Pith. sign in

REVIEW 2 major objections 5 minor 15 references

CoSiNE models antibody affinity maturation as a continuous-time Markov chain whose per-site rates are conditioned on the whole sequence, proving a first-order approximation to the true sequential point mutation process with error only quadr

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

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

T0 review · deepseek-v4-flash

2026-08-02 21:46 UTC pith:427W22GW

load-bearing objection A solid methodological contribution with an honest conditional theory; the main gap is that the rate-matching assumption behind the central bound is not verified for the fitted model, but this is addressable. the 2 major comments →

arxiv 2602.18982 v4 pith:427W22GW submitted 2026-02-21 cs.LG q-bio.PE

Conditionally Site-Independent Neural Evolution of Antibody Sequences

Stephen Zhewen Lu , Aakarsh Vermani , Kohei Sanno , Jiarui Lu , Frederick A Matsen , Milind Jagota , Yun S. Song This is my paper
classification cs.LG q-bio.PE
keywords antibody affinity maturationcontinuous-time Markov chainsomatic hypermutationepistasisvariant effect predictionphylogenetic modelclassifier guidanceprotein language model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a deep neural network can parameterize a sequence-evolution clock—a continuous-time Markov chain where each amino-acid site has its own rate matrix that depends on the entire rest of the sequence—and that this tractable factorized model faithfully approximates the biologically realistic process in which mutations occur one at a time. It proves the approximation error is bounded by a term quadratic in evolutionary branch length, so for the short branches typical of affinity maturation the model captures first-order mutation dynamics and the dominant epistatic corrections. It then shows the model outperforms state-of-the-art antibody and protein language models on zero-shot variant effect prediction, by subtracting the neutral somatic hypermutation background and reading the residue as selection. Finally, it introduces classifier-guided Gillespie sampling that steers the chain toward higher predicted binding affinity for a chosen antigen while preserving humanness and structural plausibility. If right, this gives protein engineers a generative model that is both temporally grounded and epistatic—closing a gap between phylogenetic models and language models.

Core claim

The central claim is that a conditionally site-independent CTMC—whose generator is the Kronecker sum of per-site rate matrices Qθ(x)_ℓ each output by a neural network from the full parent sequence x—provides a first-order approximation to any sequential point-mutation process over protein sequence space. Concretely, if the learned per-site rates equal the true instantaneous rates on all single-residue changes, the L1 error between factorized and true transition probability vectors is at most (λt)², i.e., quadratic in branch length; and Gillespie sampling using these rates produces trajectories exactly distributed according to the true process. The paper further claims this evolutionary groun

What carries the argument

The core object is the conditionally site-independent generator: a function Qθ(x)_ℓ mapping a full antibody sequence to a 20×20 rate matrix per site, giving the factorized transition likelihood pθ(y|x,t) = ∏_ℓ exp(tQθ(x)_ℓ)_{xℓ,yℓ}. This relaxes the classical independent-sites assumption by making each site's rates depend on the entire sequence, so epistasis enters through conditioning. The mathematical engine is uniformization: comparing the Kronecker-sum generator to the true generator row by row, so the n=0 and n=1 terms match and the error starts at n≥2, yielding the O(t²) bound. The second instrument is the selection score, a log-likelihood ratio between CoSiNE and a pre-trained neutral

Load-bearing premise

The theoretical guarantees assume the per-site rates learned by the model exactly match the true instantaneous mutation rates of the point-mutation process; because the model is trained on factorized likelihoods that marginalize intermediate states on long branches, the paper concedes this match is not guaranteed (Section C.1), and the only supporting evidence is indirect—synthetic codon experiments and a marginal 52% vs 38.8% advantage on held-out trees.

What would settle it

Train CoSiNE on a ground-truth epistatic process (e.g., the synthetic codon setup at epistasis ε=1) using long branches, then compare the fitted per-site rates to the true rates: Figure S2 already shows unweighted MLE's relative Frobenius error grows with ε, so a direct measurement on real antibody clonal trees—or a demonstration that transition-probability error grows faster than (λt)² at long t—would settle whether the rate-matching assumption holds.

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

If this is right

  • CoSiNE's VEP performance on the FLAb2 assays (Table 1) is the first demonstration that an explicitly evolutionary model, rather than a marginal-distribution language model, gives the best zero-shot predictions of antibody expression and binding fitness.
  • Because the error bound is O(t²), the model is principled precisely in the short-branch regime that dominates affinity maturation, where the first-order signal O(λt) dominates the quadratic correction.
  • Gillespie sampling (Algorithm S1) yields transition distributions closer to the true process than per-site matrix exponentiation at all epistasis levels, and on real clonal trees its leaves are closer to held-out sequences (52.0% vs 38.8% better Hamming distance, r=0.64 vs 0.56).
  • Guided Gillespie with TAG produces affinity gains matching exact guidance at 488–928× lower oracle cost, enabling in silico affinity maturation from naive sequences with structural and humanness scores comparable to natural binders.
  • The model disentangles selection from mutation by design, so fitness inference does not require clamping or heuristics as DASM does.

Where Pith is reading between the lines

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

  • If the approach generalizes beyond antibodies to other proteins, the same conditional site-independent CTMC architecture could replace i.i.d. training objectives wherever phylogenetic trees or time-stamped sequence data exist—making the language-model prior unnecessary for evolutionary inference.
  • The rate-matching failure mode suggests a concrete research program: training objectives that are unbiased for instantaneous rates (e.g., likelihood weighting by 1/t, or two-step estimation from short-branch transitions) could restore the theoretical guarantees in the long-branch regime—the paper's own Figure S2 shows SNR weighting works in data-rich synthetic settings.
  • The guidance machinery is not specific to antibodies: any sequence-to-property predictor trained on experimental data can steer the chain, so CoSiNE could serve as a general scaffold for constrained protein design where the oracle is not differentiable or is noisy.
  • A testable consequence of the selection-score interpretation: if the score really isolates fixation probability, then across a DMS library the score should correlate with experimentally measured enrichment for binding but not with non-selected mutation rates; the paper's Figure S6 already shows the edit-distance bias removed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces CoSiNE, a conditionally site-independent neural CTMC for antibody sequence evolution. The transition probability factorizes as a product of per-site matrix exponentials whose rate matrices are produced by a neural network conditioned on the parent sequence. The authors prove (Prop. 4.1) that, under an exact rate-matching assumption, the factorized transition is a first-order approximation of a full sequential point-mutation process with O(t^2) error, and (Lemma 4.2) that a Gillespie algorithm using the model's instantaneous rates samples exactly from the target process. On top of this, the paper derives a selection score from a log-likelihood ratio with a Thrifty SHM model for zero-shot variant effect prediction, and proposes Guided Gillespie sampling for affinity optimization. Empirically, CoSiNE outperforms several antibody/protein language models and DASM on four DMS assays (Table 1), and guided sampling improves predicted binding affinity while preserving humanness and structural plausibility.

Significance. If the theoretical guarantees held for the fitted model, this would be a valuable contribution: it connects expressive sequence models with phylogenetic CTMC theory, provides a principled selection/mutation decomposition, and gives a concrete sampling algorithm. The paper is also commendable for explicitly acknowledging the central assumption in Prop 4.1/Lemma 4.2 and for attempting an SNR-weighted training correction (Sec. C.1) even though it did not improve real-data results. The synthetic codon experiments are a careful, controlled validation of the theory under different epistasis levels. The VEP results are competitive and include a useful ablation showing the benefit of the SHM correction.

major comments (2)
  1. [Sections 4, C.1] The central theoretical claims (Prop. 4.1 and Lemma 4.2) are conditional on the assumption that the fitted per-site rates exactly match the instantaneous rates of the true sequential point-mutation process for every single-site mutation. The paper itself states in Sec. C.1 that training with the factorized likelihood (Eq. 2) biases rates on long branches because unobserved intermediate states are marginalized, and that the proposed SNR-weighted correction did not improve results on real antibody data. Therefore the O(t^2) bound and the Gillespie exactness are not established for the model actually used in the experiments. The only direct supporting evidence comes from synthetic codon systems (L=3, 64 states), which are far from antibody-scale; the real-tree comparison (Sec. C.3) is indirect and modest (52.0% vs. 38.8%). This is a load-bearing gap: if rate matching fails, the theoretical
  2. [Appendix D.1, Eq. (13)-(14)] In the proof of Proposition D.1, the uniformization rate is defined as λ = max_z{-Q_zz}, and it is asserted that this is 'also equal to max_z{-Qθ(x)_zz}'. This equality does not follow from the rate-matching assumption, which only matches row x of Q and Qθ(x) for single-site mutations; the exit rates of Qθ(x) from other states z ≠ x are determined by per-site matrices conditioned on x, not by Q, and can exceed λ. If λ is smaller than an exit rate of Qθ(x), then S = I + Qθ(x)/λ is not a legitimate stochastic matrix for uniformization. The proof can likely be repaired by taking λ' = max(λ_Q, λ_Qθ(x)) as the common uniformization rate, but as written the argument is not valid.
minor comments (5)
  1. [Section 5.4, Figure 5] The guidance strength γ=5 is adopted after inspecting the same evaluation figure that shows the effect of γ on predicted affinity. The claim that γ=5 generates profiles overlapping real binders is therefore partly by construction. Please state whether γ was selected on a held-out set or provide a sensitivity analysis that does not rely on the final evaluation plot.
  2. [Section C.3, Figure S4 caption] There are inconsistencies between the caption and the text: the caption reports 'Gillespie better 48.8%' and r=0.62, while the text reports 52.0% and r=0.64. Please correct the discrepancy.
  3. [Section 3 / Table 1] SiteRM is described as a strong baseline in related work but is not included in the VEP comparison. Adding it (or explaining its omission) would strengthen the empirical claims.
  4. [Notation throughout] The model name is written inconsistently as COSINE and CoSiNE. Please standardize.
  5. [Table S4] The table formatting appears to have misaligned columns, making some entries difficult to parse. Please reformat.

Circularity Check

0 steps flagged

No circular derivation: central theorems are conditional, and the only minor self-citations are not load-bearing.

full rationale

I find no step in the claimed derivation chain that reduces to its own inputs. Proposition 4.1 and Lemma 4.2 are explicit conditional statements: they assume the per-site rates satisfy (Q_theta(x)_ell)_{x_ell,y_ell} = Q_{x,y} for all single-mutant pairs, and from that premise they prove the O(t^2) bound and Gillespie exactness. The premise is an accuracy/identifiability condition on the fitted rates, not the conclusion being derived. The paper itself flags the key limitation in Section C.1: factorized MLE training can bias rates on long branches because intermediate states are marginalized, and the proposed SNR-weighted loss did not improve real-data results. That is an acknowledged correctness/validation risk for the fitted model, but it does not make the theorem circular; the theorem would remain valid as a mathematical statement even if the trained model does not satisfy the premise. The VEP selection score is evaluated on held-out FLAb2 DMS fitness values, and the Thrifty likelihood is obtained from a separately trained SHM model based on out-of-frame sequences, so the log-likelihood ratio is not constructed from the target fitness labels. The self-citations to DASM and Thrifty share co-authors, but they provide the data protocol, benchmark splits, and a pretrained baseline rather than the core derivation; they are not invoked as a uniqueness theorem or as justification that the model choice is forced. The choice of gamma by inspecting Figure 5 is a hyperparameter-tuning concern and a potential source of over-optimism, but it is not a fitted input renamed as a prediction. No other pattern — self-definition, ansatz-smuggling via citation, or renaming a known result — is present. Accordingly, no specific circular step can be exhibited; the modest score reflects only minor self-reference and tuning concerns that do not rise to circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The model contributes a neural parameterization of rates; the main free parameters are the network weights and two evaluation hyperparameters (t and γ). No new physical entities are postulated. The central theory rests on standard CTMC and population-genetics axioms plus domain assumptions about tree inference and the neutral SHM baseline.

free parameters (5)
  • θ (neural network weights) = unknown; trained on ~2M transitions
    All rates Qθ(x)_ℓ are outputs of a neural network whose weights are fit by maximum likelihood on the clonal-tree transition dataset.
  • VEP branch length t = 0.2
    Fixed for all DMS assays; no search over t reported (Section 5.3).
  • guidance strength γ = 5
    Chosen because γ≥10 exceeded biological binders and γ=5 overlapped real binders (Section 5.4, Figure 5). Tuned on the evaluation set.
  • SNR weighting δ = not used
    Explored in Appendix C.1 but abandoned because no improvement on real data; final model uses unweighted loss.
  • MC dropout masks for σθz = 10
    Used to estimate oracle uncertainty (Appendix B.4).
axioms (6)
  • domain assumption Antibody evolution proceeds by sequential single-nucleotide point mutations; insertions/deletions are negligible
    Invoked in Section 2.3 to justify the point-mutation CTMC state space; standard for SHM but excludes indels.
  • domain assumption IQ-TREE under K80 + 4-category FreeRate reconstructs correct clonal trees and ancestral sequences
    Section A.1: all training transitions come from this phylogenetic pipeline; errors in ASR would bias learned rates.
  • standard math Halpern-Bruno decomposition Qxy = k·μxy·Pfix
    Section 4.1 Eq. (4); standard mutation-selection model.
  • domain assumption Thrifty trained on out-of-frame sequences captures neutral SHM rates μxy
    Section 4.1; used to subtract mutation bias. If Thrifty is misspecified, the selection score inherits bias.
  • standard math Kimura fixation probability is valid for B-cell affinity maturation
    Appendix D.3 Eq. (15); population-genetics approximation.
  • standard math Small-t linear approximation pθ(y|x,t) ≈ t·Qxy for x≠y
    Section 4.1; used to derive the log-ratio selection score; error at t=0.2 unquantified.

pith-pipeline@v1.3.0-alltime-deepseek · 25665 in / 15388 out tokens · 135371 ms · 2026-08-02T21:46:36.265450+00:00 · methodology

0 comments
read the original abstract

Common deep learning approaches for antibody engineering focus on modeling the marginal distribution of sequences. By treating sequences as independent samples, however, these methods overlook affinity maturation as a rich and largely untapped source of information about the evolutionary process by which antibodies explore the underlying fitness landscape. In contrast, classical phylogenetic models explicitly represent evolutionary dynamics but lack the expressivity to capture complex epistatic interactions. We bridge this gap with CoSiNE, a continuous-time Markov chain parameterized by a deep neural network. Mathematically, we prove that CoSiNE provides a first-order approximation to the intractable sequential point mutation process, capturing epistatic effects with an error bound that is quadratic in branch length. Empirically, CoSiNE outperforms state-of-the-art language models in zero-shot variant effect prediction by explicitly disentangling selection from context-dependent somatic hypermutation. Finally, we introduce Guided Gillespie, a classifier-guided sampling scheme that steers CoSiNE at inference time, enabling efficient optimization of antibody binding affinity toward specific antigens.

Figures

Figures reproduced from arXiv: 2602.18982 by Aakarsh Vermani, Frederick A Matsen, Jiarui Lu, Kohei Sanno, Milind Jagota, Stephen Zhewen Lu, Yun S. Song.

Figure 1
Figure 1. Figure 1: Overview of COSINE. Given an antibody sequence x, the neural network outputs site-specific rate matrices conditioned on the full sequence. Each matrix is evolved for duration t to yield a per-site transition distribution p(yl | x, t). Assuming conditional independence, we take the product of the per-site transition probabilities to yield the full sequence transition probability p(y | x, t). sumption enable… view at source ↗
Figure 2
Figure 2. Figure 2: Mean per-site likelihood of COSINE versus DASM+Thrifty on held out evolutionary transitions from the test set. COSINE achieves better model fit, especially on transi￾tions with longer branch lengths (t ≥ 0.25). 1 25 50 75 100 1 25 50 75 100 Mutated Position 1 25 50 75 100 1 25 50 75 100 Measured Position Heavy Chain Light Chain Heavy Chain Light Chain 0 2 4 6 8 10 12 M e a n jj¢ Qµ(x)` jjF [PITH_FULL_IMAG… view at source ↗
Figure 3
Figure 3. Figure 3: Categorical Jacobian for antibody 47D11 from CoV￾AbDab. The heatmap displays the sensitivity of the model’s output predictions (y-axis) to specific mutations in the input sequence (x￾axis). Sensitivity is measured as the Frobenius norm of the change in predicted rate matrix, averaged over all possible mutations. Cyan squares denote the CDR regions for both chains. of the antibody, so changes in one CDR res… view at source ↗
Figure 4
Figure 4. Figure 4: DMS evaluation results for COSINE across expres￾sion (green) and binding (purple) assays. A solid color indicates the log-likelihood, log pθ(y | x, t), and hatching indicates the selection score defined in Equation (5), which utilizes Thrifty like￾lihoods to separate selection from neutral mutation. 2 1 0 1 2 3 Fitness from Naive Unguided (N=62) Guided ( =1) (N=62) Guided ( =5) (N=64) Guided ( =10) (N=63) … view at source ↗
Figure 5
Figure 5. Figure 5: Guided Gillespie consistently steers the predicted bind￾ing affinity against SARS-CoV-1 of the sampled leaf sequences. We plot the change in predicted binding affinity from the naive root sequence used to start sampling. Known binders from CoV-AbDab are plotted for reference in red. We compare the affinity gain of our generated leaf sequences against 415 SARS-CoV-1 and 766 SARS-CoV-2 binders from the Cov-A… view at source ↗
Figure 6
Figure 6. Figure 6: Antibodies sampled by COSINE under guidance (γ = 5) maintain high structural quality (AbodyBuilder3 pLDDT) and humanness (OASis). We compare against unguided leaf samples (black) and CoV-AbDab binders (red). 5.5. Local Optimization of Antibody CDRs Beyond simulating affinity maturation over long branch lengths, we evaluated whether COSINE could perform con￾strained local optimization. We tasked the model w… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

15 extracted references · 5 linked inside Pith

  1. [1]

    Masked language models (AbLang-2, ESM-2 150M, ESM-2 650M) are evaluated via pseudo-perplexity in accordance with the FLAb2 benchmark

  2. [2]

    Autoregressive models (ProGen2 Small, ProGen2 Medium) are evaluated via perplexity, also in accordance with the FLAb2 benchmark

  3. [3]

    TheDASMmodel is evaluated by summing its log selection factors as described in Matsen IV et al. (2025). B.3.2. CALCULATINGSEQUENCELIKELIHOODS WITHTHRIFTYSHM MODEL To estimate sequence transition likelihoods under SHM, we use ThriftyHumV0.2-59-hc-tangshm, which provides per-codon transition probabilities with a multihit correction described in Matsen IV et...

    arXiv 2025

  4. [6]

    org/CorpusID:268064869

    URL https://api.semanticscholar. org/CorpusID:268064869. Olsen, T. H., Boyles, F., and Deane, C. M. Observed an- tibody space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences. Protein Science, 31(1):141–146, 2022a. Olsen, T. H., Moal, I. H., and Deane, C. M. Ablang: an antibody language model for complet- in...

    2001

  5. [7]

    org/CorpusID:236971421

    URL https://api.semanticscholar. org/CorpusID:236971421. Prillo, S., Wu, W., and Song, Y . S. Ultrafast classi- cal phylogenetic method beats large protein language models on variant effect prediction.Advances in neu- ral information processing systems, 37:130265–130290,

  6. [10]

    org/CorpusID:245144689

    URL https://api.semanticscholar. org/CorpusID:245144689. Shanehsazzadeh, A., Bachas, S., McPartlon, M., Kasun, G., Sutton, J. M., Steiger, A. K., Shuai, R., Kohnert, C., Rakocevic, G., Gutierrez, J. M., et al. Unlocking de novo antibody design with generative artificial intelligence. BioRxiv, pp. 2023–01, 2023. Shanker, V . R., Bruun, T. U. J., Hie, B. L....

    2023

  7. [1996]

    org/CorpusID:14356869

    URL https://api.semanticscholar. org/CorpusID:14356869. Gal, Y . and Ghahramani, Z. Dropout as a bayesian ap- proximation: Representing model uncertainty in deep learning, 2016. URL https://arxiv.org/abs/ 1506.02142. Gesell, T. and von Haeseler, A. In silico se- quence evolution with site-specific interactions along phylogenetic trees.Bioinformatics, 22 6:716–22,

    Pith/arXiv arXiv 2016

  8. [2003]

    org/CorpusID:20676562

    URL https://api.semanticscholar. org/CorpusID:20676562. Zhang, Z., Wayment-Steele, H. K., Brixi, G., Wang, H., Kern, D., and Ovchinnikov, S. Protein language models learn evolutionary statistics of interacting sequence mo- tifs.Proceedings of the National Academy of Sciences, 121(45):e2406285121, 2024. 12 Conditionally Site-Independent Neural Evolution of...

    2024

  9. [2006]

    org/CorpusID:260854163

    URL https://api.semanticscholar. org/CorpusID:260854163. Gordon, C., Raghu, A., Greenside, P., and Elliott, H. Gener- ative humanization for therapeutic antibodies, 2024. URL https://arxiv.org/abs/2412.04737. Graves, J., Byerly, J., Priego, E., Makkapati, N., Parish, S. V ., Medellin, B. P., and Berrondo, M. A review of deep learning methods for antibodie...

    Pith/arXiv arXiv 2024

  10. [2014]

    org/CorpusID:16191489

    URL https://api.semanticscholar. org/CorpusID:16191489. Nijkamp, E., Ruffolo, J., Weinstein, E. N., Naik, N., and Madani, A. Progen2: Exploring the boundaries of protein language models, 2022. URL https://arxiv.org/ abs/2206.13517. Nisonoff, H., Xiong, J., Allenspach, S., and Listgarten, J. Unlocking guidance for discrete state-space diffusion and flow mo...

    Pith/arXiv arXiv 2022

  11. [2021]

    org/CorpusID:244060395

    URL https://api.semanticscholar. org/CorpusID:244060395. Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y ., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Can- dido, S., and Rives, A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1...

    doi:10.1126/science.ade2574 2023

  12. [2022]

    org/CorpusID:247596752

    URL https://api.semanticscholar. org/CorpusID:247596752. Raybould, M. I. J., Kovaltsuk, A., Marks, C., and Deane, C. M. Cov-abdab: the coronavirus antibody database.Bioinformatics, 2020. URL https: //api.semanticscholar.org/CorpusID: 218764844. Rodriguez, O. L., Safonova, Y ., Silver, C. A., Shields, K., Gibson, W. S., Kos, J. T., Tieri, D., Ke, H., Jacks...

    Pith/arXiv arXiv 2020

  13. [2023]

    org/CorpusID:263364541

    URL https://api.semanticscholar. org/CorpusID:263364541. Jaffe, D. B., Shahi, P., Adams, B. A., Chrisman, A. M., Finnegan, P. M., Raman, N., Royall, A. E., Tsai, F., V oll- brecht, T., Reyes, D. S., et al. Functional antibodies ex- hibit light chain coherence.Nature, 611(7935):352–357, 2022. Jin, W., Wohlwend, J., Barzilay, R., and Jaakkola, T. Iterative ...

    Pith/arXiv arXiv 2022

  14. [2024]

    org/CorpusID:276318499

    URL https://api.semanticscholar. org/CorpusID:276318499. Ralph, D. K. and Matsen IV , F. A. Inference of b cell clonal families using heavy/light chain pair- ing information.PLOS Computational Biology, 18,

  15. [2025]

    org/CorpusID:277464195

    URL https://api.semanticscholar. org/CorpusID:277464195. Whelan, S. and Goldman, N. A general empiri- cal model of protein evolution derived from multi- ple protein families using a maximum-likelihood ap- proach.Molecular biology and evolution, 18 5:691– 9, 2001. URL https://api.semanticscholar. org/CorpusID:44418374. Wu, X., Feng, J., Komori, A., Kim, E....

    2001