Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Pharmacophore-guided de novo drug design with diffusion bridge

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read PharmacoBridge generates 3D drug-like molecules directly from pharmacophore arrangements using an SE(3)-equivariant diffusion bridge, and its docked hits mostly beat the reference ligand's binding affinity.

desk verdict Sensible application of diffusion bridges to pharmacophore-guided 3D molecule generation, but the headline binding-affinity claim rests on an uncalibrated docking comparison. read the letter →

arxiv 2412.19812 v2 pith:M67NTWLH submitted 2024-12-10 q-bio.BM cs.LG

Conghao Wang , Jagath C. Rajapakse This is my paper · ORCID
classification q-bio.BMcs.LG
keywords denovodrugdesignpharmacophore-guidedgenerationdiffusionbridgeSE(3)-equivariantgraphneuralnetwork3Dmoleculemoleculardockingbindingaffinitystructure-based
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

PharmacoBridge sets out to prove that a pharmacophore — a sparse 3D arrangement of a few chemical feature types — carries enough information to specify a drug-like molecule, and that a diffusion bridge can perform that translation directly in 3D. The paper builds a stochastic bridge pinned at one end to a molecular point cloud and at the other to a pharmacophore point cloud, then trains an SE(3)-equivariant score model to reverse the bridge, turning pharmacophore coordinates into atom coordinates and types. If the reported results hold, drug designers could condition generation on the few interactions that matter for binding instead of on the whole protein pocket, and obtain one-shot 3D hits rather than 2D graphs needing conformer search. On the CrossDocked test set, generated molecules match the conditioning pharmacophores at high rates and, in Gnina docking, most score better than the original ligand on nearly every target.

What carries the argument

The central object is the equivariant denoising diffusion bridge: a stochastic process whose forward law is pinned at both ends, molecule $G_0$ and pharmacophore $G_T$, through Doob's h-transform, and whose reverse ODE samples molecules from pharmacophores. The score $\nabla_{G_t}\log q(G_t|G_T)$ is learned by score matching with a closed-form Gaussian transition kernel $q(G_t|G_0,G_T)=\mathcal{N}(\hat{\mu}_t,\hat{\sigma}_t^2 I)$; the denoiser is an EGNN applied to the concatenated molecular and pharmacophore point clouds, with a mask so only molecule nodes are updated, and the VP bridge with aromatic atom features is the configuration used for the pharmacophore-guided experiments.

What would settle it

Synthesize a sample of generated hits that match their input pharmacophores, measure binding to the intended targets, and compare the measured hit rate with the Gnina ranking; if most high-scoring, pharmacophore-matching molecules fail to bind, the pharmacophore-sufficiency premise is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that molecular generation can be treated as distribution translation between two paired 3D point clouds: a ligand (atom coordinates and types) and a pharmacophore (feature-point coordinates and types). PharmacoBridge uses Doob's h-transform to define a forward bridge that starts at the molecule and is guaranteed to end at the pharmacophore, together with a reverse denoising bridge ODE (Eq. 4) whose score is learned by an EGNN. The paper reports that this recovers the conditioning pharmacophore much better than pocket-conditioned baselines, with average matching scores of 0.71–1.00 across ten structure-based targets, and that Gnina docking finds 76–100% of generated molecules beating the reference ligand's Vina score on nine of the ten targets. Equivariance is handled by centering the combined point cloud and using an SE(3)-equivariant graph neural network, so rotating or translating the pharmacophore rotates or translates the generated molecule in the same way.

Load-bearing premise

The load-bearing premise is that the few positioned feature types of a pharmacophore encode enough of the binding interaction that molecules satisfying those points will actually bind; if shape complementarity, solvation, or induced-fit effects dominate, generated hits could match the pharmacophore yet be inactive.

Editorial extensions

If this is right

  • Pharmacophore conditioning transfers to the generated molecules: average pharmacophore matching scores on ten structure-based targets are 0.71–1.00, well above Pocket2Mol and TargetDiff.
  • Docking favors PharmacoBridge hits over the original ligand: on nine of ten targets, 76–100% of generated molecules get a better Vina score, and on five targets the rate is 98–100%.
  • The generator preserves drug-like chemistry without conditioning: with the VP bridge and aromatic features, 99.96% of sampled molecules are valid, 91.94% unique, and 100% novel, with SA and QED distributions closer to the data than the baselines.
  • Ligand-based design works without protein structures: pharmacophores extracted from known actives alone produce molecules whose pharmacophore recovery far exceeds unconstrained generation and TargetDiff.

Reading between the lines

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

  • Beyond the paper: if a pharmacophore carries the essential interaction information, pocket-based conditioning is largely redundant, so combining a pharmacophore prior with shape or solvation terms could sharpen selectivity while keeping the explicit control.
  • Beyond the paper: the bridge construction is not pharmacophore-specific; the same equivariant translation could pair other 3D inputs and outputs, such as fragment elaboration, scaffold hopping, or ligand–site co-design.
  • Beyond the paper: the 'high binding affinity' claim is a docking prediction; the decisive next step is to synthesize a sample of generated hits and measure binding, checking whether the docking advantage survives wet-lab assay.
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 / 5 minor

Summary. The paper introduces PharmacoBridge, a diffusion-bridge model that maps a pharmacophore point cloud (spatial arrangement of pharmacophore features) to a molecular 3D point cloud, using an SE(3)-equivariant EGNN denoiser. The authors derive a denoising diffusion bridge via Doob's h-transform, train with score matching on paired (molecule, pharmacophore) data from CrossDocked2020, and evaluate on unconditional generation as well as pharmacophore-guided hit design. The pharmacophore-guided evaluation reports that generated molecules match the input pharmacophores substantially better than baselines and, in a structure-based design task, achieve Gnina Vina scores lower (better) than the reference ligand in most targets.

Significance. If the reported results hold, PharmacoBridge is a valuable addition to 3D pharmacophore-conditioned generation: the idea of using a diffusion bridge to translate a pharmacophore into a molecule is natural, the training objective is standard denoising score matching, and the unconditional generation results (validity, uniqueness, SA, QED) are competitive. The theoretical derivation in Section 3 and Appendix A is standard and correct in its main steps, and the model uses well-established equivariant components. However, the paper's central applied claim — that PharmacoBridge generates hit candidates with high binding affinity — rests on a docking comparison (Table 3, Figure 5) whose calibration is unverified, and the sampling algorithm as written contains an integration error. These issues are fixable but currently prevent the evidence from supporting the headline claim at full strength.

major comments (4)
  1. [Section 4.3.2 and Table 3] The high-affinity ratios in Table 3 compare Gnina Vina scores of generated molecules against reference scores of the original ligands 'provided by the CrossDocked dataset.' These two score sets are produced by different docking/scoring pipelines, so the comparison is not calibrated: a systematic offset between Gnina and the CrossDocked pipeline would make the reported ratios an artifact of protocol mismatch rather than genuine superiority. The authors should re-dock the original ligands with the same Gnina invocation (same receptor preparation, box, exhaustiveness, and scoring mode) used for the generated molecules and report both the reference distribution and a parity/calibration check. Additionally, Appendix C.2 states that Vina and CNN scores of 'both generated and original molecules' are shown in Figure 9, which appears to contradict the statement that reference scores come from CrossDocked; this ambiguity must be resolved.
  2. [Algorithm 1] The Heun's second-order correction in Algorithm 1 updates G_{i-1} with the step size (t_{i+1} - t_i), which is the wrong integration interval and references the undefined t_{N+1} when i = N. The correct step size for the update from step i to step i-1 is (t_{i-1} - t_i). As written, the pseudocode does not describe a valid ODE integrator, and this undermines reproducibility of the sampling procedure. Please correct the algorithm and, ideally, provide a reference implementation or pseudocode consistent with the reported experiments.
  3. [Section 4.3.2, Table 3, Figure 5] The claim that 'our method consistently generated molecules with higher binding affinities than the original ligand across each group' is not supported by the table: the high-affinity ratio is 48% for 5LPJ, 76% for 5LSA, 80% for 5FE6, and Pocket2Mol outperforms PharmacoBridge on 5UEV (94.87% vs. 91.00%) and 5FE6 (90.99% vs. 80.00%). The text should be revised to report these exceptions and to present the docking results as a distributional comparison rather than a blanket statement of consistent superiority.
  4. [Section 4.3.1, Figure 4, Table 2] The pharmacophore matching score is measured against the very pharmacophore point cloud used as the conditioning input, so high matching scores are partly guaranteed by construction. The authors should state explicitly that this metric is a controllability/recovery check, not an independent measure of bioactivity. The independent evidence for bioactivity is the docking analysis, and since that analysis has the calibration issue raised above, the paper's overall claim of generating bioactive hit candidates is currently over-strong.
minor comments (5)
  1. [Abstract and Section 1] The word 'phamacophore' appears in the abstract and in Section 1; it should be 'pharmacophore'.
  2. [Algorithm 1] In the drift expression, the score model s_theta is called with the third argument T (e.g., s_theta(G_i, G_N, T)), but the score model is time-dependent and should be evaluated at t_i. This appears to be a typographical error, but it should be corrected for clarity.
  3. [Figure 4 caption] The caption says 'Pharmacophore matching sore distribution'; 'sore' should be 'score'.
  4. [Section 4.1] The sentence 'ensuring molecules with similar structures or biological targets occur either in the training or the sampling dataset' is ambiguous; it should say that such molecules do not occur in both sets.
  5. [Table 3] The TargetDiff sample sizes are very small (3, 8, 9, 10, 12, 16, 20, 28, 32, 45), so the reported high-affinity ratios for that baseline are not directly comparable with the 100-sample evaluations of the other methods. Please add confidence intervals or note the limitation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the core generation is standard score-matched diffusion bridging, and the headline affinity claim rests on external docking rather than on fitted inputs or self-citations.

full rationale

The central derivation is self-contained as a learning objective: Eq. (5) trains s_theta to match the score of the Gaussian bridge kernel q(G_t|G_0,G_T) given in Eq. (6), and sampling uses the denoising bridge ODE in Eq. (4) with the h-transform drift computed from the fixed pharmacophore endpoint. Theorems 3.1 and 3.2 are proved in Appendices A.1 and A.2 from the Fokker-Planck equation, Doob's h-transform, and the SE(3)-equivariance of EGNN; the citations to Zhou et al. (2023) and Peluchetti (2023) are standard background, not self-citations, and are not load-bearing author-specific claims. The only self-citation (Wang et al., 2024) appears in the introduction as one of several examples of gene-expression-conditioned generation and plays no role in the method or evaluation. The pharmacophore matching scores in Sections 4.3.1 and 4.3.2 are fidelity checks: because the model is conditioned on the exact pharmacophore point cloud and the bridge is pinned to that endpoint, the matching metric measures how well generated molecules reinstantiate their own conditioning input. That is an intended controllability property, not a circular derivation of bioactivity; it is not used as the sole evidence for the hit claim. The hit claim in Section 4.3.2 and Table 3 rests on Gnina docking, which is an external program, so the comparison does not reduce to a fitted parameter or to the training objective. A real caveat is that the reference ligand scores are taken from the CrossDocked dataset while generated molecules are re-scored with Gnina, so the two sets may be on different scales; this is an experimental calibration risk, not circularity. Overall, no load-bearing derivation reduces to its own input by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard diffusion bridge theory (Doob's h-transform, Fokker-Planck) and on domain assumptions about pharmacophore sufficiency and the availability of paired data. The only free parameters are the usual hyperparameters of the diffusion model, set by grid search or hand, plus the pharmacophore matching threshold used in evaluation. No new physical entities are introduced.

free parameters (7)
  • sigma0_pos = 0.1
    Grid search over {0.1n | n in [1:10]} in Appendix B.1; controls initial noise for atom positions.
  • sigmaT_pos = 0.3
    Grid search over {0.1n | n in [1:10]} in Appendix B.1; controls terminal noise for atom positions.
  • sigma0_feat = 0.7
    Grid search over {0.1n | n in [1:10]} in Appendix B.1; controls initial noise for node features.
  • sigmaT_feat = 1.0
    Grid search over {0.1n | n in [1:10]} in Appendix B.1; controls terminal noise for node features.
  • omega_h = 10
    Empirically set feature loss weight in Eq. (11) to keep training stable.
  • rho = 7
    Time discretization exponent in Eq. (29), following EDM default.
  • pharmacophore matching threshold = 1.5 angstrom
    Distance threshold for counting a pharmacophore match in evaluation, chosen by the authors in Section 4.3.1.
assumptions (5)
  • standard math Doob's h-transform characterizes the diffusion bridge process used in Eq. (3).
    The denoising bridge ODE (Eq. 4) relies on the h-transform, cited from Doob & Doob 1984 and Zhou et al. 2023; assumed without reproof.
  • standard math Fokker-Planck equation governs the time evolution of conditional densities.
    The proof of Theorem 3.1 uses the Fokker-Planck equation for p(Gt|G0, GT), treated as a standard result.
  • domain assumption Pharmacophore sufficiency: a sparse set of feature points encodes the essential binding interactions.
    The central approach assumes pharmacophore point clouds carry enough information to guide generation of bioactive molecules; if the pharmacophore omits essential interactions, generated molecules can match the points but not bind. Invoked throughout Section 4.3.
  • domain assumption Paired (molecule, pharmacophore) data are available and strictly coupled.
    The method assumes that pharmacophores can be reliably extracted from ligands in CrossDocked2020 and that molecule-pharmacophore pairs define an SE(3)-invariant joint distribution; stated in Section 2.1.
  • domain assumption Gaussian transition kernel is a valid model for the continuous point-cloud representation of discrete molecules.
    The bridge is constructed with q(Gt|G0, GT) Gaussian in Eq. (6); this is a modeling choice that treats molecular structures as continuous clouds, which can produce invalid structures (validity 88-100%).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pharmacophore-guided de novo drug design with diffusion bridge." pith.science (2026). https://pith.science/paper/M67NTWLH

@misc{pith2026241219812,
  author       = {Pith},
  title        = {Pith review of: Pharmacophore-guided de novo drug design with diffusion bridge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M67NTWLH}},
  note         = {Machine review of arXiv:2412.19812}
}
read the original abstract

De novo design of bioactive drug molecules with potential to treat desired biological targets is a profound task in the drug discovery process. Existing approaches tend to leverage the pocket structure of the target protein to condition the molecule generation. However, even the pocket area of the target protein may contain redundant information since not all atoms in the pocket is responsible for the interaction with the ligand. In this work, we propose PharmacoBridge, a phamacophore-guided de novo design approach to generate drug candidates inducing desired bioactivity via diffusion bridge. Our method adapts the diffusion bridge to effectively convert pharmacophore arrangements in the spatial space into molecular structures under the manner of SE(3)-equivariant transformation, providing sophisticated control over optimal biochemical feature arrangements on the generated molecules. PharmacoBridge is demonstrated to generate hit candidates that exhibit high binding affinity with potential protein targets.

Figures

Figures reproduced from arXiv: 2412.19812 by the authors.

Figure 1
Figure 1. Overview of PharmacoBridge. The diffusion bridge process is devised to map the molecule data G0 to the extracted pharma￾cophore data GT via Doob’s h-transforms. Reversely, a score matching model is trained to estimate the score function, which composes the denoising bridge process that recovers molecule data from the pharmacophore data. transition density. 2.2. Diffusion Model with SDEs A diffusion process is to inj… view at source ↗
Figure 2
Figure 2. SA score (↓) distribution. SA scores of the molecules generated by EDM and GruM are concentrated between 6.0 and 8.0, which is obviously larger than the distribution of the original dataset. The SA scores achieved by our model are lower than the baselines and more evenly distributed like the original dataset distribution. 4.3. Pharmacophore-guided Hit Molecule Design We further demonstrate PharmacoBridge in pharmaco… view at source ↗
Figure 3
Figure 3. QED score (↑) distribution. Molecules generated by our model and EDM exhibit higher QED scores than GruM. The distribution achieved by our model shows similar pattern to the dataset distribution [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Pharmacophore matching sore distribution. Compared with Unconstrained generation and TargetDiff, pharmacophore-guided generation significantly enhanced the matching of pharmacophores extracted from original ligands and generated molecules. Notably, TargetDiff suffered …
Figure 5
Figure 5. Figure 5: Distribution of Vina scores, with lower Vina score representing higher binding affinity. The reference scores given by the original ligands are indicated by red dashed lines in the figure. PharmacoBridge consistently generated molecules with higher binding affinities. …
Figure 6
Figure 6. Figure 6: Flowchart of the EGCL module. ⊗ and ⊕ denote element-wise multiplication and addition, respectively. P represents summation over all neighbors. The Equivariant Graph Convolutional Layer (EGCL) that we used for feature updating is illustrated in [PITH_FULL_IMAGE:figure…
Figure 7
Figure 7. Figure 7: Ablation studies on SA analysis. PharmacoBridge with the VP design achieved similar SA scores with or without aromatic features. PharmacoBridge with the VE design without aromatic features achieved the most similar SA score distribution compared to the dataset distribu…
Figure 8
Figure 8. Figure 8: Ablation studies on QED analysis. Similar to SA analysis, PharmacoBridge with the VP design achieved comparative QED scores with or without aromatic features. PharmacoBridge with the VE design with aromatic features achieved the highest range of QED scores, but the QED…
Figure 9
Figure 9. Figure 9: Molecules generated with the pharmacophore models of ligands from PDB structure 5JOY, 4CPI, 1DOD. Docked binding complexes and chemical structures, together with the Vina and CNN scores of both generated and original molecules are shown. 18 [PITH_FULL_IMAGE:figures/fu…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 10 canonical work pages

  1. [1]

    SE (3)-equivariant Network As illustrated in Eq

    B.4. SE (3)-equivariant Network As illustrated in Eq. (8) and (9), the target of our score matching model is finalized to predict the mixture of signal and noise with Fθ, which is, in our case, to predict the molecular point cloud with certain noise. A molecular point cloud is defined as g = (xmol, hmol), where xmol ∈ RN ×3 represents the coordinates of N...

    work page 2021

  2. [2]

    and Kipf, T

    De Cao, N. and Kipf, T. Molgan: An implicit genera- tive model for small molecular graphs. arXiv preprint arXiv:1805.11973,

    arXiv

  3. [4]

    To meet T (3)-invariance, we calculate the center of mass of the system (g, Γ) and move the center to zero

    are apparently SO(3)-equivariant. To meet T (3)-invariance, we calculate the center of mass of the system (g, Γ) and move the center to zero. Thus, we have a diffusion process with equivariant transition density p(GT |Gt). Based on this equivariant diffusion process, we can further devise the equivariant denoising diffusion bridge: Theorem 3.2 (Equivarian...

    work page 2023

  4. [5]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

    arXiv

  5. [6]

    dij = ∥xi −xj∥ is the euclidean distance between nodei and j, and rij = xl i −xl j is the vector difference between node i and j

    At l-th 15 PharmacophoreBridge layer, the updating procedure is given by mij = ϕe(hl i, hl j, d2 ij, aij), hl+1 i = hl i + ϕh(hl i, X j∈N (i) eijmij), xl+1 i = xl i + X j∈N (i) rijϕx(mij) · Mmol i (32) where i and j refer to the node index. dij = ∥xi −xj∥ is the euclidean distance between nodei and j, and rij = xl i −xl j is the vector difference between ...

    work page 2023

  6. [7]

    Bridging Geometric States via Geometric Diffusion Bridge

    Luo, S., Xu, Y ., He, D., Zheng, S., Liu, T.-Y ., and Wang, L. Bridging geometric states via geometric diffusion bridge. arXiv preprint arXiv:2410.24220,

    work page Pith review arXiv

  7. [8]

    Learning to Extend Molecular Scaffolds with Structural Motifs

    Maziarz, K., Jackson-Flux, H., Cameron, P., Sirockin, F., Schneider, N., Stiefl, N., Segler, M., and Brockschmidt, M. Learning to extend molecular scaffolds with structural motifs. arXiv preprint arXiv:2103.03864,

    work page Pith review arXiv

  8. [9]

    Non-denoising forward-time diffusions

    Peluchetti, S. Non-denoising forward-time diffusions. arXiv preprint arXiv:2312.14589,

    arXiv

Show all 26 references
  1. [10]

    9 PharmacophoreBridge Structure-based drug design with equivariant diffusion models

    Schneuing, A., Du, Y ., Harris, C., Jamasb, A., Igashov, I., Du, W., Blundell, T., Li´o, P., Gomes, C., Welling, M., et al. 9 PharmacophoreBridge Structure-based drug design with equivariant diffusion models. arXiv preprint arXiv:2210.13695,

    arXiv

  2. [11]

    and Kim, W

    Seo, S. and Kim, W. Y . Pharmaconet: Accelerating structure- based virtual screening by pharmacophore modeling. arXiv preprint arXiv:2310.00681,

    arXiv

  3. [13]

    and Komodakis, N

    Simonovsky, M. and Komodakis, N. Graphvae: Towards generation of small graphs using variational autoencoders. In Artificial Neural Networks and Machine Learning– ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceed...

    2018

  4. [16]

    Dual diffusion implicit bridges for image-to-image translation

    Su, X., Song, J., Meng, C., and Ermon, S. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382,

    arXiv

  5. [17]

    Digress: Discrete denoising diffusion for graph generation

    Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V ., and Frossard, P. Digress: Discrete denoising diffusion for graph generation. arXiv preprint arXiv:2209.14734,

    arXiv

  6. [19]

    Geodiff: A geometric diffusion model for molecular con- formation generation

    Xu, M., Yu, L., Song, Y ., Shi, C., Ermon, S., and Tang, J. Geodiff: A geometric diffusion model for molecular con- formation generation. arXiv preprint arXiv:2203.02923,

    arXiv

  7. [20]

    Denoising dif- fusion bridge models

    Zhou, L., Lou, A., Khanna, S., and Ermon, S. Denoising dif- fusion bridge models. arXiv preprint arXiv:2309.16948,

    arXiv

  8. [21]

    Proofs A.1

    10 PharmacophoreBridge A. Proofs A.1. Derivation of the Denoising Diffusion Bridge Applying Doob’s h-transforms to the diffusion process with the law of (1) results in a diffusion bridge solving Eq. (3), i.e., dGt = [f (Gt, t) +g(t)2∇Gt log p(GT |Gt)]dt + g(t)dwt, which is pin...

    2023

  9. [22]

    Starting from here, we provide the proof of Theorem 3.1

    have proven that the time evolution of the conditional probability fixed at both ends p(Gt|G0 = g, GT = Γ) follows the Fokker-Planck equation which amounts to the h-transformed diffusion bridge SDE (3). Starting from here, we provide the proof of Theorem 3.1. ∂ ∂t p(Gt|G0 = g,...

    2024

  10. [24]

    Design of VP and VE bridges f (Gt, t) g2(t) p (Gt | G0) SNR t ∇Gt log p(GT |Gt) VP − 1 2 βtGt βt N αtG0, σ2 t I α2 t /σ2 t (αt/αT )GT −Gt σ2 t (SNRt/SNRT −1) VE 0 d dt σ2 t N G0, σ2 t I 1/σ2 t GT −Gt σ2 T −σ2 t B.2. Scaling Functions of Denoiser Dθ Following DDBM (Zhou et al.,...

    2023

  11. [1988]

    J., McKee, M., and Luyten, J

    Wouters, O. J., McKee, M., and Luyten, J. Estimated re- search and development investment needed to bring a new medicine to market, 2009-2018. Jama, 323(9):844–853,

    2009

  12. [2018]

    R., Pariset, M., Hsieh, Y .-P., Martinez, M

    Somnath, V . R., Pariset, M., Hsieh, Y .-P., Martinez, M. R., Krause, A., and Bunne, C. Aligned diffusion schr ¨odinger bridges. In Uncertainty in Artificial Intelligence , pp. 1985–1995. PMLR,

    1985

  13. [2019]

    P., Kumar, A., Er- mon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456,

    2011 arXiv

  14. [2020]

    W., Peng, X., Su, Y ., Peng, J., and Ma, J

    Guan, J., Qian, W. W., Peng, X., Su, Y ., Peng, J., and Ma, J. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543,

    arXiv

  15. [2021]

    Chen, T., Liu, G.-H., and Theodorou, E. A. Likelihood training of schr\” odinger bridge using forward-backward sdes theory. arXiv preprint arXiv:2110.11291,

    arXiv

  16. [2022]

    Jo, J., Kim, D., and Hwang, S. J. Graph generation with dif- fusion mixture. arXiv preprint arXiv:2302.03596,

    arXiv

  17. [2023]

    Graphaf: a flow-based autoregressive model for molec- ular graph generation

    Shi, C., Xu, M., Zhu, Z., Zhang, W., Zhang, M., and Tang, J. Graphaf: a flow-based autoregressive model for molec- ular graph generation. arXiv preprint arXiv:2001.09382,

    2001 arXiv

  18. [2024]

    Let us build bridges: Understanding and extending diffusion generative models

    Liu, X., Wu, L., Ye, M., and Liu, Q. Let us build bridges: Understanding and extending diffusion generative models. arXiv preprint arXiv:2208.14699,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.