracy, rising with the number of languages seen.
  • The transfer is fragile: only a small fraction of random seeds find a good solution, so multiple runs or ensembling are needed to obtain the reported accuracy.

Reading between the lines

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

  • A probe that tests whether the model's representations encode individual codons, rather than pair-level statistics, would directly test whether 'rediscovery of the central dogma' means codon-level knowledge.
  • Using negative examples that preserve amino-acid composition and length but shuffle codons would separate genuine codon learning from simpler distributional cues.
  • Averaging fine-tuned models across seeds, rather than selecting the best of 100, could turn the fragile transfer into a reliable method and would test whether the rare-seed behavior is real learning rather than luck.
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 / 4 minor

Summary. The paper trains GPT-2 from scratch on a mixture of DNA sequences, protein sequences, and multilingual Wikipedia text, then fine-tunes it on the PAWS-X paraphrase-similarity dataset and evaluates it on a binary DNA-protein alignment task. The authors report a best accuracy of 81% on their "dna_protein_pair_rand" test set and interpret this as zero-shot rediscovery of the genetic code via natural-language transfer. Additional experiments examine seed sensitivity, the choice of fine-tuning language, and direct prediction with the pretrained model without fine-tuning.

Significance. If the central claim were valid, this would be a striking demonstration of cross-modal transfer from natural language to the genetic code and would merit broad attention. The authors deserve credit for releasing code, models, and data, and for including a candid stability analysis in Section 3.3 that quantifies large seed-to-seed variability. However, the headline 81% is the maximum of 100 fine-tuning runs, the test set is constructed in a way that allows shallow statistical shortcuts, and no trivial feature-based baselines are reported. The current evidence therefore does not support the claim that the model has rediscovered the central dogma.

major comments (4)
  1. [§2.3 / Supplementary §1.3] The test set is simplified in a way that makes the binary classes separable without learning the codon-to-amino-acid mapping. The construction removes the 100-bp flanks, keeps only the high-confidence ORF positives (approximately 3,000 of 8,533), and uses random proteins with similarity below 0.45 as negatives. Under these conditions, global sequence statistics such as GC content, codon-usage bias, and amino-acid composition differ systematically between matched and random pairs. The paper reports no baseline using such features (e.g., DNA k-mer frequencies, amino-acid composition, or alignment scores), so the 81% accuracy does not demonstrate knowledge of the genetic code.
  2. [§3.3 (Best finetune Language), Table 5, Fig. 4] The headline 81% is the maximum accuracy over 100 fine-tuning runs with different random seeds, and Fig. 4B shows that only about 9% of seeds exceed 70% on dna_protein_pair_rand, with most runs near chance. Thus the reported number is an extreme order statistic, not a typical or reproducible transfer outcome. The paper should report the full distribution (mean, median, quartiles) and either evaluate a fixed model or select hyperparameters on a held-out validation set rather than on the test set.
  3. [§2.4 / §3.4, Table 7] The label-swap rule in §2.4 uses test labels to choose the orientation for each run whenever accuracy is below 0.5, and §3.4 states that direct-prediction results "take the best result" over random seeds. Combined with best-of-100 selection, this makes the reported accuracy a fit to the test set rather than an unbiased estimate of model performance. The authors should separate model selection from evaluation and report the seed distribution without label reorientation based on test performance.
  4. [§4 / Abstract] The claim that the model "rediscovered the central dogma" is not established by the presented evidence. The pretraining corpus includes DNA and protein sequences, so the model could exploit statistical associations such as organism-specific codon usage or amino-acid composition without learning which codon encodes which amino acid. To support the rediscovery claim, the paper would need a control that removes composition cues, a comparison to nontrivial alignment-based baselines, or a direct probe of the learned codon-amino-acid mapping.
minor comments (4)
  1. [§3.3] The section number 3.3 is used twice, for "Training Stability" and for "Best finetune Language"; renumber to avoid confusion.
  2. [Table 5] The row labeled "ja-ka" appears to be a typo; it should probably read "ja-ja" or "ja-ko."
  3. [Fig. 2 caption] The caption says the layer maps 768 hidden units to 115,000 units as the token vocabulary size, but the GPT-2 small tokenizer has 50,257 tokens; please clarify the actual vocabulary size.
  4. [Fig. 3 / Table 3 / Table 5] The Fig. 3 caption reports approximately 76% accuracy for gpt2-gene-mulit on dna_protein_pair_rand, while Table 3 reports 0.76 and Table 5 reports 0.77 for the en-en fine-tuning setting; reconcile these numbers or clarify which run is shown.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline 81% is the best of 100 test-set-selected fine-tuning runs with post-hoc label swapping, so the reported 'zero-shot' prediction is partly fitted to the test set.

  1. fitted input called prediction [Section 3.3 (Best finetune Language), Table 5 caption and Table 6]
    "Since fine-tuning performance is highly sensitive to random seed selection, we conducted 100 training runs for each dataset, with random seeds ranging from 1 to 100, and then selected the best-performing model for comparison (Table.5)."

    The 81% reported in the abstract and conclusion is the maximum test accuracy over 100 independently seed-initialized fine-tuning runs, and also over the language-pair grid summarized in Table 6. Selecting the run with the highest test accuracy makes the test labels part of the model-selection procedure; the reported number is an extreme order statistic, not the accuracy of a pre-specified model. The paper's own Table 4 and Figure 4 show that only 9% of seeds exceed 70% on dna_protein_pair_rand, so 0.81 is a rare upper-tail draw. Calling this a zero-shot transfer success is equivalent to fitting a seed hyperparameter to the test set and re-labeling the fit as a prediction.

  2. fitted input called prediction [Section 2.4 Model Evaluation]
    "Since this is a binary classification task, the meaning of classification IDs 0 and 1 may differ during transfer. Therefore, if the test accuracy is below 0.5, we can simply swap label IDs 0 and 1."

    The label orientation is chosen after observing the test accuracy, using test labels to decide whether 0 or 1 denotes alignment. For each seed this yields max(acc, 1-acc), which guarantees at least chance-level performance and inflates a random run to an expected 0.75 even with no real signal. The direction of the label mapping is therefore fitted to the same labels that are later counted as correct predictions, so the reported accuracy is partly constructed from the test labels rather than predicted from the model alone.

full rationale

The central derivation is not circular in the sense of an equation reducing to its own input or a result borrowed from a self-citation: the paper trains and evaluates its own models, and reference [24] is background motivation rather than the evidence for the 81% claim. However, the key quantitative claim fails the 'fitted input called prediction' test. The 81% figure is selected as the best of 100 fine-tuning seeds, and also as the best of the fine-tuning language pairs in Table 6, while the binary-label orientation is decided by swapping labels whenever test accuracy is below 0.5. Both choices use the test labels to select the reported model and its mapping, so the headline number is an optimized test-set statistic rather than the accuracy of a fixed zero-shot model. Additional concerns about the test construction, such as using only the clearest positive ORFs and random-protein negatives with similarity below 0.45, weaken the inference that the model learned the genetic code, but those are experimental confounds rather than circular reductions. Overall, the paper has independent empirical content, but its marquee 'rediscover the central dogma' result is substantially fitted to the test set; this is partial circularity, not a full self-referential derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new scientific entities. The central claim rests on several assumptions: a structural analogy between PAWS-X pairs and DNA-protein pairs, the adequacy of random proteins as negatives, the sufficiency of BPE tokenization, and the validity of selecting the best random seed for reporting. None of these are independently verified, and the free parameters are the seed and the test-set construction choices.

free parameters (4)
  • fine-tuning random seed (best of 100) = not reported (selected to maximize test accuracy)
    The model fine-tuned with PAWS-X pairs was trained 100 times with seeds 1-100, and the best-performing model on the DNA-protein test set was selected (Section 3.3). The reported 81% is the maximum of this distribution, not a typical run.
  • negative protein similarity threshold = 0.45
    Negative examples in dna_protein_pair_rand were created by sampling a random protein whose similarity to the correct protein was below 0.45 (Section 2.3), making negatives easy to distinguish.
  • positive example filtering = approx. 3,000 of 8,533 original positives
    Only the most clearly matching DNA-protein pairs were retained as positives after ORF filtering (Supplementary 1.3), reducing test difficulty.
  • number of pre-training epochs = 3 to 5
    The model was trained for 3-5 epochs (Section 2.4); the exact epoch per reported model is not specified.
assumptions (5)
  • domain assumption PAWS-X sentence-pair similarity judgment is structurally analogous to DNA-protein alignment.
    The whole transfer design depends on this analogy (Section 2.2), but no empirical argument is given beyond surface similarity of 'sequence pair classification'.
  • domain assumption BPE tokenization of DNA and protein sequences preserves the information needed for alignment.
    The authors state tokenizer choice does not affect validation (Section 2.4), but this is an unverified assumption.
  • ad hoc to paper The model can determine the boundary between concatenated DNA and protein sequences without a separator.
    The input is 'sequence1+sequence2' with no separator (Section 2.4), relying on the model's implicit segmentation ability, which is not demonstrated.
  • ad hoc to paper A random protein with similarity <0.45 is a valid negative example for a DNA-protein alignment task.
    This construction removes hard negatives and may allow classification via global sequence statistics instead of codon-level mapping (Section 2.3).
  • domain assumption The central dogma can be faithfully represented as binary DNA-protein pair matching.
    The paper reduces a rich mapping to a binary classification, and labels any successful classifier as having 'rediscovered' the central dogma (Section 2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Find Central Dogma Again: Leveraging Multilingual Transfer in Large Language Models." pith.science (2026). https://pith.science/paper/DZAJQS75

@misc{pith2026250206253,
  author       = {Pith},
  title        = {Pith review of: Find Central Dogma Again: Leveraging Multilingual Transfer in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZAJQS75}},
  note         = {Machine review of arXiv:2502.06253}
}
read the original abstract

In recent years, large language models (LLMs) have achieved state-of-the-art results in various biological sequence analysis tasks, such as sequence classification, structure prediction, and function prediction. Similar to advancements in AI for other scientific fields, deeper research into biological LLMs has begun to focus on using these models to rediscover important existing biological laws or uncover entirely new patterns in biological sequences. This study leverages GPT-like LLMs to utilize language transfer capabilities to rediscover the genetic code rules of the central dogma. In our experimental design, we transformed the central dogma into a binary classification problem of aligning DNA sequences with protein sequences, where positive examples are matching DNA and protein sequences, and negative examples are non-matching pairs. We first trained a GPT-2 model from scratch using a dataset comprising protein sequences, DNA sequences, and sequences from languages such as English and Chinese. Subsequently, we fine-tuned the model using the natural language sentences similarity judgment dataset from PAWS-X. When tested on a dataset for DNA and protein sequence alignment judgment, the fine-tuned model achieved a classification accuracy of 81%. The study also analyzed factors contributing to this zero-shot capability, including model training stability and types of training data. This research demonstrates that LLMs can, through the transfer of natural language capabilities and solely relying on the analysis of sequences themselves, rediscover the central dogma without prior knowledge of it. This study bridges natural language and genetic language, opening a new door for AI-driven biological research.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    MachineTranslation–TheDNA→proteinencodingprocessisanalogoustotranslating asourcelanguageintoatargetlanguage

    work page

  2. [2]

    Semantic Similarity Texts – Similar to machine translation, this includes sentence pairs with equivalent meanings, such as two semantically similar English sentences or an English sentencepairedwithitssemanticallyequivalentFrenchtranslation

    work page

  3. [3]

    Speech-to-Text – The process of encoding DNA into proteins resembles converting speechsignalsintorecognizedtext

    work page

  4. [4]

    sentence1

    Code Generation – DNA-to-protein conversion follows strict rule-based mappings, similartothetransformationbetweennaturallanguageandprogrammingcode. Given the availability and diversity of datasets, we chose semantic similarity text datasets as candidates for language transfer learning. One of the most representative datasets in this category is PAWS-X, a ...

    work page

  5. [5]

    Multimodal Data Training During the pretraining phase, using a diverse set of training data—including multiple natural languages (not just English) as well as DNA andprotein sequences—enablesthe model to learna more comprehensive representation across different languages and symbolic systems. This approach allows the model not only to grasp English gramma...

    work page

  6. [6]

    For example, Byte Pair Encoding (BPE) can segment both English text and DNA/protein sequences into subwords or sub-sequences, forming a shared vocabulary or encoding space

    Unified Tokenization and Encoding Methods (e.g., BPE) Using a unified tokenization (or encoding) method ensures consistencywhen converting different types of data (natural language and genetic sequences) into model inputs. For example, Byte Pair Encoding (BPE) can segment both English text and DNA/protein sequences into subwords or sub-sequences, forming ...

    work page

  7. [7]

    accidentalalignmentofthinkingpatterns

    Fine-tuning on Natural Language Datasets with Structural Similarity to Genetic Sequence Tasks The structure of the dataset used for fine-tuning is critical for transfer effectiveness. Genetic sequence tasks often follow fixed patterns, such as DNA-protein alignment problems. Selecting a natural language dataset with a similar structure for fine-tuning all...

    work page

  8. [8]

    However, such solutions are extremely scarce in the optimization space andonlyappearunderafewspecificrandomseeds

    Cross-domain transfer is highly sensitive: Transferring a model between English and biological sequences requires finding those rare solutions that can accommodate both language structures. However, such solutions are extremely scarce in the optimization space andonlyappearunderafewspecificrandomseeds

    work page

Show all 38 references
  1. [9]

    crack the code

    The alignment of thinking patterns is accidental: The model's English-based reasoning onlyproducesgoodtransferresultswhenit"happens"tooverlapwiththepatternsofDNAor protein sequences under certain conditions. This suggests that cross-domain transfer may require additional regul...

  2. [10]

     Training agents to autonomously search and optimize sequences to discover newfunctionalgenesorproteinstructures

    CombiningPromptEngineeringandAgentsforBiologicalResearch  Utilizing Prompt Engineering and Agents to enable large models to actively explorehiddenpatternsingeneticsequences.Forexample:  Designing structured prompts for DNA or protein sequences to predict potentialfunctionalr...

  3. [11]

     Utilizing cross-modal AI to integrate imaging (e.g., protein structures), text, and sequence data for more precise disease prediction and personalized medicine analysis

    ExpandingAIApplicationsinBiologyThroughCross-ModalLearning  Integrating biological text data (e.g., research papers, patents) with experimental datatoenhancethemodel’sreasoningcapabilities.Forexample:  Enabling large models to automatically read biomedical literature and gen...

  4. [12]

    Zhou, et al., Dnabert-2: Efficient foundation model and benchmark for multi-species genome,arXivpreprintarXiv:2306.15006(2023)

    Z. Zhou, et al., Dnabert-2: Efficient foundation model and benchmark for multi-species genome,arXivpreprintarXiv:2306.15006(2023)

    2023 arXiv

  5. [13]

    Nguyen, et al., Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution,Advancesinneuralinformationprocessingsystems36(2024)

    E. Nguyen, et al., Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution,Advancesinneuralinformationprocessingsystems36(2024)

    2024

  6. [14]

    F.Yang,etal.,scbertasalarge-scalepretraineddeeplanguagemodelforcelltypeannotation ofsingle-cellrna-seqdata,NatureMachineIntelligence4,852–866(2022)

    2022

  7. [15]

    A. Elnaggar, et al., Prottrans: Toward understanding the language of life through selfsupervised learning, IEEE transactions on pattern analysis and machine intelligence 44, 7112–7127(2021)

    2021

  8. [16]

    N.Brandes,D.Ofer,Y.Peleg,N.Rappoport,M.Linial,Proteinbert:auniversaldeeplearning modelofproteinsequenceandfunction,Bioinformatics38,2102–2110(2022)

    2022

  9. [17]

    Lin, et al., Evolutionary-scale prediction of atomic-level protein structure with a language model,Science379,1123–1130(2023)

    Z. Lin, et al., Evolutionary-scale prediction of atomic-level protein structure with a language model,Science379,1123–1130(2023)

    2023

  10. [18]

    & Davuluri, R

    Ji, Y., Zhou, Z., Liu, H. & Davuluri, R. V. DNABERT: Pre-trained bidirectional encoder representations from transformers model for DNA-language in genome. Bioinformatics 37, 2112–2120(2021).doi:doi.org/10.1093/bioinformatics/btab083

    2021 doi

  11. [19]

    doi: https://doi.org/10.1101/2023.06.12.544594

    Veniamin Fishman, et al.GENA-LM: A Family of Open-Source Foundational DNA Language Models for Long Sequences.BioRxiv 2023.06.12.544594. doi: https://doi.org/10.1101/2023.06.12.544594

    2023 doi

  12. [20]

    Daoan Zhang, et al.DNAGPT: A Generalized Pretrained Tool for Multiple DNA Sequence AnalysisTasks.BioRxiv2023.07.11.548628;doi:https://doi.org/10.1101/2023.07.11.548628

    2023 doi

  13. [21]

    Wang, B. et al. Pre-trained Language Models in Biomedical Domain: A Systematic Survey. ACM Comput. Surv.56,1–52(2023).doi:doi.org/10.1145/3611651

    2023 doi

  14. [22]

    & Gao, G

    Chen, Z., Wei, L. & Gao, G. Foundation models for bioinformatics. Quant. Biol. (2024) doi:10.1002/qub2.69

    2024 doi

  15. [24]

    Nguyen E, Poli M, Durrant MG, Kang B, Katrekar D, Li DB, Bartie LJ, Thomas AW, King SH,BrixiG,SullivanJ.Sequencemodelinganddesignfrommoleculartogenomescalewith Evo.Science.2024Nov15;386(6723):eado9336

  16. [25]

    DNA language models are powerful predictors of genome-wide variant effects

    Benegas G, Batra SS, Song YS. DNA language models are powerful predictors of genome-wide variant effects. Proceedings of the National Academy of Sciences. 2023 Oct 31;120(44):e2311219120

    2023

  17. [26]

    ShaoB,YanJ.Along-context languagemodelfordecipheringandgenerating bacteriophage genomes.NatureCommunications.2024Oct30;15(1):9392

  18. [27]

    Language models enable zero-shot prediction of the effects of mutations on protein function

    Meier J, Rao R, Verkuil R, Liu J, Sercu T, Rives A. Language models enable zero-shot prediction of the effects of mutations on protein function. Advances in neural information processingsystems.2021Dec6;34:29287-303

  19. [28]

    DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA

    Patel A, Singhal A, Wang A, Pampari A, Kasowski M, Kundaje A. DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA. arXiv preprintarXiv:2412.05430.2024Dec6

  20. [29]

    Fei Yuan,etc.How Vocabulary Sharing Facilitates Multilingualism in LLaMA?arXiv (2024) doi:https://doi.org/10.48550/arXiv.2306.06688

    2024 doi arXiv

  21. [30]

    Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020. EmergingCross-lingualStructureinPretrainedLanguageModels.In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6022–6034, Online. AssociationforCompu...

    2020

  22. [31]

    Specialists: An Empirical RevisitingonMultilingualTransferAbility

    Jiacheng Ye, Xijia Tao, Lingpeng Kong.Language Versatilists vs. Specialists: An Empirical RevisitingonMultilingualTransferAbility. arXiv(2024)doi:10.48550/arxiv.2407.11435

    2024 doi arXiv

  23. [32]

    Hetong Wang, Pasquale Minervini, and Edoardo Ponti. 2024. Probing the Emergence of Cross-lingual Alignment during LLM Training. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12159–12173,Bangkok, Thailand. Association forComputationalLinguistics

    2024

  24. [33]

    Junjie Hu, etc.XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingualGeneralization.arXiv(2020)https://arxiv.org/abs/2003.11080

    2020 arXiv

  25. [34]

    Exploring the relationship between alignment and cross-lingual transfer in multilingual transformers

    Gaschi F, Cerda P, Rastin P, Toussaint Y. Exploring the relationship between alignment and cross-lingual transfer in multilingual transformers. Findings of the Association for ComputationalLinguistics:ACL2023

  26. [35]

    LiangW.CanlinguistsbetterunderstandDNA?.arXivpreprintarXiv:2412.07678.2024Dec 10

    arXiv

  27. [36]

    LucaOne: Generalized Biological Foundation Model with Unified Nucleic Acid and Protein Language

    Yong He, et al. LucaOne: Generalized Biological Foundation Model with Unified Nucleic Acid and Protein Language. BioRxiv 2024.05.10.592927; doi: https://doi.org/10.1101/2024.05.10.592927

    2024 doi

  28. [37]

    Zeyuan Wang, etc. 2024. InstructProtein: Aligning Human and Protein Language via Knowledge Instruction. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1114–1136, Bangkok, Thailand. AssociationforComputa...

    2024

  29. [38]

    paperhuggingface,https://huggingface.co/dnagpt/central_dogma Data availability section Sequence data that support the findings of this study have been deposited in huggingface.The fine-tuning/testdataislocatedathttps://huggingface.co/datasets/dnagpt/gene_lan_transfer,andthe pr...

  30. [39]

    A custom-constructed dataset, where protein sequences are randomly sampled from UniProtKB/Swiss-Prot, and their corresponding DNA sequences are retrieved. For non-aligned pairs, the DNA sequence remains unchanged, while the protein sequence is randomly sampled from the protein...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.