Pith. sign in

REVIEW 3 major objections 4 minor 45 references

Predicting and generating antibiotics against future pathogens with ApexOracle

T0 review · 3 major / 4 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 ApexOracle combines a diffusion language model with pathogen genome and literature embeddings to predict antibiotic potency and design new molecules for bacterial strains it has never encountered.

desk verdict Clever architecture and real open resources, but the 'unseen pathogen' claim leaks labels through LLM-generated text prompts, so the headline result needs re-earning. read the letter →

arxiv 2507.07862 v1 pith:XBEULTCU submitted 2025-07-10 cs.LG q-bio.QM

Tianang Leng , Fangping Wan , Marcelo Der Torossian Torres , Cesar de la Fuente-Nunez This is my paper
classification cs.LGq-bio.QM
keywords antimicrobialresistanceantibioticdiscoverydiffusionlanguagemodelminimuminhibitoryconcentrationpathogengenomeembeddingdenovomoleculegenerationzero-shottransfersynergyprediction
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

ApexOracle is a machine-learning platform that claims to do two things at once: predict how well any molecule—peptide or small-molecule antibiotic—will work against a bacterial or fungal strain, and design new molecules predicted to kill strains it has never encountered. To do this it fuses three representations: a diffusion-language-model embedding of the molecule, a full-genome embedding, and an embedding of a text description of the strain's traits. The paper reports that in held-out evaluations that hide entire strains, species groups, or broad taxonomic divisions, ApexOracle predicts minimum inhibitory concentrations substantially better than the alternatives, and that its predictor-guided generative module produces molecules with low predicted MIC against drug-resistant E. coli and P. aeruginosa strains on which it was never trained. If these claims hold, the model offers a route to 'future-pathogen' preparedness: estimating susceptibility and generating candidate therapies without waiting for strain-specific experimental data.

What carries the argument

The load-bearing mechanism is a cross-attention fusion of three embeddings: a molecule representation from a masked discrete diffusion language model (a Diffusion Transformer trained to denoise randomly masked SELFIES tokens and simultaneously regress 209 RDKit molecular descriptors), a fragment-averaged Evo2 genome embedding of the pathogen, and a Me-LLaMA text embedding of a Qwen2.5-Max-generated trait description. The molecule embedding serves as the query that attends in parallel to genome-derived and text-derived key–value banks; the resulting fused vector feeds task heads for MIC regression, small-molecule activity classification, and FICI synergy prediction, and also conditions the same DLM's reverse process during generation. Predictor-guided generation reweights each denoising step by the activity classifier's probability (and by a peptide classifier), with a remasking loop that lets early token mistakes be corrected mid-generation.

What would settle it

Inspect the Qwen2.5-Max-generated trait descriptions for the held-out strains (the one-third test strains and the fully held-out species clusters) and count how many mention a numeric MIC or a resistance/sensitivity statement matching the DBAASP test label. If the text for a held-out strain contains its MIC, the model's text encoder has direct access to the answer; a second check is to rerun the strain-wise and 3-species-wise splits after scrubbing all MIC numbers and susceptibility phrases from the text inputs and observe whether the reported $R^2$ values (0.5793 and 0.3744) survive unchanged or collapse.

Watch

Extended reading notes

Core claim

The central claim is that generalizing to novel pathogens requires pathogen context, not just molecular features, and that a single model can both regress MIC and generate molecules conditioned on that context. ApexOracle represents a pathogen by a genome embedding (Evo2) and a literature-derived text embedding (Me-LLaMA), and lets a molecule query attend to both via cross-attention; the molecule itself is embedded by a masked discrete diffusion language model trained on SELFIES strings with an auxiliary multi-target regression on 209 RDKit properties. In the paper's strain-wise test (two-thirds of a species' strains for training, one-third for testing), the ensembled model reaches $R^2 = 0.5793$ for held-out strains; with an entire taxonomic cluster of species held out in the coarsest '3 species-wise' split, the ensemble still reaches $R^2 = 0.3744$. The same architecture beats three small-molecule antibiotic classifiers under five-fold cross-validation and even in zero-shot transfer, and it extends to fractional-inhibitory-concentration (FICI) synergy prediction. Finally, using the internal MIC regressor to guide the diffusion sampler produces generated SELFIES molecules with predicted MIC distributions shifted toward low values and with lower maximum Tanimoto similarity to training compounds than unguided sampling, which the paper reads as evidence that the model designs genuinely new chemotypes rather than retrieving stored examples.

Load-bearing premise

The evaluation assumes the text description fed to the model for each strain does not already contain the MIC values or resistance/sensitivity outcomes the model is being asked to predict; the prompt used to generate those descriptions, however, explicitly asks for the corresponding MIC value and known susceptibility, and if that information ends up in the text the held-out results largely reflect retrieval rather than generalization.

Editorial extensions

If this is right

  • Given a new genome sequence (or even a text description) of an emerging pathogen, ApexOracle outputs MIC predictions immediately, without collecting strain-specific assay data, shrinking the usual experimental lag.
  • Small-molecule antibiotic screening becomes zero-shot: the model transfers knowledge from peptide MIC data and other antibiotics to strains with no small-molecule labels, which could cut the cost of building bespoke strain-specific classifiers.
  • The generator, conditioned on the predicted-MIC guidance, can propose candidate molecules specifically for resistant strains such as colistin-resistant E. coli and multidrug-resistant P. aeruginosa, providing a starting point for synthesis and testing.
  • The same architecture extends to synergy prediction, so combination regimens could be prioritized for a newly encountered pathogen based on predicted FICI values.
  • Because the model is not tied to a fixed molecular class, its design logic could in principle be carried over to antifungals or antivirals, as the paper itself suggests.

Reading between the lines

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

  • The strongest risk to the generalization claim is label leakage through the text modality: the Qwen2.5-Max prompt explicitly asks for MIC values and known resistance/sensitivity of the strain, and if those values are present in the text the model's text encoder has access to the very labels the held-out tests ask it to predict; the paper does not measure how much of the reported $R^2$ depends on t
  • A decisive check would be to rerun the strain-wise and species-wise splits with text descriptions stripped of all numeric MICs and susceptibility statements (or with a non-search LLM prompt that never retrieves them) and compare the resulting $R^2$; if it falls to near zero, the 'generalization' collapses to retrieval from the textual knowledge.
  • The generation experiments measure only predicted MIC and structural novelty (Tanimoto distance from training compounds); without wet-lab synthesis and susceptibility testing, high predicted potency is not evidence of actual activity—a gap the paper acknowledges in its limitations as missing toxicity and synthetic feasibility modeling.
  • Because the text embeddings come from a model told to cross-reference alternative strain identifiers, the textual descriptions may be richer for well-studied species, making the 'unseen species' test easier for common pathogens and harder for genuinely novel ones—a potential confound worth quantifying in follow-up work.
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

3 major / 4 minor

Summary. The paper presents ApexOracle, a multimodal model that combines a SELFIES-based diffusion language model with Evo2 genome embeddings and Me-LLaMA text embeddings of LLM-generated strain descriptions, aiming to predict antimicrobial activity (MIC) for held-out pathogens and to generate de novo antibiotic candidates. The main claims are state-of-the-art MIC prediction under strain-wise and species-wise held-out settings, competitive small-molecule antibiotic classification, synergy prediction, and predictor-guided molecular generation. The manuscript releases code, model weights, and data links.

Significance. If the held-out generalization results were valid, ApexOracle would be a notable advance: it unifies prediction and generation, handles non-canonical peptides via SELFIES, integrates genomic and literature-derived pathogen context, and ships code and weights. However, the central evaluation is compromised by label leakage through the LLM-generated text input, and the headline species-wise number includes a post hoc cluster exclusion. The practical and scientific value of the manuscript therefore depends on a re-run with leakage controls and a complete per-cluster report.

major comments (3)
  1. [Appendix C; §4.2] The central claim of transferability to unseen strains (§2.3, abstract) is undermined by label leakage through the text modality. The prompt in Appendix C instructs Qwen2.5-Max to provide 'Antibiotics and antimicrobial peptides Resistance' and 'Sensitivity' and explicitly says 'Make sure to include the corresponding MIC value if you can find it.' The resulting text is embedded by Me-LLaMA and used as the text key/value bank in the cross-attention fusion (§4.2 'Text Embeddings'; Fig.1a). No filtering or masking of MIC values, drug names, or susceptibility statements is reported. For a held-out strain whose description contains the very MIC value that the model is asked to predict, the model can retrieve the answer from its input, so the R² values in §2.3 do not establish generalization to novel pathogens. A leakage audit is required — for example, repeating the evaluation with the resistance/sensitivity sections removed, or with all MIC values and drug names masked — before the transferability claim can be assessed.
  2. [§2.3, Fig.2f] The 11 species-wise average R² of 0.4337 is reported only after excluding cluster 9 (Mycoplasmatota) post hoc: the text says 'excluding the sole exception of cluster 9.' The paper does not report the average including cluster 9 or a pre-specified criterion for exclusion. Since the claim is that the model 'remains capable of accurately predicting antibiotic efficacy' for held-out species, dropping the hardest cluster changes the headline number, and the full per-cluster results with and without cluster 9 should be given.
  3. [§2.4, Fig.2h] The zero-shot small-molecule comparison is not apples-to-apples: ApexOracle is compared against baseline models that were fine-tuned on the target strain's data, while the text states it 'matched or outperformed two of the four fine-tuned baseline models.' It is unclear why there are four baselines when three are listed, and Fig.2h's caption refers to 'two of three strains.' More importantly, the zero-shot claim requires a precise statement of which training data included or excluded the test strains; the phrase 'test-strain-irrelevant small-molecule antibiotics datasets' does not rule out partial overlap, and leakage from the small-molecule classification head could inflate the zero-shot numbers.
minor comments (4)
  1. [§2.2] The heading contains a typo: 'crutial' should be 'crucial.'
  2. [§4.2] There are several typos in the Methods text, including 'weigthed' for 'weighted' and 'SEIFIES' for 'SELFIES'; these should be corrected.
  3. [Eq. (8)] Equation (8) appears malformed as printed: the expression for E[y1, y'1] contains an ambiguous integral with mismatched limits and does not clearly define the expected value used for guidance; it should be rewritten in standard notation.
  4. [§2.4, Fig.2h] The main text and the Fig.2h caption disagree on the zero-shot result ('two of the four fine-tuned baseline models' vs. 'two of three strains'); this discrepancy should be reconciled.

Circularity Check

2 steps flagged · score 7.0 of 10

Held-out 'unseen' strains are not label-free: the Appendix C prompt asks the LLM to include MIC values in the text that ApexOracle embeds as input; generation is also evaluated with the same predictor that guides it.

  1. self definitional [Section 2.3 ('ApexOracle generalizes to unseen strains'); Appendix C ('Strain knowledge searching prompt'); Section 4.2 ('Text Embeddings')]
    "Even without exposure to examples from certain strains or species, the model remains capable of accurately predicting antibiotic efficacy against them. ... Make sure to include the corresponding MIC value if you can find it. ... For each strain, the descriptive text is first preprocessed by replacing the specific strain name with “this strain” to prevent overfitting to particular strain names."

    A held-out strain counts as 'unseen' only because its (molecule, MIC) rows are excluded from training, but the text input for that strain is generated by Qwen2.5-Max with search and explicitly instructed to report MIC values and known antibiotic/AMP resistance and sensitivity. This text is embedded by Me-LLaMA and used as the text key/value bank in cross-attention. Replacing the strain name with 'this strain' does not remove the numerical MIC values or drug names. For any test pair whose MIC is documented in the literature, the answer is therefore present in the model input, so the 'prediction' is partly retrieval from the LLM-written description, and the claimed absence of antimicrobial data fails by construction.

  2. self definitional [Section 2.7 ('Designing novel antibiotics for unseen strains via predictor-guided generation'); Section 4.2 ('Predictor-Based Guidance')]
    "Relative to unconditional sampling, our predictor-guided generation produced predicted MIC distributions strongly shifted toward lower values (Fig.3a, 3b). ... we utilize these noisy predictors, while reserving the clean-data-trained predictors solely for reporting the final properties of the generated molecules."

    The generated molecules are selected by the MIC regressor through predictor-based guidance, and the same family of predictors (clean-data version) is then used to report their 'predicted MIC' and to support the claim of designing active molecules. The shift toward low predicted MIC is therefore enforced by the optimization objective rather than being an independent finding: the evaluation metric is the same function that guides generation, so the result is true by construction rather than evidence of de novo activity.

full rationale

The paper's core novelty claim is transferability to novel pathogens. That claim depends on held-out strains having no antimicrobial labels in the model inputs. Appendix C instructs the LLM to include MIC values in the resistance/sensitivity sections, and Section 4.2 feeds this text into Me-LLaMA embeddings without masking. Numeric MIC values and drug names survive the 'this strain' replacement, so for many test pairs the label is in the input and the 'unseen' condition is unmet. This is a concrete, quotable reduction of the Section 2.3 generalization result to retrieval from LLM-written text. The generation experiments are similarly self-evaluated: predictor-based guidance minimizes predicted MIC, and the same predictor reports the outcome, so the low predicted MIC is by construction. The remaining technical components—DLM molecular representations, RDKit multitask pretraining, small-molecule baseline comparisons, and modality ablations—are internally consistent and not circular. Because the leak affects the central 'novel pathogen' claim, the circularity score is high, though not maximal: not every test pair necessarily has its exact MIC in the text, and the architecture itself is not inherently circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The architecture relies entirely on pretrained external components (Evo2, Me-LLaMA, Qwen2.5-Max, SELFIES, RDKit) and on DBAASP/ATCC/NCBI-derived labels. The hand-chosen hyperparameters listed above are needed to reproduce the reported training and generation behavior. No new physical entities are introduced; the generated molecules are model outputs, not inputs to the method.

free parameters (6)
  • Auxiliary loss weight lambda = 0.1
    Chosen by hand in Eq. (5) to weight the multi-task regression loss in DLM training.
  • Guidance strength gamma_1 = 15
    Chosen by hand in Section 4.2 to steer generation toward low predicted MIC.
  • Guidance strength gamma_2 = 15
    Chosen by hand in Section 4.2 to steer generation toward peptide-like molecules during the remasking stage.
  • Genome embedding scaling factor = 1e14
    Applied to Evo2-40B embeddings in Section 4.2 to reconcile a scale mismatch; chosen by hand.
  • Remasking window (t_on, t_off) = (0.55, 0.45)
    Selected in Section 4.2 to schedule remasking during generation.
  • Target MIC for guidance = 1
    Set in Section 4.2 to guide generation toward low MIC values; the scale is arbitrary on the transformed MIC.
assumptions (5)
  • domain assumption Evo2 genome embeddings contain sufficient cross-taxa signal to infer antimicrobial susceptibility.
    Used in Sections 2.1 and 2.3. The claim that the model can generalize to unseen strains from genomic context rests entirely on this.
  • ad hoc to paper Qwen2.5-Max generated strain descriptions are factually accurate and do not leak the prediction labels.
    The prompt in Appendix C asks for MIC values and sensitivity/resistance data. The description is the second pathogen-context channel; if labels leak, the out-of-distribution result is circular.
  • domain assumption DBAASP MIC measurements and the operator-preprocessing rules in Table 1 are reliable.
    All AMP training and evaluation build on these labels (Section 4.1).
  • domain assumption RDKit 209 descriptors are a sufficient self-supervised target to learn useful molecular representations.
    Used in the DLM multi-task regression objective (Section 4.2, Eq. 5).
  • standard math Masked diffusion language model equations from MDLM apply to SELFIES token diffusion.
    The forward/reverse process and loss in Section 4.2 are adopted from Sahoo et al. [34] without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting and generating antibiotics against future pathogens with ApexOracle." pith.science (2026). https://pith.science/paper/XBEULTCU

@misc{pith2026250707862,
  author       = {Pith},
  title        = {Pith review of: Predicting and generating antibiotics against future pathogens with ApexOracle},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBEULTCU}},
  note         = {Machine review of arXiv:2507.07862}
}
read the original abstract

Antimicrobial resistance (AMR) is escalating and outpacing current antibiotic development. Thus, discovering antibiotics effective against emerging pathogens is becoming increasingly critical. However, existing approaches cannot rapidly identify effective molecules against novel pathogens or emerging drug-resistant strains. Here, we introduce ApexOracle, an artificial intelligence (AI) model that both predicts the antibacterial potency of existing compounds and designs de novo molecules active against strains it has never encountered. Departing from models that rely solely on molecular features, ApexOracle incorporates pathogen-specific context through the integration of molecular features captured via a foundational discrete diffusion language model and a dual-embedding framework that combines genomic- and literature-derived strain representations. Across diverse bacterial species and chemical modalities, ApexOracle consistently outperformed state-of-the-art approaches in activity prediction and demonstrated reliable transferability to novel pathogens with little or no antimicrobial data. Its unified representation-generation architecture further enables the in silico creation of "new-to-nature" molecules with high predicted efficacy against priority threats. By pairing rapid activity prediction with targeted molecular generation, ApexOracle offers a scalable strategy for countering AMR and preparing for future infectious-disease outbreaks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages

  1. [1]

    Antimicrobial resistance: tackling a crisis for the health and wealth of nations

    O’neill, J. Antimicrobial resistance: tackling a crisis for the health and wealth of nations. Rev. Antimicrob. Resist.(2014)

    work page 2014

  2. [2]

    & Collins, J

    Wong, F., de la Fuente-Nunez, C. & Collins, J. J. Leveraging artificial intelligence in the fight against infectious diseases.Science381, 164–170 (2023)

    work page 2023

  3. [3]

    R., Torres, M

    Maasch, J. R., Torres, M. D., Melo, M. C. & de la Fuente-Nunez, C. Molecular de- extinction of ancient antimicrobial peptides enabled by machine learning.Cell Host & Microbe31, 1260–1274 (2023)

    work page 2023

  4. [4]

    D.et al.Discovery of antimicrobial peptides in the global micro- biome with machine learning.Cell187, 3761–3778 (2024)

    Santos-J´ unior, C. D.et al.Discovery of antimicrobial peptides in the global micro- biome with machine learning.Cell187, 3761–3778 (2024)

    work page 2024

  5. [5]

    D., Peng, J

    Wan, F., Torres, M. D., Peng, J. & de la Fuente-Nunez, C. Deep-learning-enabled antibiotic discovery through molecular de- extinction.Nature Biomedical Engineering8, 854–871 (2024)

    work page 2024

  6. [6]

    D.et al.Mining for encrypted peptide antibiotics in the human proteome

    Torres, M. D.et al.Mining for encrypted peptide antibiotics in the human proteome. Nature Biomedical Engineering6, 67–75 (2022)

    work page 2022

  7. [7]

    Xu, J.et al.Comprehensive assessment of machine learning-based methods for pre- dicting antimicrobial peptides.Briefings in bioinformatics22, bbab083 (2021)

    work page 2021

  8. [8]

    Sidorczuk, K.et al.Benchmarks in antimi- crobial peptide prediction are biased due to the selection of negative data.Briefings in Bioinformatics23, bbac343 (2022)

    work page 2022

Show all 45 references
  1. [9]

    Wang, R.et al.Diff-amp: tailored designed antimicrobial peptide framework with all-in- one generation, identification, prediction and optimization.Briefings in Bioinformatics25, bbae078 (2024)

    2024

  2. [10]

    & Dong, G

    Xing, W., Zhang, J., Li, C., Huo, Y. & Dong, G. iamp-attenpred: a novel antimicro- bial peptide predictor based on bert feature extraction method and cnn-bilstm-attention combination model.Briefings in bioinformat- ics25, bbad443 (2024). 13

    2024

  3. [11]

    Briefings in Bioinformatics26, bbae644 (2025)

    Cao, J.et al.Tg-cddpm: text-guided antimi- crobial peptides generation based on condi- tional denoising diffusion probabilistic model. Briefings in Bioinformatics26, bbae644 (2025)

    2025

  4. [12]

    Chen, N.et al.Tp-lmmsg: a peptide pre- diction graph neural network incorporat- ing flexible amino acid property representa- tion.Briefings in Bioinformatics25, bbae308 (2024)

    2024

  5. [13]

    Nature626, 177–185 (2024)

    Wong, F.et al.Discovery of a structural class of antibiotics with explainable deep learning. Nature626, 177–185 (2024)

    2024

  6. [14]

    Liu, G.et al.Deep learning-guided discov- ery of an antibiotic targeting acinetobacter baumannii.Nature Chemical Biology19, 1342–1350 (2023)

    2023

  7. [15]

    Chung, C.-R.et al.An ensemble deep learn- ing model for predicting minimum inhibitory concentrations of antimicrobial peptides against pathogenic bacteria.Iscience27 (2024)

    2024

  8. [16]

    Jin, S.et al.Ampgen: an evolution- ary information-reserved and diffusion-driven generative model for de novo design of antimi- crobial peptides.Communications Biology8, 1–14 (2025)

    2025

  9. [17]

    Brixi, G.et al.Genome modeling and design across all domains of life with evo 2.BioRxiv 2025–02 (2025)

    2025

  10. [18]

    Research squarers–3 (2024)

    Xie, Q.et al.Me-llama: Foundation large language models for medical applications. Research squarers–3 (2024)

    2024

  11. [19]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115(2024)

    Team, Q. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115(2024)

    2024 arXiv

  12. [20]

    & Aspuru-Guzik, A

    Krenn, M., H¨ ase, F., Nigam, A., Friederich, P. & Aspuru-Guzik, A. Self-referencing embed- ded strings (selfies): A 100% robust molecular string representation.Machine Learning: Science and Technology1, 045024 (2020)

    2020

  13. [21]

    Pirtskhalava, M.et al.Dbaasp v3: database of antimicrobial/cytotoxic activity and struc- ture of peptides as a resource for development of new therapeutics.Nucleic acids research 49, D288–D297 (2021)

    2021

  14. [22]

    M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702 (2020)

    Stokes, J. M.et al.A deep learning approach to antibiotic discovery.Cell180, 688–702 (2020)

    2020

  15. [23]

    & Tang, J

    Liu, S., Du, W., Ma, Z.-M., Guo, H. & Tang, J. A group symmetric stochastic differen- tial equation model for molecule multi-modal pretraining.International Conference on Machine Learning21497–21526 (2023)

    2023

  16. [24]

    Feng, S.et al.Unigem: A unified approach to generation and property prediction for molecules.arXiv preprint arXiv:2410.10516 (2024)

    2024 arXiv

  17. [25]

    & Ramsundar, B

    Ahmad, W., Simon, E., Chithrananda, S., Grand, G. & Ramsundar, B. Chemberta-2: Towards chemical foundation models.arXiv preprint arXiv:2209.01712(2022)

    2022 arXiv

  18. [26]

    Kim, S.et al.Pubchem substance and com- pound databases.Nucleic acids research44, D1202–D1213 (2016)

    2016

  19. [27]

    Genomics, proteomics & bioinformatics19, 602–610 (2021)

    Li, Y.et al.Smprot: a reliable reposi- tory with comprehensive annotation of small proteins identified from ribosome profiling. Genomics, proteomics & bioinformatics19, 602–610 (2021)

    2021

  20. [28]

    E.et al.Uniref clusters: a compre- hensive and scalable alternative for improving sequence similarity searches.Bioinformatics 31, 926–932 (2015)

    Suzek, B. E.et al.Uniref clusters: a compre- hensive and scalable alternative for improving sequence similarity searches.Bioinformatics 31, 926–932 (2015)

    2015

  21. [29]

    Uniprot: a worldwide hub of protein knowledge.Nucleic acids research 47, D506–D515 (2019)

    Consortium, U. Uniprot: a worldwide hub of protein knowledge.Nucleic acids research 47, D506–D515 (2019)

    2019

  22. [30]

    J.et al.Cyclops: generating virtual libraries of cyclized and constrained peptides including nonnatural amino acids.Journal of chemical information and modeling51, 829– 836 (2011)

    Duffy, F. J.et al.Cyclops: generating virtual libraries of cyclized and constrained peptides including nonnatural amino acids.Journal of chemical information and modeling51, 829– 836 (2011)

    2011

  23. [31]

    Feller, A. L. & Wilke, C. O. Peptide- aware chemical language model successfully predicts membrane diffusion of cyclic pep- tides.Journal of Chemical Information and Modeling(2025)

    2025

  24. [32]

    M., Corbett, P

    Lowe, D. M., Corbett, P. T., Murray-Rust, P. & Glen, R. C. Chemical name to structure: Opsin, an open source solution (2011)

    2011

  25. [33]

    Ross, J.et al.Large-scale chemical language representations capture molecular structure and properties.Nature Machine Intelligence 4, 1256–1264 (2022)

    2022

  26. [34]

    Sahoo, S.et al.Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems37, 130136–130184 (2024). 14

    2024

  27. [35]

    & Xie, S

    Peebles, W. & Xie, S. Scalable diffusion models with transformers.Proceedings of the IEEE/CVF international conference on computer vision4195–4205 (2023)

    2023

  28. [36]

    Schiff, Y.et al.Simple guidance mechanisms for discrete diffusion models.arXiv preprint arXiv:2412.10193(2024)

    2024 arXiv

  29. [37]

    qwen - max -0125

    Wang, G., Schiff, Y., Sahoo, S. S. & Kuleshov, V. Remasking discrete diffusion models with inference-time scaling.arXiv preprint arXiv:2503.00307(2025). Appendix A Non-canonical Amino Acids a: C[C@H](C(=O)O)N r: C(C[C@H](C(=O)O)N)CN=C(N)N n: C([C@H](C(=O)O)N)C(=O)N d: C([C@H](...

    2025

  30. [38]

    Describe its notable p h y s i o l o g i c a l traits

    Species I n f o r m a t i o n : Identify the species to which this strain belongs , s p e c i f y i n g whether it is Gram - positive , Gram - negative , Fungi , Archaea , or Protozoa . Describe its notable p h y s i o l o g i c a l traits . 22

  31. [39]

    Unique Mu ta tio ns : Describe any d i s t i n c t i v e genetic m ut at io ns i d e n t i f i e d in this strain compared to the wild - type strain of the same species , p a r t i c u l a r l y those af fe ct in g v ir ul en ce factors , me ta bo li c pathways , or the plasma...

  32. [40]

    Describe the mo le cu la r m e c h a n i s m s by which these confer r e s i s t a n c e to p a r t i c u l a r a n t i b i o t i c s

    A n t i b i o t i c s and a n t i m i c r o b i a l peptides R e s i s t a n c e : Outline any known a n t i b i o t i c s and a n t i m i c r o b i a l peptides r e s i s t a n c e m e c h a n i s m s a s s o c i a t e d with this strain , i nc lu di ng specific r e s i s t a...

  33. [41]

    Explain the m e c h a n i s m s by which these a n t i b i o t i c s exert their effects on the strain

    A n t i b i o t i c s and a n t i m i c r o b i a l peptides S e n s i t i v i t y : Identify a n t i b i o t i c s and a n t i m i c r o b i a l peptides to which this strain is known to be s en si ti ve . Explain the m e c h a n i s m s by which these a n t i b i o t i c s e...

  34. [42]

    Ensure that the d e s c r i p t i o n is based on current s c i e n t i f i c k no wl ed ge and includes relevant r e f e r e n c e s where applicable , do not insert r e f e r e n c e s before the R ef er en ce section

  35. [43]

    Please make sure to cross - r ef er en ce these a l t e r n a t i v e ID systems when se ar ch in g for relevant i n f o r m a t i o n

    Although the strain ID I provided is from ATCC , the same strain may also be ca ta lo ge d under di ff er en t i d e n t i f i e r s in other databases , such as DSM , KCTC , NCTC , JCM , or other unique nu mb er in g systems beyond these examples . Please make sure to cross -...

  36. [44]

    Do not include any bulletin points in your response

  37. [45]

    " " }] , e x t r a _ b o d y ={

    If i n f o r m a t i o n about the strain is u n a v a i l a b l e or cannot be found , JUST respond with ’ None ’ in the c o r r e s p o n d i n g section ! Do not respond with any further e x p l a n a t i o n ! " " " }] , e x t r a _ b o d y ={ " e n a b l e _ s e a r c h "...

Pith tools

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