Pith. sign in

REVIEW 4 major objections 7 minor 44 references

Preference-based Antibody Expression Ranking: Scaling with Large-scale Weak Supervision

T0 review · 4 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Antibody expression ranking can be learned from scarce yield labels plus millions of weakly labeled immunization-derived sequences; the recipe beats supervised baselines and improves as weak data grows.

desk verdict A practical, well-ablated approach to antibody expression ranking, but the weak-supervision signal likely teaches 'camelid-likeness' rather than 'expressibility', and all results rest on an unreleased private dataset. read the letter →

arxiv 2607.16263 v1 pith:DWPP7VBK submitted 2026-06-29 cs.LG cs.CEq-bio.QM

Josh Qixuan Sun , Morteza Babaie , Wenyang Hou , Mark Crowley , David Young This is my paper
classification cs.LGcs.CEq-bio.QM
keywords antibodyexpressionrankingdirectpreferenceoptimizationweaksupervisionproteinlanguagemodelspseudo-log-likelihoodcamelidVHHantibodiesIMGTalignmentdevelopabilityprediction
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

Measuring antibody expression yields is expensive, so antibody campaigns typically have thousands of candidate sequences but only hundreds of quantitative labels. The paper tries to establish that the ranking question — which candidates will express at usable yields — can be learned by converting scarce quantitative yields and massive weakly labeled immunization data into preference pairs, then fine-tuning a protein language model with Direct Preference Optimization. Its empirical claim is that a two-stage recipe — continual masked-language pretraining on 4.2 million camelid-derived VHH sequences, then preference optimization on strong pairs (yield gaps of at least 30 mg/L) and weak pairs (immunization-derived positives against zero-yield designs) — beats standard supervised baselines on the internal Exp-1K test set across three model backbones, with the composite ranking/classification score rising as the weak corpus grows from 1K to 4.2M sequences. A sympathetic reader would care because the approach converts the routine byproduct of immunization-based discovery into usable training signal, offering a scalable path to candidate prioritization in data-constrained antibody design.

What carries the argument

The preference pair is the load-bearing object. Strong pairs compare yields only past a 30 mg/L margin; weak pairs declare each of 4.2 million immunization-derived camelid VHHs preferred over a zero-yield design, converting the assumption that such sequences are mostly expressible into training signal. DPO becomes tractable for masked protein language models via IMGT alignment to a shared 128-position grid and a union-masked pseudo-log-likelihood, scoring only differing positions in one forward pass as identical positions cancel. Around this: Stage I masked-language pretraining on the 4.2M corpus, Stage II DPO with a KL penalty keeping the model near its Stage I reference.

What would settle it

Measure recombinant expression for a few hundred randomly sampled sequences from Camel-4M under the same SDS-PAGE assay used for Exp-1K: if the detected fraction is well below the assumed ~90%, the weak preference pairs carry corrupted supervision and the Figure 4 scaling curve is an artifact of a clean weak pool rather than of preference learning itself. A complementary check is to run the exact two-stage pipeline on an external VHH expression dataset; if the ranking gains over supervised baselines do not transfer beyond the internal Exp-1K test set, the claim is distribution-specific.

Watch

Extended reading notes

Core claim

The paper's central claim: antibody expressibility becomes learnable when supervision is expressed as relative preferences rather than pointwise labels. Measured yields form pairs only past a 30 mg/L margin; each of 4.2 million immunization-derived camelid VHHs is preferred over a zero-yield design, assuming such sequences are mostly expressible. IMGT alignment to 128 positions and a union-masked pseudo-log-likelihood score only differing positions in one pass, making DPO tractable for masked language models. On Exp-1K, the two-stage pipeline posts the best composite score on all three backbones (AntiBERTa2 45.0 vs 33.1/29.0); without the weak corpus DPO underperforms regression (Table 5).

Load-bearing premise

The framework rests on the assumption that the 4.2 million immunization-derived camelid VHH sequences in Camel-4M are predominantly expressible — supported only by 16 of 17 internally measured sequences (Appendix C.3) and an evolutionary argument — because if a substantial share of those sequences are actually secretion-incompetent or misfolded under the authors' assay, the weak preference pairs encode a corrupted signal and the scaling gains of Figure 4 would not transfer to

Editorial extensions

If this is right

  • Discovery campaigns that routinely generate immunization-derived sequence libraries gain a new use for them: as weak expressibility supervision, candidate prioritization improves without additional wet-lab yield measurements.
  • The union-masked score approximation removes the main computational obstacle to preference optimization for masked protein language models, so the same recipe transfers to other protein properties that come with relative rather than absolute labels, such as stability, aggregation, or affinity.
  • Because weak preference pairs narrow the validation-to-test generalization gap and smooth the optimization trajectory, small labeled datasets trained this way should generalize more reliably than regression heads trained on the same labels alone.
  • The scaling result implies that accumulating more immunization data — a byproduct of normal discovery work — directly translates into better ranking, so unmeasured industrial sequence archives become economically valuable assets.
  • The paper's own Appendix B shows a supervised binary classifier augmented with the same weak positives still wins on pure expressibility classification; the preference-based model's demonstrated advantage is ranking, so its natural deployment is as a prioritizer inside screening pipelines rather than as a hard express/no-express filter.

Reading between the lines

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

  • The weak-positive premise is the hinge: support for '>90% of camelid VHHs are expressible' is 16 of 17 internal measurements plus an evolutionary argument. A direct assay of a few hundred randomly sampled Camel-4M sequences would reveal whether the Figure 4 scaling curve is a property of the learning method or of an unusually clean weak pool; if the true positive rate is lower, the framework exten
  • The union-masked approximation weights every IMGT position equally, so hypervariable loops and gap positions carry as much preference signal as conserved framework residues; weighting positions by conservation, or masking gaps, is a concrete testable extension.
  • Because the fine-tuned model scores sequences by average pseudo-log-likelihood, it can serve as a ranking layer inside generative antibody design loops, reranking model-sampled candidates rather than only screening pre-existing libraries.
  • Pair construction discards most quantitative information: intermediate yields matter only when they clear the 30 mg/L margin. Graded preferences across multiple yield bins would let the scarce labeled data contribute more, and the paper's own sensitivity analysis shows the result is sensitive to how such signals are mixed.
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 / 7 minor

Summary. The paper proposes a two-stage framework for ranking antibody sequences by expression yield. Stage I performs continual masked-language-model pretraining on Camel-4M, a proprietary corpus of 4.2M camelid VHH sequences. Stage II applies Direct Preference Optimization (DPO) to protein language models, using a union-masked pseudo-log-likelihood approximation and IMGT alignment to handle variable-length sequences. Preference pairs come from two sources: strong pairs constructed from Exp-1K yield measurements with a margin of 30 mg/L, and weak pairs that pair Camel-4M sequences (assumed expressible) against zero-yield Exp-1K sequences. On the internal Exp-1K test set (n=126), the full pipeline outperforms supervised regression/classification baselines on most metric-by-backbone cells of Table 2, and Figure 4 shows increasing performance with weak-data scale. The paper also includes ablations on weak-data removal, sampling ratio, and similarity filtering.

Significance. If the central claims hold, the framework would be a practical way to exploit large industrial immunization datasets to improve expression ranking without large labeled sets. The paper's strengths include evaluation across three backbones, a sensible ablation structure (Tables 3-6, Appendix B), a clear statement of the binary-supervision comparison in Appendix B, and an analysis of the generalization gap (Table 4). However, two load-bearing issues need to be resolved before the claims are fully established: the validity of the union-masked PLL approximation and the source/expression confound in the weak preference signal. The evaluation is also based on a single split of a proprietary dataset, with no error bars or external validation.

major comments (4)
  1. [Section 3.3.2] Union-masked approximation: The claim that 'positions outside M need not be evaluated, as their contributions cancel in the score difference' is not correct under the PLL definition in Section 3.2. For a pair (yw, yl), the exact difference PLL(yw)-PLL(yl) includes, for positions i not in M, terms where the token is identical but the conditioning contexts differ because the M positions are different in yw and yl; these terms do not cancel. In addition, the proposed single-forward-pass evaluation of positions i in M uses a context in which all other M positions are masked, whereas exact PLL masks only position i. The approximation therefore has an unquantified bias. Because the efficiency contribution of the paper rests on this approximation, please provide a theoretical justification or an empirical comparison with exact PLL on a subset of pairs, showing that the relative ordering of pref
  2. [Section 3.3.3 / Section 4.1 / Appendix C.3] Weak supervision signal: The weak pairs (yk in Camel-4M, y0 in Exp-1K with zero yield) differ in two correlated attributes: expression status and origin (immunized camelid PBMC vs synthetic design). DPO only requires the preferred sequence to receive a higher score, so the model can satisfy the weak-pair signal by learning a 'camelid-likeness' boundary rather than an 'expressible vs non-expressible' ranking. The n=17 expressibility measurement in Appendix C.3 does not rule this out, and its 95% CI is wide (16/17, approximately 71-100%). Because the Exp-1K test set is entirely synthetic, the scaling gains in Figure 4 could reflect source discrimination that transfers only if expressible synthetic designs are systematically closer to natural camelid VHHs. Please add a control that separates origin from expression status, e.g., weak pairs whose positives are expressible synthetic sequences,
  3. [Section 4.2 / Table 2] Statistical evaluation: The main conclusion rests on a single 80/10/10 split with n=126 test sequences and no variance estimates, repeated seeds, or significance tests. Some comparisons are close or go in the opposite direction (e.g., IgBERT AUC 74.8 vs 78.6 for the best supervised baseline), so the 'most metrics' claim is hard to interpret without uncertainty quantification. Please report bootstrapped confidence intervals or multiple-seed means for the central results, and consider evaluating on one of the public benchmarks in Table 1 to support external validity.
  4. [Section 4.7 / Table 5] Inconsistent claim: The text states that 'without the inclusion of large-scale weak supervision (Camel-4M), the pure DPO approach yields slightly inferior results compared to direct regression.' The table shows the opposite for the paper's own composite metric: DPO without Camel-4M has Avg = sqrt(32.8 * 35.1) ~ 33.9, while supervised regression has Avg = sqrt(29.1 * 37.6) ~ 33.1, and MCC is higher (32.8 vs 29.1). Please correct the claim or specify which metric 'inferior' refers to, since this paragraph is used to argue that weak supervision is necessary.
minor comments (7)
  1. [Table 2] IgBERT row: '78.67.5' appears to be a formatting error; should read '78.6 7.5' for AUC and MCC.
  2. [Table 1] Header 'NYield Range (mg/L)' is ambiguous; clarify which columns are N, yield range, mean, and standard deviation.
  3. [Figure 4] Report numerical values or confidence intervals for the scaling curves; the current figure cannot be quantitatively assessed.
  4. [Appendix A.4] Typo: 'n the preference optimization phase' should be 'In the preference optimization phase'.
  5. [Appendix C.3] The 94.1% expressibility estimate from n=17 should be reported with a confidence interval, and the random sampling procedure should be specified.
  6. [Section 4.2] Clarify how 'Recall at top 20%' is computed: ranking by predicted score, and which score (PLL or DPO score)?
  7. [Appendix A.5] Unify notation Nsup/Nweak (Section 4.5) with N_S/N_L in Appendix A.5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical comparisons rest on measured labels and held-out data; the weak-supervision premise is an assumption with empirical support, not a definitional reduction.

full rationale

The paper's central claims are empirical: Stage I CPt and Stage II DPO are evaluated against measured yields in a held-out Exp-1K test split (§4.3), and strong preference pairs are constructed directly from quantitative yield differences (§3.3.3). No equation in the method defines the target ranking in terms of the fitted parameters or vice versa; the DPO objective is a standard learning objective, and the PLL/union-mask approximation is cited to external work (Ferragu et al., 2025; Zhao et al., 2024), not to a self-citation chain. The weak-supervision premise that camelid-derived sequences are mostly expressible is an input assumption supported by internal n=17 measurements (Appendix C.3) and external biology; this is a data-quality or confound concern (e.g., weak pairs may encode source identity rather than expressibility), not a circular reduction of the claimed result to its inputs. The paper even reports that removing weak data or using pure DPO without it does not yield the same gains (Table 5), which confirms the result is data-driven rather than forced by construction. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The ledger is dominated by empirical and domain choices rather than invented entities: five tuned or labeling-defined parameters (δ, β, α, the percentile operating point, the <1 mg/L cutoff) and six assumptions, of which two are load-bearing — the n=17-backed camelid-expressibility premise and the unquantified union-masked PLL approximation. No new physical or biological entities are postulated.

free parameters (5)
  • yield margin δ = 30 mg/L
    Preference pairs from labeled data are created only when yield difference ≥ δ; set to reduce experimental-noise ambiguity (§3.3.3). Affects the size and difficulty of the supervised pair set.
  • DPO KL regularization β = 0.1
    Controls deviation from the reference model in the DPO objective (§4.2); a tuned hyperparameter.
  • supervised oversampling rate α = 1000 (p_weak ≈ 74%)
    Balances strong vs weak pair exposure; Table 3 shows strong sensitivity (Avg ranges 13.9→40.1→31.1 across α=10..10000), and the headline results use the best value on the 125-sample validation set (§4.5).
  • binary decision percentile threshold = ≈73rd percentile of PLL scores (train positive fraction)
    PLL scores are converted to binary labels by taking the training-set percentile equal to the known expressible fraction and applying it to val/test (§3.2). Legitimate use of label frequency, but a fitted operating point.
  • non-expressible cutoff = <1 mg/L
    Exp-1K labeling convention: no detectable SDS-PAGE band is assigned yield <1 mg/L (App. C.1). Defines the binary ground truth and the zero-yield D0 set used in weak pairs.
assumptions (6)
  • domain assumption Immunization-derived camelid VHH sequences are predominantly expressible (>90%)
    The load-bearing premise that turns Camel-4M into weak positivity. Direct support: the authors' internal n=17 expression test, 16/17 (94.1%) expressible (App. C.3), plus cited camelid HCAb biology (Flajnik 2011; Zhang 2018). Used at §1, §4.1, and in weak-pair construction §3.3.3.
  • domain assumption Pseudo-log-likelihood (PLL) is a serviceable surrogate for sequence log-likelihood inside the DPO objective
    DPO's derivation (Rafailov 2023) presumes tractable likelihoods; masked LMs require the PLL substitution (§3.2, §3.3.2), following Salazar 2020 / Meier 2021. No task-specific error analysis is given.
  • domain assumption The union-masked single-forward-pass estimate preserves the DPO preference gradients
    Scores use only positions in M with 1/|M| normalization; shared positions are dropped (§3.3.2). Cancellation is exact only inside the approximation; bias from 1/|M| vs 1/L, IMGT gap tokens, and pairs with small |M| is unquantified.
  • domain assumption IMGT numbering provides valid positional homology across variable-length VHHs
    Alignment, masking, and gap placement all assume IMGT residue positions are comparable (ANARCI, §3.3.2, App. A.1). Standard in the field, but structurally required for pair construction.
  • domain assumption Semi-quantitative SDS-PAGE yield estimates support pair construction at margin δ=30
    Yields are estimated from band intensity relative to internal standards (App. C.1); the 30 mg/L margin (§3.3.3) is meant to absorb this noise but is not independently validated.
  • standard math The Bradley-Terry preference model underlying DPO maps yields and weak positives to valid pairwise preferences
    The DPO loss (Rafailov 2023) is adopted as-is; its validity after the PLL substitution is assumed without re-derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preference-based Antibody Expression Ranking: Scaling with Large-scale Weak Supervision." pith.science (2026). https://pith.science/paper/DWPP7VBK

@misc{pith2026260716263,
  author       = {Pith},
  title        = {Pith review of: Preference-based Antibody Expression Ranking: Scaling with Large-scale Weak Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWPP7VBK}},
  note         = {Machine review of arXiv:2607.16263}
}
read the original abstract

Antibody expression ranking is a critical task in antibody design, yet its modelling is severely hindered by the scarcity of labeled experimental data. To address this, we propose a unified preference-based learning framework that integrates scarce quantitative expression data with large-scale weak positive supervision from immunization data. We adapt Direct Preference Optimization (DPO) to protein language models by introducing a union-masked log-likelihood approximation and IMGT-based alignment, enabling efficient training on variable-length sequences. Evaluating on a diverse internal dataset of 1254 labeled sequences and 4 million unlabeled camelid-derived antibodies, we show that our method consistently outperforms baselines on most metrics. Our results demonstrate that preference learning can effectively learn from weak supervision, providing a scalable solution for antibody expressibility optimization in data-constrained settings. Project page: https://kisoji-biotechnology-inc.github.io/Preference-Expression-Ranking/.

Figures

Figures reproduced from arXiv: 2607.16263 by the authors.

Figure 1
Figure 1. Construction of preference pairs. Strong supervision (top) uses yield differences with margin δ; weak supervision (bot￾tom) pairs immunization-derived sequences with zero-yield anti￾bodies. Both contribute to the unified preference dataset Dpair. To make the definition of M(yw, yl) well-defined for an￾tibody sequences with variable lengths, we first align se￾quences using a standardized numbering scheme. Specifi￾cal… view at source ↗
Figure 2
Figure 2. Histogram of Exp-1K yield score [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Scaling with weak supervision. Results are shown with and without Stage I continual pretraining (CPT), while keeping a consistent exposure ratio between supervised and weak pairs. uniquely synergistic with Stage II preference optimization, significantly outperforming the use of Stage II alone. 4.4. Scaling with Weak Supervision To investigate how our framework benefits from increasing scales of weak supervision, we … view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Validation performance trajectories over training for different methods. Method MCC ↑ Tau ↑ SCC ↑ Supervised 29.1 37.6 42.2 w/o Camel-4M 32.8 35.1 36.2 Ours 43.7 46.3 50.8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Pairwise similarity distribution of the Exp-1K dataset. The similarity is calculated as the IMGT-aligned Hamming identity across all sequence pairs. Pairwise Similarity Distribution. To facilitate the analysis of preference construction (Section 4.7), we compute the pa…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 1 canonical work pages

  1. [1]

    BioRxiv , pages=

    FLAb: Benchmarking deep learning methods for antibody fitness prediction , author=. BioRxiv , pages=. 2024 , publisher=

    2024

  2. [2]

    Proceedings of the National Academy of Sciences , volume=

    Biophysical properties of the clinical-stage antibody landscape , author=. Proceedings of the National Academy of Sciences , volume=. 2017 , publisher=

    2017

  3. [3]

    and Borhani, David , title =

    Arsiwala, Ammar and Bhatt, Rebecca and Yang, Yaoyu and Quintero Cadena, Porfi and Anderson, KC and Ao, Xiang and van Niekerk, Lood and Rosenbaum, Adam and Bhatt, Aanal and Smith, Alex and Grippo, Lucia and Cao, Xing and Cohen, Rich and Patel, Jay and Allen, Olga and Faraj, Ali and Nandy, Anisha and Hocking, Jason and Tural, Berk and Salvador, Sara and Jac...

    2025

  4. [4]

    Nett and Beth Sharkey and Beata Bobrowicz and Isabelle Caffry and Yao Yu and Yuan Cao and Heather Lynaugh and Michael Brown and Hemanta Baruah and Laura T

    Tushar Jain and Tingwan Sun and Stéphanie Durand and Amy Hall and Nga Rewa Houston and Juergen H. Nett and Beth Sharkey and Beata Bobrowicz and Isabelle Caffry and Yao Yu and Yuan Cao and Heather Lynaugh and Michael Brown and Hemanta Baruah and Laura T. Gray and Eric M. Krauland and Yingda Xu and Maximiliano Vásquez and K. Dane Wittrup , title =. Proceedi...

    doi:10.1073/pnas.1616408114 2017

  5. [5]

    and Lee, Kelvin H

    Szkodny, Alana C. and Lee, Kelvin H. , title =. Biotechnology Progress , volume =. doi:https://doi.org/10.1002/btpr.3466 , url =. https://aiche.onlinelibrary.wiley.com/doi/pdf/10.1002/btpr.3466 , abstract =

    work page doi:10.1002/btpr.3466

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    A SARS-CoV-2 interaction dataset and VHH sequence corpus for antibody language models , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    arXiv preprint arXiv:2510.19474 , year=

    g-DPO: Scalable Preference Optimization for Protein Language Models , author=. arXiv preprint arXiv:2510.19474 , year=

    arXiv

  8. [8]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

Show all 44 references
  1. [9]

    Advances in neural information processing systems , volume=

    Language models enable zero-shot prediction of the effects of mutations on protein function , author=. Advances in neural information processing systems , volume=

  2. [10]

    arXiv preprint arXiv:2004.03497 , year=

    Progen: Language modeling for protein generation , author=. arXiv preprint arXiv:2004.03497 , year=

    2004 arXiv

  3. [11]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Masked language model scoring , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  4. [12]

    Nature methods , volume=

    Unified rational protein engineering with sequence-based deep representation learning , author=. Nature methods , volume=. 2019 , publisher=

    2019

  5. [13]

    Nature methods , volume=

    Low-N protein engineering with data-efficient deep learning , author=. Nature methods , volume=. 2021 , publisher=

    2021

  6. [14]

    Nature communications , volume=

    An alpaca nanobody neutralizes SARS-CoV-2 by blocking receptor interaction , author=. Nature communications , volume=. 2020 , publisher=

    2020

  7. [15]

    International Conference on Research in Computational Molecular Biology , pages=

    Contrastive fitness learning: Reprogramming protein language models for low-n learning of protein fitness landscape , author=. International Conference on Research in Computational Molecular Biology , pages=. 2024 , organization=

    2024

  8. [16]

    Likelihood-based fine-tuning of protein language models for few-shot fitness prediction and design , author=

  9. [17]

    Nature biotechnology , volume=

    Learning protein fitness models from evolutionary and assay-labeled data , author=. Nature biotechnology , volume=. 2022 , publisher=

    2022

  10. [18]

    Protein Language Model Fitness is a Matter of Preference , author=

  11. [19]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  12. [20]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Protein Inverse Folding From Structure Feedback , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  13. [21]

    arXiv preprint arXiv:2506.00297 , year=

    Improving Protein Sequence Design through Designability Preference Optimization , author=. arXiv preprint arXiv:2506.00297 , year=

    arXiv

  14. [22]

    Advances in Neural Information Processing Systems , volume=

    Antigen-specific antibody design via direct energy-based preference optimization , author=. Advances in Neural Information Processing Systems , volume=

  15. [23]

    arXiv preprint arXiv:2412.20984 , year=

    Alignab: Pareto-optimal energy alignment for designing nature-like antibodies , author=. arXiv preprint arXiv:2412.20984 , year=

  16. [24]

    arXiv preprint arXiv:2512.17815 , year=

    Structure-Aware Antibody Design with Affinity-Optimized Inverse Folding , author=. arXiv preprint arXiv:2512.17815 , year=

  17. [25]

    arXiv preprint arXiv:2510.21161 , year=

    RiboPO: Preference Optimization for Structure-and Stability-Aware RNA Design , author=. arXiv preprint arXiv:2510.21161 , year=

  18. [26]

    Bioinformatics , volume=

    Protein--Sol: a web tool for predicting protein solubility from sequence , author=. Bioinformatics , volume=. 2017 , publisher=

    2017

  19. [27]

    Bioinformatics Advances , volume=

    G-RANK: an equivariant graph neural network for the scoring of protein--protein docking models , author=. Bioinformatics Advances , volume=. 2023 , publisher=

    2023

  20. [28]

    Khade and Michael Maser and Vladimir Gligorijevic and Andrew Watkins , title =

    Pranav M. Khade and Michael Maser and Vladimir Gligorijevic and Andrew Watkins , title =. 2023 , doi =. https://www.biorxiv.org/content/early/2023/06/28/2023.06.26.546331.full.pdf , journal =

    2023

  21. [29]

    BioRxiv , pages=

    Enhancing antibody language models with structural information , author=. BioRxiv , pages=. 2024 , publisher=

    2024

  22. [30]

    PLOS Computational Biology , volume=

    Large scale paired antibody language models , author=. PLOS Computational Biology , volume=. 2024 , publisher=

    2024

  23. [31]

    BioRxiv , volume=

    Language models of protein sequences at the scale of evolution enable accurate structure prediction , author=. BioRxiv , volume=

  24. [32]

    Bioinformatics , volume=

    ProteinBERT: a universal deep-learning model of protein sequence and function , author=. Bioinformatics , volume=. 2022 , publisher=

    2022

  25. [33]

    International Journal of Molecular Sciences , volume=

    DeepSTABp: a deep learning approach for the prediction of thermal protein stability , author=. International Journal of Molecular Sciences , volume=. 2023 , publisher=

    2023

  26. [34]

    Synthetic and Systems Biotechnology , volume=

    Transfer learning with pre-trained language models for protein expression level prediction in Escherichia coli , author=. Synthetic and Systems Biotechnology , volume=. 2026 , publisher=

    2026

  27. [35]

    Bioinformatics , volume=

    NetSolP: predicting protein solubility in Escherichia coli using language models , author=. Bioinformatics , volume=. 2022 , publisher=

    2022

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    Proteingym: Large-scale benchmarks for protein fitness prediction and design , author=. Advances in Neural Information Processing Systems , volume=

  29. [37]

    Developmental & Comparative Immunology , volume=

    IMGT unique numbering for immunoglobulin and T cell receptor variable domains and Ig superfamily V-like domains , author=. Developmental & Comparative Immunology , volume=. 2003 , publisher=

    2003

  30. [38]

    Bioinformatics , volume=

    ANARCI: antigen receptor numbering and receptor classification , author=. Bioinformatics , volume=. 2016 , publisher=

    2016

  31. [39]

    2009 , publisher=

    The ELISA guidebook , author=. 2009 , publisher=

    2009

  32. [40]

    Clinical Chemistry , volume=

    The ELISA, enzyme-linked immunosorbent assay , author=. Clinical Chemistry , volume=. 2010 , publisher=

    2010

  33. [41]

    Bioinformatics , pages=

    RP3Net: a deep learning model for predicting recombinant protein production in Escherichia coli , author=. Bioinformatics , pages=. 2026 , publisher=

    2026

  34. [42]

    PLoS biology , volume=

    A case of convergence: why did a simple alternative to canonical antibodies arise in sharks and camels? , author=. PLoS biology , volume=. 2011 , publisher=

    2011

  35. [43]

    Frontiers of Agricultural Science and Engineering , volume=

    A comprehensive analysis of immunoglobulin heavy chain genes in the Bactrian camel (Camelus bactrianus) , author=. Frontiers of Agricultural Science and Engineering , volume=

  36. [44]

    Frontiers in Immunology , volume=

    Genetic removal of the CH1 exon enables the production of heavy chain-only IgG in mice , author=. Frontiers in Immunology , volume=. 2018 , publisher=

    2018

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.