Pith. sign in

REVIEW 4 major objections 5 minor 44 references

3D molecule generation needs no equivariant layers: a plain causal transformer over local-frame position tokens reaches ~50% valid yield, and a data-free xTB-rewarded RL loop lifts it to ~95%.

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 21:55 UTC pith:EGP7SB23

load-bearing objection Solid, honest, and reproducible architecture paper; the 95% yield claim should be labeled 'xTB-valid' not 'chemically valid' — the authors say as much in the body, but the abstract oversells. the 4 major comments →

arxiv 2607.15918 v1 pith:EGP7SB23 submitted 2026-07-17 physics.comp-ph

Atomic Design Transformer: Scaffold-Conditioned 3D Molecule Generation via xTB-Reward Reinforcement Learning

Takao Kotani This is my paper
classification physics.comp-ph
keywords Atomic Design Transformer3D molecule generationlocal coordinate framesSE(3) invariancexTB topology preservation (XTP)GFN2-xTBreinforcement learningscaffold-conditioned generation
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 claims that direct 3D molecule generation does not require equivariant network layers — a design assumption much of the field has treated as necessary. Its answer is to describe a molecule as a stream of discrete tokens in which each new atom's position is expressed in the local coordinate frame of an already-placed parent atom; the stream is SE(3)-invariant by construction and carries the molecule's complete bond graph with no bond-order assignment. A plain causal transformer learns this token grammar. To judge its own output, the paper introduces XTP, the fraction of generated molecules whose declared bond topology survives a GFN2-xTB relaxation, and then uses XTP as a reward in a data-free reinforcement-learning loop. The pretrained model reaches about 54% XTP and about 50% end-to-end valid yield on seven scaffold-conditioned benchmarks; the RL-refined model reaches about 98% XTP and about 95% distinct-valid yield while roughly preserving the reference size and element composition. An auxiliary inverse-kinematics transformer corrects the per-step placement error that accumulates in molecules beyond roughly 34 heavy atoms.

Core claim

In the paper's own terms: a molecule's 3D structure and bond graph can be fully specified by a discrete local-frame token stream — each new atom placed in its parent's frame, so the stream is SE(3)-invariant by construction and carries the graph with no bond orders. A plain causal transformer with no equivariant layers learns the grammar from GEOM-Drugs and reaches ~54% XTP and ~50% yield on seven scaffold-conditioned benchmarks of 10,000 samples. A data-free RL stage (RLVR) rewarded by GFN2-xTB topology preservation lifts XTP to ~98% and yield to ~95%, keeping size/composition near the reference. An Inverse-Kinematics Transformer restores XTP where per-step error accumulates for larger mole

What carries the argument

The load-bearing mechanism is the local-frame tokenizer–detokenizer round-trip. Each placed atom stores an orthonormal frame from its bond to its parent and the parent's bond to its grandparent; the next atom is a distance bin plus HEALPix direction in that frame, so rigid motions never change the tokens. The transformer is plain: 12 layers, no positional encoding. The XTP funnel: learned heads add hydrogens; GFN2-xTB relaxes first with the declared graph restrained, then freely; acceptance means the free minimum realizes G_ADT. Graded by per-atom strain, that binary reward drives the REINFORCE loop (KL trust region, virtual size/composition targets) behind the ~95% yield.

Load-bearing premise

The load-bearing premise is that a 'valid molecule' is one that passes the paper's own proxy: an xTB GFN2 relaxation that preserves the bond graph declared in the token stream, plus RDKit readability. The headline ~95% yield is literally the share of generated molecules passing that proxy, and the paper itself notes that ~5% of accepted molecules contain ylide-like charged pairs or strained rings that human curation (and likely more accurate quantum chemistry) would reject.

What would settle it

Relax a random sample of RLVR XTP-accepted molecules with a more accurate electronic-structure method (DFT or correlated wavefunction) and count how many keep the declared bond graph as neutral, closed-shell minima. A large failure rate means the xTB verifier is over-permissive and the ~95% yield is inflated; near-total success means the proxy is sound. Cheaper partial check: reject the ylide-like charged pairs, in-ring cumulenes, and arynes the paper flags (~4.9% of accepted molecules) and see how much of 95% remains. Also test reproducibility with the released checkpoint, given the paper's n

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

If this is right

  • A standard causal transformer with no equivariant layers and no positional encoding is enough to emit valid 3D molecular geometry, because invariance lives in the local-frame tokenization, not in the network.
  • Data-free reinforcement learning against GFN2-xTB — no external molecules — raises the share of distinct, topology-preserving, readable molecules from roughly 52% to roughly 95% on every one of seven scaffolds.
  • The RL-refined geometries sit close to the xTB minimum: median relaxation RMSD falls from 0.47 Å to 0.22 Å and median strain from 29.5 to 8.7 kcal/mol, so accepted molecules barely move when relaxed.
  • The Inverse-Kinematics Transformer lifts XTP from 83% to 99% at 40 heavy atoms and from 31% to 80% at 55, showing the long-molecule failure is an accumulation of small per-step discretization errors rather than a fundamental limit.
  • One trained model handles any scaffold as an inference-time token prefix, generating >99% novel molecules and thousands of distinct Bemis–Murcko scaffolds per scaffold without scaffold-specific retraining.

Where Pith is reading between the lines

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

  • The 95% yield is only as strong as the verifier: XTP plus RDKit readability is a proxy, and the paper documents ~4.9% of accepted molecules that GEOM-style curation would reject (ylide-like charged pairs, in-ring cumulenes, arynes). A stricter chemistry filter would shrink the headline number, though the pretrained-versus-RL ordering would likely survive.
  • The 'propose, relax, accept what physics confirms' loop is verifier-agnostic: the same architecture could be trained against a force field, a higher-level quantum method, or a property oracle (binding, solubility, synthesizability), making it a general reward-conditioned 3D-design engine rather than an xTB-specific one.
  • If the tokenization claim transfers, 'put the symmetry in the tokens, not the network' becomes a general recipe: discrete local-frame tokens plus a plain transformer could replace bespoke equivariant architectures in other geometry-generation settings, such as conformer ensembles or materials.
  • The paper's suggested step-by-step size climb — generate, correct with IKT, raise the target size, repeat — is a concrete testable curriculum for pushing past the ~44-atom ceiling the RLVR model currently tops out at; IKT is only demonstrated to repair errors in the 35–55-atom regime.

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

4 major / 5 minor

Summary. The paper introduces ADT, an autoregressive transformer that generates 3D molecules by emitting discrete tokens in local coordinate frames anchored on previously placed atoms, thereby avoiding equivariant network layers. A pretrained model is evaluated on seven scaffold-conditioned generation tasks with N=10,000 each, using a new metric, XTP, defined as the fraction of molecules whose declared bond graph G_ADT survives GFN2-xTB relaxation. A second model, refined by data-free reinforcement learning (RLVR) against an xTB-based reward, is reported to raise XTP from ~54% to ~98% and the end-to-end yield of distinct, topology-preserving, RDKit-readable molecules from ~52% to ~95%. The paper also proposes an Inverse-Kinematics Transformer to mitigate error accumulation for large molecules. Code, checkpoints, and evaluation tools are publicly released.

Significance. If the claims are taken at face value, the central architectural contribution—discrete local-frame tokenization plus a plain causal transformer without positional encoding—is a notable simplification over equivariant diffusion and flow-matching generators, and the RLVR scheme is an interesting example of data-free self-improvement against a quantum-chemical verifier. The paper is also unusually transparent: it publishes code and data, quantifies a full evaluation funnel, and explicitly discloses several limitations, including a ~4.9% tail of chemically unrealistic but XTP-accepted molecules. However, the significance is substantially tempered by the fact that the headline yield is measured with the same verifier that serves as the RL reward, and by the paper's own evidence that this verifier is over-permissive for some charge-separated and strained motifs. The architecture claim is plausible and well supported; the yield claim requires re-framing or additional independent validation.

major comments (4)
  1. [§4.2, §4.3, Table 2, Eq. (13)] The central yield claim is circular. The RLVR reward in Eq. (13) is exactly the XTP criterion, with a strain-dependent shaping term, and the reported Ngen/N in Table 2 is computed by applying the same XTP funnel plus an RDKit readability check. Thus the improvement from ~54% to ~98% XTP and from ~52% to ~95% Ngen/N shows, by construction, that the model learned to satisfy the verifier used in training; it is not an independent test of chemical validity. The abstract and §1 use phrases like 'chemically valid' and 'valid-molecule yield' without this qualification. The authors should either re-label these numbers as 'XTP-valid under the xTB/RDKit funnel' or provide an independent validity assessment (e.g., against a curated set of drug-like structures, or with a stricter post-hoc filter).
  2. [§4.3, §4.4, Table 6(c)] The paper itself documents that 4.9% of the accepted Ngen molecules are ylide-like carbanion–cation pairs, in-ring cumulenes, or arynes—structures that GEOM-Drugs curation excluded. These structures pass the same XTP+RDKit acceptance that defines the reported ~95% yield. Because the reward only checks whether a GFN2-xTB relaxation preserves G_ADT, xTB's approximate treatment of open-shell/charge-separated motifs is exploitable. The manuscript acknowledges this in §4.3, but the acknowledgment does not resolve the load-bearing problem: the headline 'valid-molecule yield' includes molecules the authors themselves characterize as 'unlikely to be real.' At minimum, the paper should report Ngen/N after excluding the block-(c) and block-(b) categories, or incorporate a formal-charge/strained-ring penalty into the reward and re-evaluate.
  3. [§4.3, Appendix D, Eq. (11), Table 5] The claim that RLVR 'approximately preserves the GEOM-Drugs size and composition distributions' is enforced by construction, not emergent. The loss in Eq. (11) includes KL control terms on heavy-atom-count and element-composition marginals, and Appendix D explains that the virtual targets (μ,σ) and f_tgt are stepped so that the realized distribution lands on the GEOM values. Table 5 lists these as active controls. This is a legitimate and transparent mechanism, but the wording 'approximately preserving' should be replaced or qualified: the match is a design target, not an independent prediction of the RLVR model. The paper would be stronger if it reported the size/composition statistics of an RLVR run without L_ctrl, to show the reward's intrinsic bias.
  4. [§3.2, §4.4] The claim that the token stream 'fully specifies a 3D structure together with its chemical-bond graph G_ADT, without any bond-order assignment' is only partially accurate. The stream specifies a heavy-atom skeleton and a bond graph, but not bond orders, valence states, or hydrogen counts; hydrogen completion is delegated to learned MLnH and MLHplacer modules (§4.1, Appendix B). For the purposes of the XTP relaxation protocol this is internally consistent, but the abstract's phrase 'fully specifies' overstates what the model actually determines. Suggest rephrasing to 'fully specifies a heavy-atom 3D skeleton and a bond graph' and moving the hydrogen-completion step into the method description more prominently.
minor comments (5)
  1. [§2] Typos: 'easiliy' and 'relaxiation' should be corrected.
  2. [§4.1] The paragraph after 'the survivor counts ... are tabulated in Section 4.2' ends with the sentence fragment 'Note that' followed by nothing. This appears to be a missing sentence or a formatting error.
  3. [Table 3] The notation 'model/GEOM' in each diversity cell is terse; a footnote or caption sentence explaining that GEOM is randomly sub-sampled to the same Ngen would improve readability.
  4. [§4.5, Figure 7] The title and text say IKT 'recovers XTP' for large molecules, but the red curve still drops to ~80% at 55 heavy atoms and the authors note it does not generalize beyond the trained regime. Consider rephrasing to 'partially recovers' or 'mitigates the decline'.
  5. [Abstract and §4.3] The reported numbers vary slightly between the abstract (~98% XTP, ~95% Ngen/N) and Table 2 (97.7% average XTP, 95.3% average Ngen/N). The abstract should cite the Table 2 averages or state that these are rounded from the scaffold-average rows.

Circularity Check

2 steps flagged

RLVR's headline yield is the optimized XTP reward, and GEOM-matching is enforced by virtual targets; the architecture claim remains independent.

specific steps
  1. fitted input called prediction [§4.3 Eq. (13); §4.1 Eq. (5); Table 2 RLVR block]
    "The reward is the physical XTP criterion (§4.1): a GFN2-xTB relaxation must realize the graph G_ADT the model declared... rτ = {0, not XTP, 0.6 + 0.4 exp(−ε/Tε), XTP} (Eq. 13)... The topology-preserved set NXTP nearly doubles (5422→9771; 54.2%→97.7%)."

    The paper's headline XTP and Ngen/N improvements are computed by the same GFN2-xTB topology-preservation funnel that defines the RL reward: Eq. (13) gives rτ=1 exactly when Eq. (5), G(Xrelaxed)=G_ADT, holds, and Table 2's NXTP/N is the fraction of samples passing that same funnel (with RDKit readability and duplicate removal for Ngen/N). Thus 'RLVR raises XTP to ~98% and Ngen/N to ~95%' is a report of the optimized objective, not an independent prediction or external validation. The paper is transparent that the reward is XTP, but the abstract presents the post-RL numbers as a method result rather than a restatement of the training signal.

  2. self definitional [Appendix D (Virtual targets, Table 5); §4.3 Results; Table 4]
    "Size and composition are well fit to the reference of GEOM-Drugs thanks to Lctrl... The size and composition targets are virtual: we do not fix them at the GEOM-Drugs values but step them so the realized distribution equilibrates there... The realized size (25.4, 5.4) and element marginal (Table 4) thus land on GEOM even under the reward's size/composition pull."

    The claim that RLVR 'approximately preserves the GEOM-Drugs size and composition distributions' is not an emergent or independently tested property: Lctrl contains KL terms whose virtual targets are stepped by a ratchet/PID until the realized heavy-atom-count moment and element marginal match GEOM. The reported GEOM similarity is therefore written into the loss by construction, so reporting it as a preservation result is reporting the control objective rather than a prediction.

full rationale

The ADT architecture/tokenization claim — that a plain causal transformer over discrete local-frame tokens can generate 3D molecules without equivariant layers — is self-contained and not circular: SE(3)-invariance is built into the tokenization (Eqs. 1–3) and is verified by reconstruction and per-bond error distributions. The pretrained model's XTP (~54%) and diversity/novelty numbers are also independent measurements against GEOM-Drugs and are not the reward objective. However, the central RLVR claim has two construction-level reductions. First, the validity reward of Eq. (13) is exactly the XTP acceptance criterion used to report NXTP/N and Ngen/N in Table 2; the 54%→98% XTP and ~95% yield are the policy optimizing its own scoring function, so they cannot serve as an independent test of the method. Second, the GEOM-Drugs size/composition 'preservation' cited in the abstract is enforced by virtual targets in Lctrl (Appendix D), so the match is a control term, not a discovered property. The paper's own caveat that ~4.9% of accepted molecules are ylide-like charge-separated species, in-ring cumulenes, or arynes is not itself a circularity, but it shows that the optimized verifier is permissive, which amplifies the first issue: the reward-satisfaction numbers include molecules a chemist would reject. There are no load-bearing self-citations or imported uniqueness theorems. Score 6 reflects that the central yield claim reduces by construction to the reward and control objectives, while the architecture and pretrained-model evaluation remain independent content.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

All central numbers are defined through GFN2-xTB relaxation (XTP, strain, RMSD) and RDKit readability; the GEOM-Drugs pretraining set was itself relaxed by GFN2-xTB. Several discrete choices (distance bins, HEALPix Nside, bond cutoff, RL controls) are hand-set, and the RLVR size/composition targets are stepped to GEOM moments, so 'preserving distributions' is partly enforced rather than emergent. No new physical entities are introduced.

free parameters (6)
  • Distance-bin grid R_min/R_max/B = 0.80 Å / 2.50 Å / 200 bins (log-spaced)
    Hand-set discretization of bond length used in every ADD/LINK token; resolution limits geometric accuracy and hence XTP.
  • HEALPix direction resolution Nside = Nside=16, 3072 pixels (~3.7° edges)
    Hand-set angular discretization of local-frame direction; drives the large-molecule error accumulation that IKT is introduced to fix.
  • Parent-offset truncation max_offset = 50
    Hand-set truncation on the 50-way softmax; the paper says it is a practical convenience and may bias parent choice for long molecules.
  • Bond-perception cutoff factor = 1.3 × (Cordero covalent radii), Eq. (10)
    Hand-set threshold that defines the perceived graph G(X) used in the XTP equality G(X_relaxed)=G_ADT; paper claims a gap but does not derive it from physics.
  • RLVR reward strain temperature T_epsilon = 2.0 kcal/mol per atom
    Hand-set reward-shaping constant in Eq. (13); affects which XTP-passing molecules score higher and thus shapes the final distribution.
  • RLVR control weights and virtual targets = lambda_t=0.1, lambda_s=1.0, lambda_c PID base 2.0, scaffold floor 0.96; size target stepped to (25.4, 5.4), composition
    Chosen and stepped controls in Eq. (11) and Table 5. The size and composition matching to GEOM-Drugs is fitted by the virtual-target ratchet, not an emergent property of xTB RL.
axioms (6)
  • domain assumption GFN2-xTB is an adequate physical verifier for drug-like organic molecules.
    XTP, strain, RMSD, and the GEOM-Drugs reference energies are all defined by GFN2-xTB; if xTB mis-relaxes or accepts unphysical minima, every headline number is affected (paper's own caution in §1 and §4.3, ref. [38]).
  • domain assumption The Eq. (10) distance-cutoff graph perception with 1.3×(Cordero radii) recovers the true chemical bond graph.
    XTP compares G_ADT with perceived G(X_relaxed); the paper states the gap never closes on their data, but the rule is heuristic (Appendix C).
  • domain assumption MLnH and MLHplacer hydrogen completion is correct and lossless.
    The funnel's hydrogen stages are asserted to have N_noclash=N_nH=N_Hplacer; only a short Appendix B description is given, so this premise cannot be independently checked.
  • domain assumption A single ground-state GFN2-xTB conformer per SMILES represents the molecule for pretraining.
    GEOM-Drugs is ingested as one lowest-energy conformer per SMILES; conformer diversity is not modeled (§4.2, Figure 5).
  • ad hoc to paper A causal transformer without positional encoding can recover sequence position from the 7-slot token grammar and causal mask.
    The paper's 'no positional encoding' claim (§4) is an architectural hypothesis validated only by training curves; no proof is offered that order information is unambiguous.
  • domain assumption Random-frontier free-order training over rooted spanning trees preserves the 3D structure and gives the model enough order diversity.
    Every molecule has combinatorially many tokenizations; the model is assumed to learn invariant placement from this augmentation (§4).

pith-pipeline@v1.3.0-alltime-deepseek · 23857 in / 19409 out tokens · 203735 ms · 2026-08-01T21:55:57.521883+00:00 · methodology

0 comments
read the original abstract

We present an SE(3)-invariant transformer for 3D-molecule generation, the Atomic Design Transformer (ADT). ADT places atoms one at a time, autoregressively. SE(3) invariance is achieved by tokenization: each new atom's position is encoded in the local coordinate frame of a previously placed atom. The backbone is a plain causal transformer. The token stream fully specifies a 3D structure together with its chemical-bond graph G, without any bond-order assignment. To score generated molecules we introduce the xTB topology-preservation rate (XTP): the fraction of molecules for which an xTB GFN2 relaxation preserves G specified by the token stream. For XTP-accepted molecules we also report the relaxation energy and the root mean square of the atomic displacement (RMSD). We evaluate two ADT models. The first is ADT pretrained on the GEOM-Drugs $\le\!30$-heavy-atom dataset; we benchmark scaffold-conditioned 3D generation across seven drug-like scaffolds from the model. It reaches an XTP of ${\sim}54\%$ and a valid-molecule yield $N^{\mathrm{gen}}/N$ of ${\sim}50\%$, where $N^{\mathrm{gen}}/N$ is the fraction of samples that are distinct, topology-preserving, and chemically valid. The second model continues from the first by reinforcement learning against the verifiable xTB reward (RLVR), using no external molecules. RLVR raises XTP to ${\sim}98\%$ and $N^{\mathrm{gen}}/N$ to ${\sim}95\%$, while approximately preserving the GEOM-Drugs size and composition distributions. Finally, we present an Inverse-Kinematics Transformer that recovers XTP for large molecules, where discretization error accumulates. ADT thus enables direct 3D generation.

Figures

Figures reproduced from arXiv: 2607.15918 by Takao Kotani.

Figure 1
Figure 1. Figure 1: Scaffold-conditional generation, one atom at a time. ADT is prompted with a fixed seed scaffold (here the benzene ring) and then autoregressively places one heavy atom per step; the orange circle marks the atom added at that step (selected steps shown: 1, 3, 6, 9, 12, and the final molecule at step 15). The model decides when to stop, terminating generation on its own. See https://tkotani.github.io/ADT/ben… view at source ↗
Figure 2
Figure 2. Figure 2: ADT pipeline as a two-row horizontal data flow. Top row (left → right): 3D coordinates with atomic species → tokenizer → token sequence (t1, . . . , tN ) → encoder (each token ti is embedded into its own 768-dim hidden vector, so the transformer sees a sequence of length N, not N/7). Pullback (right): hidden vectors flow down through the causal transformer into the decoder, which emits next-token logits. B… view at source ↗
Figure 3
Figure 3. Figure 3: Local coordinate frame Fk at atom Ak. Two ancestor-chain edges u1 = pk − pπ(k) (parent→ Ak) and u2 = pπ(k) − pπ2(k) (grandparent→parent) define an orthonormal triad {e1, e2, e3} via Gram–Schmidt (Eq. 2): e1 is the unit vector along u1, e2 is the residual of u2 orthogonalized against e1, and e3 closes the right-handed frame. When the next atom Ak′ is added, the offset slot selects Ak as its parent, and Ak′ … view at source ↗
Figure 4
Figure 4. Figure 4: The molecule-evaluation funnel (protocol §4.1). A batch of N generated heavy-atom skeletons descends a hydrogen-completion phase (noclash, MLnH, MLHplacer) and a relaxation-and-acceptance phase (H-prerelax, TP Relax) to the XTP-accepted set NXTP; the number beside each arrow is that stage’s survivor count, and the two learned stages are lossless (Nnoclash=NnH=NHplacer). Acceptance is tested against the ADT… view at source ↗
Figure 5
Figure 5. Figure 5: GEOM-Drugs heavy-atom distribution of the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-bond and per-angle structural error introduced by the GFN2-xTB relaxation. Per-bond and per-angle errors. Left: bond-length displacement ∆r = post − pre, binned at 0.01 Å. Right: bond-angle displacement ∆θ = post − pre, binned at 1 ◦ . Both distributions are sharply peaked at zero (median |∆r| = 0.008 / 0.006 Å and |∆θ| = 1.9 ◦ / 1.4 ◦ for pretrained / RLVR). error accumulation as follows. ADT places o… view at source ↗
Figure 7
Figure 7. Figure 7: The Inverse-Kinematics Transformer repairs the large-molecule XTP decline. XTP versus heavy-atom count for ADT (blue) and ADT+IKT (red), with Wilson 95% confidence intervals. Generation is biased toward long molecules to populate the 35–55-atom regime; the shaded histogram (right axis) is the per-bin sample count that sets each interval. ADT alone falls from ∼100% below 34 atoms to 31% at 55. Together thes… view at source ↗
Figure 8
Figure 8. Figure 8: Training curves. (a) Pretraining: training and validation cross-entropy loss versus epoch for the Drugs model (§4.2). (b) RLVR baseline run ( [PITH_FULL_IMAGE:figures/full_fig_p017_8.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

44 extracted references · 2 canonical work pages

  1. [1]

    GEOM, energy-annotated molecular conformations for property prediction and molecular generation

    Simon Axelrod and Rafael Gómez-Bombarelli. GEOM, energy-annotated molecular conformations for property prediction and molecular generation. Scientific Data, 9:185, 2022. doi:10.1038/s41597- 022-01288-4. URL https://arxiv.org/abs/2006. 05531

    doi:10.1038/s41597- 2022

  2. [2]

    Scalable Autoregressive 3D Molecule Gen- eration.arXiv:2505.13791, 2025

    Austin H Cheng, Chong Sun, and Alán Aspuru- Guzik. Scalable Autoregressive 3D Molecule Gen- eration.arXiv:2505.13791, 2025. URL https: //arxiv.org/abs/2505.13791

    Pith/arXiv arXiv 2025

  3. [3]

    InertialAR: Autoregressive 3D Molecule Generation with Inertial Frames

    Haorui Li, Weitao Du, Yuqiang Li, Hongyu Guo, and Shengchao Liu. InertialAR: Autoregressive 3D Molecule Generation with Inertial Frames. arXiv:2510.27497, 2025. URL https://arxiv. org/abs/2510.27497

    Pith/arXiv arXiv 2025

  4. [4]

    Christoph Bannwarth, Sebastian Ehlert, and Stefan Grimme. GFN2-xTB—An accurate and broadly 12 parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions.Journal of Chemical Theory and Computation, 15(3):1652– 1671, 2019. doi:10.1021/acs.jctc.8b01176

    doi:10.1021/acs.jctc.8b01176 2019

  5. [5]

    MiDi: Mixed Graph and 3D De- noising Diffusion for Molecule Generation

    Clément Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. MiDi: Mixed Graph and 3D De- noising Diffusion for Molecule Generation. InECML PKDD, 2023. URLhttps://arxiv.org/abs/2302. 09048

    2023

  6. [6]

    Tuan Le, Julian Cremer, Frank Noé, Djork-Arné Clevert, and Kristof T. Schütt. Navigating the Design Space of Equivariant Diffusion-Based Gener- ative Models for De Novo 3D Molecule Generation. InICLR, 2024. URL https://arxiv.org/abs/ 2309.17296

    Pith/arXiv arXiv 2024

  7. [7]

    SemlaFlow: Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching

    Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. SemlaFlow: Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching. InAISTATS, 2025. URL https: //arxiv.org/abs/2406.07266

    Pith/arXiv arXiv 2025

  8. [8]

    Geometric Representation Con- dition Improves Equivariant Molecule Generation

    Zian Li, Cai Zhou, Xiyuan Wang, Xingang Peng, and Muhan Zhang. Geometric Representation Con- dition Improves Equivariant Molecule Generation. InICML, 2025. URL https://arxiv.org/abs/ 2410.03655

    Pith/arXiv arXiv 2025

  9. [9]

    Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation

    Cai Zhou, Zijie Chen, Zian Li, Jike Wang, Kaiyi Jiang, Pan Li, Rose Yu, Muhan Zhang, Stephen Bates, and Tommi Jaakkola. Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation

  10. [10]

    Toward Better Geomet- ric Representations for Molecule Generative Models

    Shaoheng Yan, Zian Li, Cai Zhou, Qiaojing Huang, Kai Liu, and Muhan Zhang. Toward Better Geomet- ric Representations for Molecule Generative Models

  11. [11]

    Equivariant diffusion for molecule generation in 3D

    Emiel Hoogeboom, Víctor Garcia Satorras, Clément Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3D. InICML, 2022. URL https://arxiv.org/abs/2203.17003

    Pith/arXiv arXiv 2022

  12. [12]

    URLhttps://arxiv.org/abs/2605.07693

    Pith/arXiv arXiv

  13. [13]

    Geometry- Complete Diffusion for 3D Molecule Generation and Optimization.Communications Chemistry, 7,

    Alex Morehead and Jianlin Cheng. Geometry- Complete Diffusion for 3D Molecule Generation and Optimization.Communications Chemistry, 7,

  14. [14]

    Geometric Latent Diffusion Models for 3D Molecule Generation

    Minkai Xu, Alexander S Powers, Ron O Dror, Ste- fano Ermon, and Jure Leskovec. Geometric Latent Diffusion Models for 3D Molecule Generation. In ICML, 2023. URL https://arxiv.org/abs/2305. 01140

    2023

  15. [15]

    3D Equiv- ariant Diffusion for Target-Aware Molecule Genera- tion and Affinity Prediction

    Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3D Equiv- ariant Diffusion for Target-Aware Molecule Genera- tion and Affinity Prediction. InICLR, 2023. URL https://arxiv.org/abs/2303.03543

    Pith/arXiv arXiv 2023

  16. [16]

    Structure-based Drug Design with Equivariant Dif- fusion Models.Nature Computational Science, 4 (12):899–909, 2024

    Arne Schneuing, Yuanqi Du, Charles Harris, Ar- ian Jamasb, Ilia Igashov, Weitao Du, Tom Blun- dell, Pietro Lió, Carla Gomes, Max Welling, et al. Structure-based Drug Design with Equivariant Dif- fusion Models.Nature Computational Science, 4 (12):899–909, 2024. doi:10.1038/s43588-024-00737-x. URLhttps://arxiv.org/abs/2210.13695

    Pith/arXiv arXiv 2024

  17. [17]

    NExT-Mol: 3D Diffusion Meets 1D Language Mod- eling for 3D Molecule Generation

    Zhiyuan Liu, Yanchen Luo, Han Huang, Enzhi Zhang, Sihang Li, Junfeng Fang, Yaorui Shi, Xi- ang Wang, Kenji Kawaguchi, and Tat-Seng Chua. NExT-Mol: 3D Diffusion Meets 1D Language Mod- eling for 3D Molecule Generation. InICLR, 2025. URLhttps://arxiv.org/abs/2502.12638

    Pith/arXiv arXiv 2025

  18. [18]

    Hopcroft, Kun He, and Lijun Wu

    Kaiyuan Gao, Yusong Wang, Haoxiang Guan, Zun Wang, Qizhi Pei, John E. Hopcroft, Kun He, and Lijun Wu. Tokenizing 3D Molecule Structure with Quantized Spherical Coordinates.arXiv preprint arXiv:2412.01564, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules

    Niklas W A Gebauer, Michael Gastegger, and Kristof T Schütt. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. InNeurIPS, 2019. URL https://arxiv.org/abs/ 1906.00957

    Pith/arXiv arXiv 2019

  20. [20]

    Geometry Informed Tokenization of Molecules for Language Model Generation

    Xiner Li, Limei Wang, Youzhi Luo, Carl Edwards, Shurui Gui, Yuchao Lin, Heng Ji, and Shuiwang Ji. Geometry Informed Tokenization of Molecules for Language Model Generation. InInternational Conference on Machine Learning (ICML), 2025. arXiv:2408.10120

    Pith/arXiv arXiv 2025

  21. [21]

    Fragment and Geometry Aware To- kenization of Molecules for Structure-Based Drug Design Using Language Models.arXiv preprint arXiv:2408.09730, 2024

    Cong Fu, Xiner Li, Blake Olson, Heng Ji, and Shuiwang Ji. Fragment and Geometry Aware To- kenization of Molecules for Structure-Based Drug Design Using Language Models.arXiv preprint arXiv:2408.09730, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Language models can generate molecules, materials, and protein binding sites directly in three dimen- sions as XYZ, CIF, and PDB files.arXiv preprint arXiv:2305.05708, 2023

    Daniel Flam-Shepherd and Alán Aspuru-Guzik. Language models can generate molecules, materials, and protein binding sites directly in three dimen- sions as XYZ, CIF, and PDB files.arXiv preprint arXiv:2305.05708, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Symphony: Symmetry- Equivariant Point-Centered Spherical Harmonics for 3D Molecule Generation

    Ameya Daigavane, Song Kim, Mario Geiger, and Tess Smidt. Symphony: Symmetry- Equivariant Point-Centered Spherical Harmonics for 3D Molecule Generation. InICLR, 2024. URL https://arxiv.org/abs/2311.16199

    Pith/arXiv arXiv 2024

  24. [24]

    Reinforcement Learning for Molecular Design Guided by Quantum Mechanics

    GregorN.C.Simm, RobertPinsler, andJoséMiguel Hernández-Lobato. Reinforcement Learning for Molecular Design Guided by Quantum Mechanics. InInternational Conference on Machine Learning (ICML), 2020. arXiv:2002.07717

    Pith/arXiv arXiv 2020

  25. [25]

    Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation

    Zhijian Zhou, Junyi An, Zongkai Liu, Yunfei Shi, Xuan Zhang, Fenglei Cao, Chao Qu, and Yuan Qi. Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation. 2025. URLhttps://arxiv.org/abs/2508.16521. 13

    Pith/arXiv arXiv 2025

  26. [26]

    An Autoregressive Flow Model for 3D Molecular Geometry Genera- tion from Scratch

    Youzhi Luo and Shuiwang Ji. An Autoregressive Flow Model for 3D Molecular Geometry Genera- tion from Scratch. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  27. [27]

    Equivariant 3D-conditional diffusion model for molecular linker design.Na- ture Machine Intelligence, 6(4):417–427, 2024

    Ilia Igashov, Hannes Stark, Clément Vignac, Arne Schneuing, and others. Equivariant 3D-conditional diffusion model for molecular linker design.Na- ture Machine Intelligence, 6(4):417–427, 2024. doi:10.1038/s42256-024-00815-9. URL https:// arxiv.org/abs/2210.05274

    Pith/arXiv arXiv 2024

  28. [28]

    DiffHopp: A Graph Diffusion Model for Novel Drug Design via Scaffold Hop- ping.arXiv:2308.07416, 2023

    Jos Torge, Charles Harris, Simon V Mathis, and Pietro Lió. DiffHopp: A Graph Diffusion Model for Novel Drug Design via Scaffold Hop- ping.arXiv:2308.07416, 2023. URL https:// arxiv.org/abs/2308.07416

    Pith/arXiv arXiv 2023

  29. [29]

    Deep generative models for 3D linker design.Journal of Chemical Information and Modeling, 60(4):1983–1995, 2020

    Fergus Imrie, Anthony R Bradley, Mihaela van der Schaar, and Charlotte M Deane. Deep generative models for 3D linker design.Journal of Chemical Information and Modeling, 60(4):1983–1995, 2020. doi:10.1021/acs.jcim.9b01120

    work page doi:10.1021/acs.jcim.9b01120 1983

  30. [30]

    DiffDec: Structure-Aware Scaffold Decora- tion with an End-to-End Diffusion Model.Journal of Chemical Information and Modeling, 64(7):2554– 2564, 2024

    Junjie Xie, Sheng Chen, Jinping Lei, and Yuedong Yang. DiffDec: Structure-Aware Scaffold Decora- tion with an End-to-End Diffusion Model.Journal of Chemical Information and Modeling, 64(7):2554– 2564, 2024. doi:10.1021/acs.jcim.3c01466

    doi:10.1021/acs.jcim.3c01466 2024

  31. [31]

    Timing AI: LLM-based multimodal hypothesis gen- eration for medicinal chemistry (JST Moonshot final report).JST Moonshot Goal 3 Final Report, 2026

    Keisuke Sakaguchi, Yoshitaka Ushiku, and others. Timing AI: LLM-based multimodal hypothesis gen- eration for medicinal chemistry (JST Moonshot final report).JST Moonshot Goal 3 Final Report, 2026. https://www.jst.go.jp/moonshot/program/ goal3/files/36_ushiku_f_report.pdf

    2026

  32. [32]

    Equivariant Shape-Conditioned Generation of 3D Molecules for Ligand-Based Drug Design

    Keir Adams and Connor W Coley. Equivariant Shape-Conditioned Generation of 3D Molecules for Ligand-Based Drug Design. InICLR, 2023. URL https://arxiv.org/abs/2210.04893

    Pith/arXiv arXiv 2023

  33. [33]

    The properties of known drugs

    Guy W Bemis and Mark A Murcko. The properties of known drugs. 1. Molecular frameworks.Journal of Medicinal Chemistry, 39(15):2887–2893, 1996. doi:10.1021/jm9602928

    doi:10.1021/jm9602928 1996

  34. [34]

    Molecular Sets (MOSES): A Bench- marking Platform for Molecular Generation Mod- els.Frontiers in Pharmacology, 11:565644, 2020

    Daniil Polykovskiy, Alexander Zhebrak, Ben- jamín Sanchez-Lengeling, Sergey Golovanov, Oktai Tatanov, Stanislav Belyaev, Rauf Kurbanov, Alek- sey Artamonov, Vladimir Aladinskiy, Mark Veselov, Artur Kadurin, Simon Johansson, Hongming Chen, Sergey Nikolenko, Alán Aspuru-Guzik, and Alex Zhavoronkov. Molecular Sets (MOSES): A Bench- marking Platform for Molec...

    arXiv 2020

  35. [35]

    GEOM energy-annotated molecular conforma- tions (dataset), 2022

    Simon Axelrod and Rafael Gómez-Bombarelli. GEOM energy-annotated molecular conforma- tions (dataset), 2022. URL https://dataverse. harvard.edu/api/access/datafile/4327252

    arXiv 2022

  36. [36]

    Extended- connectivity fingerprints.Journal of Chemical Information and Modeling, 50(5):742–754, 2010

    David Rogers and Mathew Hahn. Extended- connectivity fingerprints.Journal of Chemical Information and Modeling, 50(5):742–754, 2010. doi:10.1021/ci100050t

    doi:10.1021/ci100050t 2010

  37. [37]

    Pre- diction of physicochemical parameters by atomic contributions.Journal of Chemical Informa- tion and Computer Sciences, 39(5):868–873, 1999

    Scott A Wildman and Gordon M Crippen. Pre- diction of physicochemical parameters by atomic contributions.Journal of Chemical Informa- tion and Computer Sciences, 39(5):868–873, 1999. doi:10.1021/ci990307l

    work page doi:10.1021/ci990307l 1999

  38. [38]

    Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of Cheminformatics, 7:20, 2015

    Dávid Bajusz, Anita Rácz, and Károly Héberger. Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of Cheminformatics, 7:20, 2015. doi:10.1186/s13321- 015-0069-3

    doi:10.1186/s13321- 2015

  39. [39]

    ChEMBL: towards direct depo- sition of bioassay data.Nucleic Acids Research, 47 (D1):D930–D940, 2019

    David Mendez et al. ChEMBL: towards direct depo- sition of bioassay data.Nucleic Acids Research, 47 (D1):D930–D940, 2019. doi:10.1093/nar/gky1075

    doi:10.1093/nar/gky1075 2019

  40. [40]

    HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere

    Krzysztof M Górski, Eric Hivon, Anthony J Ban- day, Benjamin D Wandelt, Frode K Hansen, Martin Reinecke, and Matthias Bartelmann. HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere. The Astrophysical Journal, 622(2):759–771, 2005. doi:10.1086/427976. URL https://arxiv.org/ abs/astro-ph/0409513

    Pith/arXiv arXiv 2005

  41. [41]

    Ex- tended tight-binding quantum chemistry methods

    Christoph Bannwarth, Eike Caldeweyher, Sebas- tian Ehlert, Andreas Hansen, Philipp Pracht, Jakob Seibert, Sebastian Spicher, and Stefan Grimme. Ex- tended tight-binding quantum chemistry methods. WIREs Computational Molecular Science, 11(2): e1493, 2021. doi:10.1002/wcms.1493

    doi:10.1002/wcms.1493 2021

  42. [44]

    Platero- Prats, Marc Revés, Jorge Echeverría, Eduard Cre- mades, Flavia Barragán, and Santiago Alvarez

    Beatriz Cordero, Verónica Gómez, Ana E. Platero- Prats, Marc Revés, Jorge Echeverría, Eduard Cre- mades, Flavia Barragán, and Santiago Alvarez. Co- valent radii revisited.Dalton Transactions, pages 2832–2838, 2008. doi:10.1039/b801115j. Appendix A Token format details A.1 Log-spaced distance binr b A continuous bond distancer (in Å) is discretized into on...

    doi:10.1039/b801115j 2008

  43. [2024]

    URLhttps://arxiv.org/abs/2302.04313

    Pith/arXiv arXiv

  44. [2026]

    URLhttps://arxiv.org/abs/2602.15022

    arXiv