Pith. sign in

REVIEW 3 major objections 5 minor 54 references

RELIABLE-BA fuses four docking engines by modeling each as an evidential expert, scaling its uncertainty by learned context-dependent reliability, and achieves up to 25.7% error reduction through selective prediction.

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

2026-08-01 17:32 UTC pith:NDUZ4DPP

load-bearing objection Sound evidential fusion with correct math, but the empirical claims need an engine-leakage audit and a bigger independent test before I'd trust the calibration gains. the 3 major comments →

arxiv 2607.17601 v1 pith:NDUZ4DPP submitted 2026-07-20 cs.LG cs.AI

Trustworthy Protein-Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion

Yongchan Hong , Defu Cao , Wenjin Liu , Thomas Ku , Jordy Homing Lam , Emily Nguyen , Willie Neiswanger , Vsevolod Katritch
show 1 more author
Yan Liu
This is my paper
classification cs.LG cs.AI
keywords binding affinity predictionuncertainty quantificationevidential regressionNormal-Inverse-Gammamodel fusionvirtual screeningselective predictionprotein-ligand complexes
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.

This paper argues that the right way to combine disagreeing binding-affinity engines is to treat each engine as an evidential expert, learn from the protein-ligand context how much to trust it, and fuse the scaled experts in closed form. The result is a predictive distribution whose aleatoric/epistemic split is calibrated, and whose uncertainty can be used to filter out low-confidence predictions, cutting mean absolute error by up to 25.7% when only the most confident 10% of predictions are kept. A sympathetic reader would care because virtual screening currently lacks confidence scores, so practitioners cannot tell a reliable docking hit from an artifact. If the claim holds, multi-engine screening pipelines gain a cheap post-hoc layer that outputs both affinity and trustworthiness.

Core claim

RELIABLE-BA models each of four engines (GNINA, BIND, FlowDock, DynamicBind) as a Normal-Inverse-Gamma evidential expert, scales each expert's epistemic uncertainty by a context-dependent reliability score learned from protein and ligand embeddings, and fuses them with the MoNIG summation operator. The scaling preserves each expert's predictive mean and aleatoric variance, so the fused distribution's epistemic component rises exactly when engines disagree or an expert is judged unreliable. On the PDBbind time-split test and the independent BDB2020+ set, the paper reports the best or tied-best point accuracy among aggregation methods and the lowest calibration error; selective prediction reac

What carries the argument

The central object is reliability-scaled MoNIG fusion: each engine's score maps through a small MLP to four Normal-Inverse-Gamma parameters; a reliability network maps frozen protein and ligand embeddings to a per-engine reliability score in (0,1); scaling (nu_tilde = r*nu, alpha_tilde = 1 + r*(alpha-1), beta_tilde = r*beta, gamma_tilde = gamma) leaves the predictive mean and aleatoric variance invariant while inflating epistemic variance by 1/r; a closed-form NIG summation then aggregates the scaled experts, with a cross-term in beta adding inter-engine disagreement into the fused epistemic uncertainty.

Load-bearing premise

The load-bearing premise is that the four pretrained docking engines never saw the test protein-ligand complexes during their own training; the paper's time-split and leak-proof datasets shield the fusion and reliability layers but do not audit the engine scores fed into them.

What would settle it

Audit the pretraining corpora of GNINA, BIND, FlowDock, and DynamicBind against the 242-complex time-split test set and the 96-complex BDB2020+ set; if any test complex or a close structural neighbor appears in an engine's training data, remove those complexes and re-run RELIABLE-BA. If the ECE drop and the 25.7% selective-prediction gain vanish or shrink markedly, the gains come from memorized engine scores rather than the reliability mechanism.

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

If this is right

  • Virtual screening pipelines can rank compounds by epistemic uncertainty and retain only high-confidence pairs, reducing MAE by up to 25.7% at 10% coverage.
  • Fusing more diverse engines improves both accuracy and calibration in the paper's tests: ECE drops 81% when going from two to four engines.
  • Calibrated uncertainty transfers to held-out targets: 5HT2A receptor ECE is 0.0233 versus 0.0149 in distribution, and the fused model outperforms all single engines on the SARS-CoV-2 Mpro dataset.
  • The modular design allows new docking engines to be added as evidential experts without retraining the fusion architecture.
  • Because reliability scaling changes only epistemic uncertainty, point predictions remain stable even when engines disagree, so uncertainty does not distort the affinity estimate itself.

Where Pith is reading between the lines

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

  • If any of the four pretrained engines were themselves trained on the test complexes, the reported calibration and selective-prediction gains could be inflated by memorized scores; this is a direct, testable threat to the central claim.
  • The same reliability-scaling idea could generalize to other multi-source regression tasks, wherever several black-box predictors have complementary blind spots and a shared contextual embedding is available.
  • Selective prediction at 10% coverage discards 90% of candidates, so the practical value depends on an explicit cost model for false negatives in a screening campaign; the paper does not quantify that trade-off.
  • The context-dependent reliability scores suggest a testable structure: analyzing reliability per protein family or binding-pocket geometry could reveal which engine dominates which chemical space, potentially guiding engine selection even without fusion.

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

3 major / 5 minor

Summary. The paper proposes RELIABLE-BA, a framework that fuses multiple docking engines for protein-ligand binding affinity prediction. Each engine score is mapped to a Normal-Inverse-Gamma (NIG) evidential distribution; a reliability network, conditioned on frozen ESM-2/ChemBERTa embeddings, modulates each expert's epistemic uncertainty while preserving the predictive mean; the scaled experts are then fused with the Mixture of NIG (MoNIG) aggregation rule. The paper proves algebraic validity of the scaling and fusion, and reports experiments on PDBbind time-split, BDB2020+, SARS-CoV-2 Mpro, and a 5HT2A case study, claiming competitive point accuracy, improved ECE/CRPS/NLL, and up to 25.7% MAE reduction under selective prediction. The authors position RELIABLE-BA as the first multi-engine affinity framework with context-dependent evidential reliability.

Significance. If the empirical claims hold, RELIABLE-BA provides a practical way to combine heterogeneous docking engines into a calibrated predictive distribution with closed-form uncertainty decomposition, which is useful for virtual screening. The manuscript's strengths include 20-seed averaged experiments, several independent benchmarks, a public code release, and a transparent algebraic proof in Appendix A. The main caveat is that the empirical support is conditional on the pretrained engines not having memorized the test complexes; this independence is asserted but not audited. The disagreement-epistemic correlation in Fig. 4 is partly tautological, and the selective-prediction claim would be stronger with explicit baseline comparison at matched retention levels.

major comments (3)
  1. [§4.1.1, §6] The time-split and 'leak-proof' BDB2020+ protocols protect only the fusion and reliability layers. The four pretrained engines listed in §2.1 (GNINA, BIND, FlowDock, DynamicBind) may have been trained on PDBbind-derived corpora that overlap the PDBbind time-split test set or BDB2020+. If an engine has seen a test complex, its score encodes the label, and the reliability network could learn to upweight that engine exactly on those complexes, inflating ECE/CRPS/NLL and the selective-prediction curve. Because Contribution (3) is the paper's headline, the authors should (a) report the exact training data and splits for each engine and demonstrate non-overlap with the test complexes, (b) add an evaluation on a target/dataset provably outside every engine's pretraining data, or (c) substantially temper the generalization claims. The current §6 statement about 'inherent biases or data leakage d
  2. [§4.2.3, Figure 3] The 25.7% MAE reduction at 10% retention is reported only for RELIABLE-BA; the text does not specify the selective-prediction performance of the baselines at matched retention levels. A selective-prediction improvement is meaningful if the proposed uncertainty ranking is shown to dominate, or at least match, existing uncertainty baselines across the coverage spectrum. If Figure 3 already contains baseline curves, the caption and text must identify them and include 20-seed error bands; as written, the claim that the estimates 'enable reliable filtering' is an internal consistency check rather than a demonstration of superiority over the baselines used elsewhere in the paper.
  3. [§4.2.4, Eq. (16), Figure 4] The reported positive correlation between expert disagreement and epistemic uncertainty is a mathematical consequence of the fusion rule. Equation (16) adds the term ½Σν̃ᵢ(γ̃ᵢ−γ)² directly to β, and Var(μ)=β/(ν(α−1)) is therefore increasing in disagreement by construction. Thus Figure 4 (left) cannot be presented as empirical evidence that the model 'appropriately assigns higher epistemic uncertainty' to disagreement; it only confirms the closed-form formula. This should be reframed as a validity check of the implementation, not as an independent empirical finding.
minor comments (5)
  1. [§3.3, Eq. (7)] The text says r_i(x) is constrained away from zero, but the sigmoid output in Eq. (7) can be arbitrarily close to zero. Please specify the clamping value or the exact epsilon used in the implementation, since this affects the numerical validity guarantee in Appendix A.
  2. [Appendix A] Theorem A.4 proves that the reliability-scaled fusion produces a valid NIG distribution, but 'validity' is not the same as 'correctness' as a calibrated Bayesian fusion rule. The paper should avoid the implication that algebraic validity alone justifies the aggregation as a principled uncertainty model.
  3. [Table 5] The hyperparameter sensitivity table shows λ=0.01 gives slightly better MAE (0.814 vs 0.824) than the chosen λ=0.001. The authors choose λ=0.001 because of calibration; please state the explicit selection criterion and whether this decision was made before or after seeing test-set results.
  4. [Abstract / Conclusion] The abstract says 'reducing prediction error by up to 25%' while §5 says 25.7%. Use one consistent rounded value.
  5. [§4.1.1, Appendix B] BDB2020+ contains only 96 retained complexes. This is a very small independent test set, and the reported R² values are highly variable. Please add confidence intervals or a sensitivity analysis, and note the small-sample limitation in the main text.

Circularity Check

1 steps flagged

One RQ4 'finding' is true by construction; the rest of the fusion and benchmark chain is self-contained.

specific steps
  1. self definitional [§4.2.4 (RQ4), Figure 4(left); Eqs. (16)–(17)]
    "Figure 4 (left) shows a clear positive correlation between inter-expert disagreement and RELIABLE-BA’s epistemic uncertainty. As expert predictions diverge, RELIABLE-BA appropriately assigns higher epistemic uncertainty... The fused scale parameter is β = Σ β̃ᵢ + 1/2 Σ ν̃ᵢ(γ̃ᵢ−γ)² (Eq. 16)... Var(μ) = β/(ν(α−1)) (Eq. 17)."

    The claimed empirical correlation is guaranteed by construction. Eq. (17) defines epistemic uncertainty as Var(μ)=β/[ν(α−1)], and Eq. (16) defines β as the sum of per-expert scales plus 1/2Σν̃ᵢ(γ̃ᵢ−γ)², which is exactly the inter-expert disagreement term. Since ν>0 and α>1, Var(μ) is monotonically increasing in that disagreement term for fixed scaled expert parameters. Therefore any divergence in expert means mechanically raises epistemic uncertainty; the 'positive correlation' in Figure 4(left) is an algebraic identity of the model, not an independent empirical validation that the model 'appropriately assigns higher epistemic uncertainty.'

full rationale

The central point-prediction and calibration claims are not circular: RELIABLE-BA is trained on the PDBbind time-split and evaluated on independent BDB2020+, Mpro, and 5HT2A data; the fusion/reliability layers do not see test labels, and the MoNIG operator is cited to non-overlapping prior work (Ma et al. 2021) with algebraic validity proofs that do not assume the target result. The reliability scaling is a design choice (Propositions A.2–A.3), not a disguised prediction. The only circular element is the RQ4 statement that epistemic uncertainty correlates with inter-expert disagreement: this is guaranteed by Eqs. (16)–(17), so it should be presented as a property of the definition rather than as an empirical validation. The self-citations present (e.g., Uncertainty Toolbox) are tool/background citations and are not load-bearing. The engine-level pretraining leakage concern raised by the skeptic is a data-validity risk, not circularity, and is not scored here.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The central claim rests on a few standard probabilistic facts, two domain assumptions about the informativeness and independence of the input features, and one ad hoc scaling rule. The most fragile is the independence/leakage assumption: the evaluation protocol controls the fusion layer but not the pretrained engines. The reliability-scaling rule is an explicit modeling choice, not an inevitable consequence of evidential theory.

free parameters (2)
  • evidential regularization coefficient λ = 0.001
    Appendix D shows calibration (ECE, NLL) varies strongly with λ; 0.001 was selected because it gives the best ECE/NLL with only marginal MAE cost. The reported calibration results are thus partially selected on the validation set.
  • reliability clamping constant for r_i = unspecified
    Section 3.3 states r_i is 'constrained away from zero' to ensure ν̃>0, β̃>0, ᾱ>1, but the exact lower bound is not reported. This constant affects the maximum epistemic uncertainty and is a free choice.
axioms (6)
  • standard math The NIG prior and its Student-t marginal likelihood yield closed-form evidential regression with the stated moment formulas.
    Invoked in §3.2 and §3.5 for expert and fused losses; standard result from Amini et al. [1].
  • standard math The MoNIG summation operator ⊕ defined by Ma et al. [27] preserves conjugacy and is a valid fusion rule for multiple NIG distributions.
    Invoked in §3.4 to define the fused NIG parameters; the paper extends it to reliability-scaled parameters in Appendix A.
  • domain assumption Each engine's scalar score s_i(x) carries information about binding affinity y that can be transformed by an MLP into evidential parameters.
    Eq. (3) maps a single scalar to four NIG parameters. If an engine's score were unrelated to affinity or had non-monotone/LUT behavior not learnable from a scalar, the expert-level loss could not be minimized. Engine scores are not normalized in the paper.
  • domain assumption The ESM-2 + ChemBERTa embeddings h(x) contain enough information to predict which engine is reliable for a given protein–ligand complex.
    Section 3.3 conditions the reliability network on h(x). The ablation gain over uniform reliability is modest (~0.01 MAE), and out-of-distribution generalization is only tested on two small case studies.
  • domain assumption The PDBbind time-split and BDB2020+ are independent of the fusion training without residual leakage through the pretrained engines.
    Section 4.1.1 describes the time-split preventing temporal leakage for the fusion/reliability layers. However, the four base engines' pretraining data are not audited; if those engines saw test complexes, the engine scores encode leakage that the reliability network could exploit.
  • ad hoc to paper The reliability-scaling rules (Eqs. 8–11) are the correct way to translate trustworthiness into evidential parameters.
    The scaling is deliberately chosen to make epistemic uncertainty scale inversely with reliability while preserving the mean and aleatoric variance. It is not derived from a physical or Bayesian first principle; it is a design choice validated only by downstream calibration metrics.

pith-pipeline@v1.3.0-alltime-deepseek · 4387 in / 4655 out tokens · 191815 ms · 2026-08-01T17:32:09.218088+00:00 · methodology

0 comments
read the original abstract

Accurate protein-ligand binding affinity prediction is central to computational drug discovery, yet modern docking engines frequently disagree without indicating which prediction to trust. Consensus scoring and ensemble methods improve mean accuracy but treat all predictions identically without interpretable confidence measures or uncertainty decomposition, ignoring the chemical context of each protein-ligand pair. To address this limitation, we introduce RELIABLE-BA (RELIABiLity-aware Evidential fusion for Binding Affinity), an evidential framework for multi-engine binding affinity prediction. Our model comprises three steps: (1) modeling each engine as an evidential expert via Normal-Inverse-Gamma distributions, (2) scaling epistemic uncertainty through learned reliability from molecular context while preserving each expert's predictive mean, and (3) fusing experts through closed-form aggregation that captures both individual uncertainty and inter-engine disagreement. Experiments on the PDBBind and BDB2020+ benchmarks demonstrate competitive point prediction with substantially improved uncertainty calibration, and additional validation on the SARS-CoV-2 Mpro dataset and 5HT2A receptor demonstrates applicability to clinically relevant drug targets. Crucially, these uncertainty estimates enable reliable filtering of protein-ligand pairs, reducing prediction error by up to 25% when retaining only high-confidence pairs. To our knowledge, RELIABLE-BA is the first multi-engine binding affinity prediction framework to combine evidential fusion with context-dependent reliability, offering a principled path toward trustworthy AI-guided drug discovery. Our code is publicly available at https://github.com/yongchand/RELIABLE-BA.

Figures

Figures reproduced from arXiv: 2607.17601 by Defu Cao, Emily Nguyen, Jordy Homing Lam, Thomas Ku, Vsevolod Katritch, Wenjin Liu, Willie Neiswanger, Yan Liu, Yongchan Hong.

Figure 1
Figure 1. Figure 1: Pairwise error correlations between four binding [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the RELIABLE-BA architecture. Evidential heads map each engine’s score to NIG parameters, which [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Selective predictive analysis on five models. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Correlation between expert disagreement and epis [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE visualization of protein-ligand embeddings [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] 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

54 extracted references · 10 canonical work pages

  1. [1]

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. 2020. Deep evidential regression.Advances in neural information processing systems33 (2020), 14927–14937

    2020

  2. [2]

    Ballester and John B

    Pedro J. Ballester and John B. O. Mitchell. 2010. A machine learning approach to predicting protein–ligand binding affinity with applications to molecular docking. Bioinformatics26, 9 (2010), 1169–1175. doi:10.1093/bioinformatics/btq112

    work page doi:10.1093/bioinformatics/btq112 2010

  3. [3]

    Clara Blanes-Mira, Pilar Fernández-Aguado, Jorge de Andrés-López, Asia Fernández-Carvajal, Antonio Ferrer-Montiel, and Gregorio Fernández-Ballester

  4. [4]

    Leo Breiman. 1996. Stacked regressions.Machine Learning24, 1 (1996), 49–64. doi:10.1007/BF00117832

    doi:10.1007/bf00117832 1996

  5. [5]

    Martin Buttenschoen, Garrett M Morris, and Charlotte M Deane. 2024. Pose- Busters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences.Chemical Science15, 9 (2024), 3130–3139

    2024

  6. [6]

    Defu Cao, James Enouen, Yujing Wang, Xiangchen Song, Chuizheng Meng, Hao Niu, and Yan Liu. 2023. Estimating treatment effects from irregular time series observations with hidden confounders. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 6897–6905

    2023

  7. [7]

    Defu Cao, Wen Ye, Yizhou Zhang, Sam Griesemer, and Yan Liu. 2026. PINFDiT: Energy-Based Physics-Informed Diffusion Transformers for General-purpose Time Series Tasks. InThe Fourteenth International Conference on Learning Repre- sentations. https://openreview.net/forum?id=EphTlUJ4XN

    2026

  8. [8]

    Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. 2020. Chem- BERTa: large-scale self-supervised pretraining for molecular property prediction. arXiv preprint arXiv:2010.09885(2020)

    Pith/arXiv arXiv 2020

  9. [9]

    Youngseog Chung, Ian Char, Han Guo, Jeff Schneider, and Willie Neiswanger

  10. [10]

    Spencer S Ericksen, Haozhen Wu, Huikun Zhang, Lauren A Michael, Michael A Newton, F Michael Hoffmann, and Scott A Wildman. 2017. Machine Learning Consensus Scoring Improves Performance Across Targets in Structure-Based Virtual Screening.Journal of Chemical Information and Modeling57, 7 (2017), 1579–1590. doi:10.1021/acs.jcim.7b00153

    work page doi:10.1021/acs.jcim.7b00153 2017

  11. [11]

    Richard A Friesner, Jay L Banks, Robert B Murphy, Thomas A Halgren, Jasna J Klicic, Daniel T Mainz, Matthew P Repasky, Eric H Knoll, Mee Shelley, Jason K Perry, et al. 2004. Glide: a new approach for rapid, accurate docking and scoring

    2004

  12. [12]

    Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning. PMLR, 1050–1059

    2016

  13. [13]

    Anna Gaulton, Louisa J Bellis, A Patricia Bento, Jon Chambers, Mark Davies, Anne Hersey, Yvonne Light, Shaun McGlinchey, David Michalovich, Bissan Al-Lazikani, et al. 2012. ChEMBL: a large-scale bioactivity database for drug discovery.Nucleic acids research40, D1 (2012), D1100–D1107

    2012

  14. [14]

    Method and assessment of docking accuracy.Journal of Medicinal Chemistry 47, 7 (2004), 1739–1749

    2004

  15. [15]

    Alexander S Hauser, Misty M Attwood, Mathias Rask-Andersen, Helgi B Schiöth, and David E Gloriam. 2017. Trends in GPCR drug discovery: new agents, targets and indications.Nature reviews Drug discovery16, 12 (2017), 829–842

    2017

  16. [16]

    James Hensman, Alexander Matthews, and Zoubin Ghahramani. 2015. Scalable variational Gaussian process classification. InArtificial intelligence and statistics. PMLR, 351–360

    2015

  17. [17]

    David E Graff, Eugene I Shakhnovich, and Connor W Coley. 2021. Accelerating high-throughput virtual screening through molecular pool-based active learning. Chemical science12, 22 (2021), 7866–7881

    2021

  18. [18]

    Tuomo Kalliokoski, Christian Kramer, Anna Vulpetti, and Peter Gedeck. 2013. Comparability of Mixed IC50 Data–A Statistical Analysis.PLoS ONE8, 4 (2013), e61007. doi:10.1371/journal.pone.0061007

    work page doi:10.1371/journal.pone.0061007 2013

  19. [19]

    Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision?. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  20. [20]

    José Jiménez, Miha Skalic, Gerard Martinez-Rosell, and Gianni De Fabritiis. 2018. K deep: protein–ligand absolute binding affinity prediction via 3d-convolutional neural networks.Journal of chemical information and modeling58, 2 (2018), 287–296

    2018

  21. [21]

    Hilbert Yuen In Lam, Jia Sheng Guan, Xing Er Ong, Robbe Pincket, and Yuguang Mu. 2024. Protein language models are performant in structure-free virtual screening.Briefings in Bioinformatics25, 6 (2024), bbae480

    2024

  22. [22]

    Jordy Homing Lam and Vsevolod Katritch. 2025. Navigating structure-based drug discovery with emerging innovations in physics-and knowledge-based approaches.npj Drug Discovery2, 1 (2025), 29

    2025

  23. [23]

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems30

    2017

  24. [24]

    Jie Li, Xingyi Guan, Oufan Zhang, Kunyang Sun, Yingze Wang, Dorian Bagni, and Teresa Head-Gordon. 2026. Leak Proof PDBBind: a reorganized data set 9 Yongchan Hong et al. of protein–ligand complexes for more generalizable binding affinity prediction. The Journal of Physical Chemistry B130, 2 (2026), 730–740

    2026

  25. [25]

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, et al

  26. [26]

    Gregory A Landrum and Sereina Riniker. 2024. Combining IC50 or K i values from different sources is a source of significant noise.Journal of chemical information and modeling64, 5 (2024), 1560–1567

    2024

  27. [27]

    Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. 2021. Trustworthy multimodal regression with mixture of normal- inverse gamma distributions.Advances in Neural Information Processing Systems 34 (2021), 6881–6893

    2021

  28. [28]

    Wesley J Maddox, Pavel Izmailov, Timur Garipov, Dmitry P Vetrov, and An- drew Gordon Wilson. 2019. A simple baseline for bayesian uncertainty in deep learning.Advances in neural information processing systems32

    2019

  29. [29]

    Language models of protein sequences at the scale of evolution enable accurate structure prediction.BioRxiv2022 (2022), 500902

    2022

  30. [30]

    Wei Lu, Qifeng Wu, Jixian Zhang, Jiahua Rao, Chengtao Li, and Shuangjia Zheng

  31. [31]

    Alex Morehead and Jianlin Cheng. 2025. FlowDock: Geometric flow matching for generative protein–ligand docking and affinity prediction.Bioinformatics41, Supplement_1 (2025), i198–i206. doi:10.1093/bioinformatics/btaf187

    doi:10.1093/bioinformatics/btaf187 2025

  32. [32]

    Flower, Subhagata Chattopadhyay, and Amit K

    Do Nhat Phuong, Darren R. Flower, Subhagata Chattopadhyay, and Amit K. Chattopadhyay. 2023. Towards effective consensus scoring in Structure-Based virtual screening.Interdisciplinary Sciences: Computational Life Sciences15, 1 (2023), 131–145. doi:10.1007/s12539-022-00546-8

    work page doi:10.1007/s12539-022-00546-8 2023

  33. [33]

    Fatemeh Panahandeh and Najme Mansouri. 2025. A comprehensive review of neural network-based approaches for drug–target interaction prediction.Molec- ular Diversity(2025), 1–48

    2025

  34. [34]

    Andrew T McNutt, Paul Francoeur, Rishal Aggarwal, Tomohide Masuda, Rocco Meli, Matthew Ragoza, Jocelyn Sunseri, and David Ryan Koes. 2021. GNINA 1.0: molecular docking with deep learning.Journal of cheminformatics13, 1 (2021), 43

    2021

  35. [35]

    Jiffriya Mohamed Abdul Cader, MA Hakim Newton, Julia Rahman, Akmal Jahan Mohamed Abdul Cader, and Abdul Sattar. 2024. Ensembling methods for protein- ligand binding affinity prediction.Scientific Reports14, 1 (2024), 24447

    2024

  36. [36]

    Milad Rayka and S Shahab Naghavi. 2025. Uncertainty quantification enables reliable deep learning for protein–ligand binding affinity prediction.Scientific Reports15, 1 (2025), 43156. doi:10.1038/s41598-025-27167-7

    work page doi:10.1038/s41598-025-27167-7 2025

  37. [37]

    Christopher J Schmidt, Stephen M Sorensen, John H Kenne, Albert A Carr, and Michael G Palfreyman. 1995. The role of 5-HT2A receptors in antipsychotic activity.Life sciences56, 25 (1995), 2209–2222

    1995

  38. [38]

    Ava P Soleimany, Alexander Amini, Samuel Goldman, Daniela Rus, Sangeeta N Bhatia, and Connor W Coley. 2021. Evidential Deep Learning for Guided Molecu- lar Property Prediction and Discovery.ACS central science7, 8 (2021), 1356–1367. doi:10.1021/acscentsci.1c00546

    work page doi:10.1021/acscentsci.1c00546 2021

  39. [39]

    Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Mateo Reveiz, Stephan Thaler, Vignesh Ram Somnath, Noah Getz, Tally Portnoi, Julien Roy, Hannes Stark, et al. 2025. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv(2025). doi:10.1101/2025.06.14.659707

    doi:10.1101/2025.06.14.659707 2025

  40. [40]

    Milad Rayka, Morteza Mirzaei, and Ali Mohammad Latifi. 2024. An ensemble- based approach to estimate confidence of predicted protein–ligand binding affinity values.Molecular Informatics43, 4 (2024), e202300292. doi:10.1002/minf. 202300292

    work page doi:10.1002/minf 2024

  41. [41]

    Oleg Trott and Arthur J Olson. 2010. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading.Journal of Computational Chemistry31, 2 (2010), 455–461

    2010

  42. [42]

    Lingle Wang, Yujie Wu, Yuqing Deng, Byungchan Kim, Levi Pierce, Goran Krilov, Dmitry Lupyan, Shaughnessy Robinson, Markus K Dahlgren, Jeremy Greenwood, et al. 2015. Accurate and reliable prediction of relative ligand binding potency in prospective drug discovery by way of a modern free-energy calculation protocol and force field.Journal of the American Ch...

    2015

  43. [43]

    Renxiao Wang, Xueliang Fang, Yipin Lu, Chao-Yie Yang, and Shaomeng Wang

  44. [44]

    Hannes Stärk, Octavian Ganea, Lagnajit Pattanaik, Regina Barzilay, and Tommi Jaakkola. 2022. Equibind: Geometric deep learning for drug binding structure prediction. InInternational conference on machine learning. PMLR, 20503–20521

    2022

  45. [45]

    Marta M Stepniewska-Dziubinska, Piotr Zielenkiewicz, and Pawel Siedlecki

  46. [46]

    Kristal, and D

    Jinn-Moon Yang, Yen-Fu Chen, Tsai-Wei Shen, Bruce S. Kristal, and D. Frank Hsu. 2005. Consensus Scoring Criteria for Improving Enrichment in Virtual Screening.Journal of Chemical Information and Modeling45, 4 (2005), 1134–1146. doi:10.1021/ci050034w

    work page doi:10.1021/ci050034w 2005

  47. [47]

    Yanpeng Zhao, Yuting Xing, Yixin Zhang, Yifei Wang, Mengxuan Wan, Duoyun Yi, Chengkun Wu, Shangze Li, Huiyan Xu, Hongyang Zhang, et al. 2025. Evi- dential deep learning-based drug-target interaction prediction.Nature communi- cations16, 1 (2025), 6915. A Proof of Validity of Reliability-Scaled MoNIG Aggregation We prove that RELIABLE-BA’s reliability scal...

    2025

  48. [51]

    Zhe Wang, Huiyong Sun, Xiaojun Yao, Dan Li, Lei Xu, Youyong Li, Sheng Tian, and Tingjun Hou. 2016. Comprehensive evaluation of ten docking programs on a diverse set of protein–ligand complexes: the prediction accuracy of sampling power and scoring power.Physical Chemistry Chemical Physics18, 18 (2016), 12964–12975. doi:10.1039/C6CP01555G

    doi:10.1039/c6cp01555g 2016

  49. [52]

    Wenzhe Xu, Xiaorong Liu, Jie Wang, Fan Zhang, Dongfeng Hu, and Liansong Zong. 2025. UAMRL: multi-granularity uncertainty-aware multimodal represen- tation learning for drug-target affinity prediction.Bioinformatics41, 10 (2025), btaf512. doi:10.1093/bioinformatics/btaf512

    work page doi:10.1093/bioinformatics/btaf512 2025

  50. [2005]

    The PDBbind database: methodologies and updates.Journal of medicinal chemistry48, 12 (2005), 4111–4119

    2005

  51. [2018]

    Development and evaluation of a deep learning model for protein–ligand binding affinity prediction.Bioinformatics34, 21 (2018), 3666–3674

    2018

  52. [2021]

    Uncertainty Toolbox: an Open-Source Library for Assessing, Visualizing, and Improving Uncertainty Quantification.arXiv preprint arXiv:2109.10254 (2021)

    Pith/arXiv arXiv 2021

  53. [2022]

    doi:10.3390/molecules28010175

    Comprehensive survey of consensus docking for high-throughput virtual screening.Molecules28, 1 (2022), 175. doi:10.3390/molecules28010175

    work page doi:10.3390/molecules28010175 2022

  54. [2024]

    DynamicBind: predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model.Nature Communications15, 1 (2024), 1071

    2024