Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Monte Carlo sampling of a molecule's own Gaussian density gives unbiased exact shape overlap and union volumes.

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 09:26 UTC pith:NSCH3ZWU

load-bearing objection The union-overlap estimator is clean and correct; the screening benchmark is overstated as 'complete DUD-E' when it used at most 1000 decoys per target — fix that and this is a solid methods paper worth refereeing. the 3 major comments →

arxiv 2607.20766 v1 pith:NSCH3ZWU submitted 2026-07-22 physics.chem-ph q-bio.QMstat.AP

Importance-Sampling Estimation of Gaussian Molecular Shape Overlap: Exact Union Volumes and Confidence-Bounded Virtual Screening

Egor I. Tuzharov , Alexandra V. Bochenkova , Yury Maximov This is my paper
classification physics.chem-ph q-bio.QMstat.AP
keywords Gaussian shape overlapimportance samplingMonte Carlounion volumeshape Tanimotovirtual screeningmolecular shapeuncertainty quantification
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 introduces the first stochastic, Monte Carlo estimator for the overlap between Gaussian descriptions of molecules. Instead of truncating an inclusion–exclusion expansion at first order, it samples points from one molecule's own Gaussian mixture and reweights them to estimate the exact union volume. The estimator is unbiased, carries an analytic standard error, and costs O(N) per sample. On drug-like molecules it matches dense grid integration to a mean relative error of 0.07%, while the standard first-order approximation over-counts the true union volume by about 3.4x. The error bars enable adaptive screening that preserves ranking while using 94% fewer samples.

Core claim

The central claim is that exact Gaussian shape overlap can be estimated without combinatorial enumeration by writing it as an expectation under one molecule's Gaussian mixture and correcting with the bounded weight wA = uA/mA. Because uA ≤ mA pointwise, this weight lies between 0 and 1, so the estimator is unbiased for the union of all inclusion–exclusion orders and has finite controlled variance. The same framework reproduces the analytic first-order overlap when the weight is omitted, giving a single algorithm for both the biased and the exact quantity.

What carries the argument

Eq. (8): O_AB^∪ = Z_A E_{q_A}[w_A(r) u_B(r)], where q_A = m_A/Z_A is the normalized Gaussian mixture of molecule A, Z_A is its integral, and w_A(r) = u_A(r)/m_A(r) is the bounded ratio of the true union occupancy to the first-order sum. Sampling r from q_A and averaging w_A·u_B gives an unbiased estimate of the exact union volume; the same fixed samples can be reused for gradients, enabling differentiable alignment and adaptive error control.

Load-bearing premise

The screening result depends on the unstated assumption that the up-to-1000 decoys per target used in the DUD-E evaluation preserve the full decoy distribution's AUROC and ranking; if they do not, the reported enrichment no longer describes the benchmark as named.

What would settle it

Compute the exact union overlap of a small molecule (e.g., benzene or aspirin) by enumerating all inclusion–exclusion terms, then run the importance-sampling estimator many times; if the average estimate deviates from the exact value by more than a few times the reported standard error, the unbiasedness claim fails. A cheaper test is to compare estimator variance to the claimed bound Var ≤ Z_A^2/(4n) on a pair where the weight saturates.

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

If this is right

  • Unbiased absolute Gaussian union volumes become available at Monte Carlo cost, replacing the 3.4x over-count of the first-order approximation.
  • Every shape score can carry a standard error, so screening can stop early for clear matches and misses; the paper reports 94% sample savings at Spearman correlation 0.99.
  • The estimator is differentiable, so gradient-based rigid alignment recovers known superpositions (reported shape-Tanimoto 1.00) and multi-start optimization handles multi-modal objectives.
  • On standard retrospective benchmarks (102 and 15 targets), shape-only enrichment matches typical single-conformer screening results while adding uncertainty estimates.
  • The bounded-weight construction removes the need to choose an inclusion–exclusion order, making the method consistent across molecules of very different sizes.

Where Pith is reading between the lines

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

  • Because the variance is controlled by the bounded weight, the estimator may extend naturally to other integrals over products of Gaussian mixtures, such as shape-complementarity potentials, not just Tanimoto similarity.
  • The 94% sampling reduction is demonstrated on a 12-molecule library; scaling it to full benchmarks would require verifying that the adaptive stopping rule preserves ranking across a broad distribution of similarities.
  • The reported 0.07% agreement with grid quadrature tests self-overlap; cross-molecule overlaps with partial overlap may converge more slowly, and the analytic standard error offers a built-in check of exactly that.
  • If the decoy subsample used in the benchmark preserves the rank distribution of the full decoy set, the mean AUROC is representative; if not, the result is a proxy rather than a benchmark answer.

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 / 4 minor

Summary. The paper introduces an unbiased importance-sampling estimator for Gaussian molecular shape overlap. The overlap is re-expressed as an expectation under a molecule's own Gaussian mixture (Eq. 5), and the exact union volume is estimated by reweighting with w_A = u_A/m_A (Eq. 8), giving O(N) cost per sample and an analytic standard error. The authors report 0.07% mean relative error against grid quadrature on 12 molecules, 94% sample savings in an adaptive screening demo, differentiable rigid alignment, and DUD-E/LIT-PCBA enrichment results. The mathematical core is simple and correct; the main weaknesses are in the scope and presentation of the empirical screening validation.

Significance. If substantiated, the paper is a useful contribution: it provides a principled stochastic route to the exact Gaussian union volume, removes the systematic first-order over-count, supplies per-estimate error bars, and enables adaptive sample allocation. The unbiasedness of Eqs. (5) and (8) follows directly from the definitions, and the bounded importance weight keeps the variance finite. The JAX implementation and released code are strengths. However, the screening benchmark is currently described in a way that overstates what was actually run, and one reported comparison stat is internally inconsistent; these issues need correction before the screening claims can be accepted.

major comments (3)
  1. [Results — Retrospective virtual screening; Abstract] The text states the evaluation was run on the 'complete DUD-E (102 targets, ~22.8k actives / 1.4M decoys)' but then says 'actives, and up to 1000 decoys per target, were used.' DUD-E has on average ~13.8k decoys per target, so this uses at most ~7% of the decoys per target. An AUROC computed on 1000 random decoys is an unbiased estimate of the full-decoy AUROC but has non-negligible variance, and per-target values can shift. The abstract and conclusions should either report results on the full decoy set or consistently label the evaluation as 'DUD-E subset with ≤1000 decoys per target.'
  2. [Results — Retrospective virtual screening, Fig. 6c] The sentence 'paired ΔAUROC = −0.01 over nine targets, Fig. 6c' is inconsistent with Fig. 6c, which plots per-target AUROC for the full DUD-E and LIT-PCBA panels (117 targets). If the union-versus-first-order comparison was made on only nine targets, the claim that enrichment is 'statistically unchanged' is underpowered; if it was made on all targets, the number 'nine' is wrong and should be corrected. Please report the actual number of targets and the distribution of paired differences.
  3. [Abstract and Results — Fig. 3] The headline claims of '94% sample savings' and Spearman ρ = 0.99 for confidence-bounded screening are demonstrated on a single 12-molecule library screened against one query (aspirin), not on a general screening benchmark. The abstract presents these numbers without this context. Please qualify them as a proof-of-concept on a small panel, or provide a broader evaluation. This is a scope issue that affects the generality of the contribution as stated.
minor comments (4)
  1. [Results — Fig. 1] Please specify the grid quadrature resolution and convergence criterion used as ground truth. The 12-molecule panel is small; consider stating explicitly that '3.4x' is a panel-specific average, not a universal constant.
  2. [References and benchmarking] The comparison 'cf. ROCS ≈ 0.6 on DUD-E' cites ref. 1 (2007), which predates the DUD-E benchmark (2012) and was not run under the same protocol. This comparison should be removed or properly qualified.
  3. [Confidence-bounded screening] The delta-method standard error SE(T) treats O_AA and O_BB as known constants. If these are estimated rather than precomputed at high precision, the variance formula should include their contribution; otherwise the manuscript should state that they are precomputed.
  4. [LIT-PCBA description] Please clarify whether LIT-PCBA was also subsampled to up to 1000 decoys per target, or whether the full decoy set was used; the current wording is ambiguous.

Circularity Check

0 steps flagged

No circularity: the union-overlap estimator is derived from the definition of u_A and verified against independent grid quadrature.

full rationale

The paper's central estimator, Eq. (8), is obtained by a direct algebraic identity from the definition of the exact union occupancy u_A (Eq. 2) and the normalized Gaussian mixture q_A = m_A/Z_A. No parameter is fitted to make the estimator reproduce its target; the unbiasedness and variance statements follow from standard Monte Carlo theory applied to a bounded integrand. The validation against grid quadrature is an independent external comparison, not a restatement of the estimator. Self-citations (refs 13-17) are used only as background on importance sampling and are not load-bearing: the proof of unbiasedness is self-contained and does not depend on those references. The noted DUD-E evaluation using up to 1000 decoys per target is a benchmark-validity limitation, not a circularity, because it concerns the strength of external screening evidence rather than the derivation of the estimator itself. The paper also explicitly acknowledges its single-conformer, single-query limitation. Thus no circular step is present.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The estimator's unbiasedness is derived from the Gaussian mixture definition with no fitted constants; the empirical claims additionally rely on the Gaussian shape model, on grid quadrature as ground truth, and on the representativeness of a 1000-decoys-per-target DUD-E subsample. The only hand-set numbers are screening tolerances.

free parameters (2)
  • Tanimoto tolerance tau = 0.01
    Hand-chosen in the confidence-bounded screening experiment; the reported 94% sample reduction depends on this threshold and the budget cap.
  • Fixed sample budget cap = 1.2e5 samples per candidate
    Defines the baseline against which the 94% savings is measured; not derived from data.
axioms (6)
  • domain assumption Molecular shape is represented by isotropic Gaussian occupancies o_i(r)=exp(-a_i||r-c_i||^2), with a_i=kappa/(lambda R_i^2), lambda=1 (Eq. 1).
    Inherited from Grant-Pickup/ROCS; the entire overlap and union-volume claim is about this model, not physical hard-sphere volume.
  • domain assumption The 'physically correct molecular occupancy' is the union u_A=1-prod(1-o_i), and shape overlap is integral u_A u_B dr (Eqs. 2-3).
    This definition is taken from prior analytic literature; the estimator is unbiased for this target, so validity of the target model is assumed.
  • standard math Importance-sampling identity: for normalized mixture q_A=m_A/Z_A, integral m_A f dr = Z_A E_{q_A}[f], and the sample mean is unbiased (Eq. 5).
    Textbook MC; requires m_A>=0 and q_A normalizable, both true for Gaussian mixtures.
  • standard math The bounded weight w_A=u_A/m_A lies in [0,1] because 0<=u_A<=m_A pointwise.
    Follows from union bound; gives finite variance bound Var <= Z_A^2/(4n).
  • domain assumption High-resolution grid quadrature of Eq. (3) is an accurate independent ground truth.
    Grid is itself an approximation; the 0.07% agreement claim assumes grid convergence and that the comparison is meaningful.
  • ad hoc to paper Subsampling DUD-E to up to 1000 decoys per target is representative of the full decoy set.
    The paper calls the benchmark 'complete DUD-E' while evaluating a subset; no evidence of representativeness is given, so the screening claim rests on an unstated assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 3746 in / 4220 out tokens · 184989 ms · 2026-08-01T09:26:23.176226+00:00 · methodology

0 comments
read the original abstract

Gaussian descriptions of molecular shape underpin 3D shape-based virtual screening, but existing methods evaluate Gaussian overlap analytically. The widely used first-order approximation is fast but systematically overestimates overlap, whereas the exact molecular volume requires a combinatorial inclusion-exclusion expansion. We introduce the first stochastic estimator of Gaussian shape overlap: an unbiased Monte Carlo method that importance-samples directly from a molecule's Gaussian mixture. The estimator reproduces analytic overlap without bias and extends to the exact union volume of all inclusion-exclusion orders with O(N) cost per sample. On drug-like molecules, the union estimator matches high-resolution grid quadrature with a mean relative error of 0.07 percent, while the first-order approximation overestimates the true union volume by 3.4x on average. The estimator provides analytic standard errors, enabling confidence-bounded screening that reduces sampling by 94 percent while preserving ranking. Implemented in JAX, it is fully differentiable and supports gradient-based rigid alignment on CPU, GPU, and TPU. On the DUD-E and LIT-PCBA benchmarks, the method achieves shape-only enrichment comparable to existing single-conformer approaches while additionally providing unbiased absolute volumes and uncertainty estimates.

Figures

Figures reproduced from arXiv: 2607.20766 by Alexandra V. Bochenkova, Egor I. Tuzharov, Yury Maximov.

Figure 1
Figure 1. Figure 1: (a) Union importance-sampling estimates versus grid ground truth for the self [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Convergence of the union estimator (aspirin self-overlap). (a) Unbiasedness across [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Confidence-bounded screening of a 12-molecule library against aspirin. (a) Adap [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Differentiable rigid-body shape alignment, shown as heavy-atom stick-and-ball [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Superposition of five ibuprofen conformers, coloured by conformer. As embedded [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Retrospective virtual screening. (a) Distribution of per-target AUROC on the [PITH_FULL_IMAGE:figures/full_fig_p014_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

27 extracted references

  1. [1]

    Andrew and Pickup, Barry T

    Grant, J. Andrew and Pickup, Barry T. , title =. J. Phys. Chem. , year =

  2. [2]

    Andrew and Gallardo, M

    Grant, J. Andrew and Gallardo, M. A. and Pickup, Barry T. , title =. J. Comput. Chem. , year =

  3. [3]

    Hawkins, Paul C. D. and Skillman, A. Geoffrey and Nicholls, Anthony , title =. J. Med. Chem. , year =

  4. [4]

    Yan, Xin and Gu, Qiong and Lu, Fang and Li, Jiabo and Xu, Jun , title =. J. Chem. Inf. Model. , year =

  5. [5]

    Yan, Xin and Li, Jiabo and Gu, Qiong and Xu, Jun , title =. J. Comput. Chem. , year =

  6. [6]

    Atwi, Rasha and Wang, Yanfei and others , title =. J. Chem. Inf. Model. , year =

  7. [7]

    Atwi, Rasha and others , title =. J. Chem. Inf. Model. , year =

  8. [8]

    and Pande, Vijay S

    Haque, Imran S. and Pande, Vijay S. , title =. J. Comput. Chem. , year =

  9. [9]

    Jebara, Tony and Kondor, Risi and Howard, Andrew , title =. J. Mach. Learn. Res. , year =

  10. [10]

    Kumar, Ashutosh and Zhang, Kam Y. J. , title =. Front. Chem. , year =

  11. [11]

    and Richards, W

    Ballester, Pedro J. and Richards, W. Graham , title =. J. Comput. Chem. , year =

  12. [12]

    and Blundell, Tom , title =

    Schreyer, Adrian M. and Blundell, Tom , title =. J. Cheminform. , year =

  13. [13]

    , title =

    Riniker, Sereina and Landrum, Gregory A. , title =. J. Chem. Inf. Model. , year =

  14. [14]

    Landrum, Greg and others , title =

  15. [15]

    2018 , howpublished =

    Bradbury, James and Frostig, Roy and Hawkins, Peter and Johnson, Matthew James and Leary, Chris and Maclaurin, Dougal and Necula, George and Paszke, Adam and Vander. 2018 , howpublished =

    2018

  16. [16]

    and Ba, Jimmy , title =

    Kingma, Diederik P. and Ba, Jimmy , title =. Int. Conf. on Learning Representations (ICLR) , year =

  17. [17]

    Mohamed, Shakir and Rosca, Mihaela and Figurnov, Michael and Mnih, Andriy , title =. J. Mach. Learn. Res. , year =

  18. [18]

    , title =

    Owen, Art B. , title =. 2013 , publisher =

    2013

  19. [19]

    Best Arm Identification in Multi-Armed Bandits , booktitle =

    Audibert, Jean-Yves and Bubeck, S. Best Arm Identification in Multi-Armed Bandits , booktitle =. 2010 , pages =

    2010

  20. [20]

    and Carchia, Michael and Irwin, John J

    Mysinger, Michael M. and Carchia, Michael and Irwin, John J. and Shoichet, Brian K. , title =. J. Med. Chem. , year =

  21. [21]

    LIT-PCBA: An Unbiased Data Set for Machine Learning and Virtual Screening , journal =

    Tran-Nguyen, Viet-Khoa and Jacquemard, C. LIT-PCBA: An Unbiased Data Set for Machine Learning and Virtual Screening , journal =. 2020 , volume =

    2020

  22. [22]

    Evaluating Virtual Screening Methods: Good and Bad Metrics for the ``Early Recognition'' Problem , journal =

    Truchon, Jean-Fran. Evaluating Virtual Screening Methods: Good and Bad Metrics for the ``Early Recognition'' Problem , journal =. 2007 , volume =

    2007

  23. [23]

    Electron

    Importance Sampling the Union of Rare Events with an Application to Power Systems Analysis , author =. Electron. J. Stat. , year =

  24. [24]

    Data-Driven Stochastic

    Mitrovic, Mile and Lukashevich, Aleksandr and Vorobev, Petr and Terzija, Vladimir and Budennyy, Semen and Maximov, Yury and Deka, Deepjyoti , journal =. Data-Driven Stochastic. 2023 , volume =

    2023

  25. [25]

    Importance Sampling Approach to Chance-Constrained

    Lukashevich, Aleksander and Gorchakov, Vyacheslav and Vorobev, Petr and Deka, Deepjyoti and Maximov, Yury , journal =. Importance Sampling Approach to Chance-Constrained. 2023 , volume =

    2023

  26. [26]

    Inference and Sampling of K_

    Likhosherstov, Valerii and Maximov, Yury and Chertkov, Misha , booktitle =. Inference and Sampling of K_

  27. [27]

    IEEE Control Syst

    A-Priori Reduction of Scenario Approximation for Automated Generation Control in High-Voltage Power Grids with Renewable Energy , author =. IEEE Control Syst. Lett. , year =