pith. sign in

arxiv: 2606.25865 · v1 · pith:HEWZPWX3new · submitted 2026-06-24 · 🧬 q-bio.BM

Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design

Haoyu Lin , Yiyan Liao , Jinmei Pan , Xinliao Ling , Luhua Lai , Jianfeng Pei This is my paper

Pith reviewed 2026-06-25 19:27 UTC · model grok-4.3

classification 🧬 q-bio.BM
keywords molecular generationdrug designmultimodal foundation modelFragment-SELFIESautoregressive decodersupervised fine-tuningproperty-conditioned generationtarget-conditioned generation
0
0 comments X

The pith

A single autoregressive decoder generates valid molecules from scalar properties, protein sequences, or binding pockets by replacing value-token embeddings during fine-tuning.

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

The paper introduces Molexar as a unified model that first pretrains an autoregressive decoder on Fragment-SELFIES to capture molecular syntax and distributions. Supervised fine-tuning then conditions the same decoder on diverse inputs by swapping in value-token embeddings for each mode, keeping all generation on one path. This produces 100 percent validity in unconditional and fragment tasks plus competitive results on property and target-conditioned benchmarks while using fewer parameters than larger alternatives. A sympathetic reader would see the practical payoff as fewer specialized models needed for drug-design constraints.

Core claim

Molexar demonstrates that one small pretrained autoregressive decoder, trained on Fragment-SELFIES and then fine-tuned by in-place replacement of value-token embeddings for scalar properties, pharmacophore fingerprints, protein sequences, and binding pockets, reaches 100 percent validity, high drug-likeness, and competitive performance on CrossDocked2020 and MolGenBench tasks.

What carries the argument

In-place replacement of value-token embeddings during supervised fine-tuning, which lets one autoregressive decoder follow multiple conditioning modes without separate paths or architectures.

If this is right

  • The pretrained model alone already yields 100 percent validity and high drug-likeness for unconditional and fragment-constrained generation.
  • The fine-tuned model follows both single-property and multi-property instructions while staying competitive on target-conditioned generation.
  • Molecules produced on MolGenBench show favorable safety and potency profiles relative to other models.
  • All modes share the same autoregressive generation path, eliminating the need for separate decoders per condition type.

Where Pith is reading between the lines

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

  • The embedding-swap approach could be tested on additional condition types such as 3D conformer constraints without retraining the core decoder.
  • Small-parameter models trained this way might be deployed directly in screening pipelines where larger multimodal systems are currently required.
  • If the replacement trick scales cleanly, it reduces the engineering cost of adding new molecular design objectives to an existing foundation model.

Load-bearing premise

That swapping value-token embeddings in place is enough for the decoder to learn and follow every listed condition without losing validity or performance.

What would settle it

A test set where the model is given combined conditions such as a target protein sequence plus a required potency value and produces molecules that violate validity or fail to satisfy both constraints at rates higher than the reported baselines.

Figures

Figures reproduced from arXiv: 2606.25865 by Haoyu Lin, Jianfeng Pei, Jinmei Pan, Luhua Lai, Xinliao Ling, Yiyan Liao.

Figure 1
Figure 1. Figure 1: Fragment-SELFIES encoding and decoding workflow. A molecule or SMILES string is BRICS-fragmented into chemically meaningful fragments with dummy attachment sites. The encoder serializes the resulting fragment tree with root [Frag] tokens, parent-side [Attach:M] selectors, child-side [Frag@N] tokens, and [pop] returns while keeping each fragment body in SELFIES￾compatible tokens. The connection-type panel d… view at source ↗
Figure 2
Figure 2. Figure 2: Molexar architecture and training flow. Stage 1 performs unconditional causal-language-model pretraining on Fragment￾SELFIES strings with the unified prompt template and a Gemma2-style molecular decoder. Stage 2 performs universal multi-condition supervised fine-tuning: scalar property encoders, pharmacophore-fingerprint projections, protein-sequence embeddings, and pocket￾geometry encoders produce vectors… view at source ↗
Figure 3
Figure 3. Figure 3: Inference efficiency under unconditional generation. Generation time (left) and peak GPU memory (right) versus batch size. Crosses (×): first OOM batch size; dotted line: 80 GB capacity. file, combining the highest quality with strong diversity and near-perfect uniqueness. By construction, Molexar decodes generated Fragment-SELFIES strings through a validity-oriented codec with RDKit sanitization and non￾s… view at source ↗
Figure 4
Figure 4. Figure 4: Single-property conditioning across nine properties. Each panel shows the property distribution of molecules generated under three target values (dashed red lines); from left to right and top to bottom: HAC, HBDC, HBAC, RotBC (discrete), WT, LogP, TPSA, QED, SAS (continuous). 4.2. Property-Controlled Generation Conditioning in prior molecular generators is often re￾stricted: some methods require additional… view at source ↗
Figure 5
Figure 5. Figure 5: Multi-property conditioning under realistic design scenarios. Left: size-flexibility decoupling, jointly conditioning on {HAC, RotBC} at rigid-extreme {25, 2}, rigid {25, 5}, flexible {25, 8}, and flexible-extreme {25, 11}. Middle: QED-SAS trade-off, conditioning on {QED, SAS} at {0.85, 1.5}, {0.85, 4.0}, {0.40, 1.5}, and {0.40, 4.0}. Right: drug-development regimes defined by {WT, LogP, HBDC}: fragment-li… view at source ↗
Figure 6
Figure 6. Figure 6: Target-conditioned molecular generation on the CrossDocked2020 test set. Left: generation metrics for methods conditioned on target structure or sequence (some methods additionally take a reference-ligand pharmacophore as input). Right: generation conditioned only on the reference-ligand pharmacophore. For Molexar, we report molecules generated under each of three conditioning modes: protein sequence, bind… view at source ↗
Figure 7
Figure 7. Figure 7: Real-world applicability results on MolGenBench. Left: fraction of generated molecules that pass the chemical filters (dark) and the fraction of molecules that also contribute a unique Bemis-Murcko scaffold (Bemis & Murcko, 1996) (light). Middle: evaluation of molecular generative models on identifying active molecules and scaffolds through de novo generation. Right: evaluation of hit-to-lead (H2L) optimiz… view at source ↗
Figure 8
Figure 8. Figure 8: Marginal property distributions of unconditionally generated molecules versus the training set. Each panel overlays the normalized histogram and kernel-density estimate of one molecular property for 10,000 molecules sampled from Molexar (pink) and for the 33.9M-molecule training corpus (green); dashed vertical lines mark the respective means. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: UMAP embedding of ECFP4 fingerprints for 10,000 unconditionally generated molecules (pink) and 100,000 molecules subsampled from the training set (green). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Single-property obedience rate versus training-set abundance for the nine conditioned properties. In each panel, the pink bars show the obedience rate at each sampling point (left axis) and the green histogram shows the abundance of that property value in the training set (right axis). 21 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Pairwise Pearson correlation coefficients (r) between the nine conditioning properties in the training set. D. Target-Conditioned Generation on the CrossDocked2020 Benchmark We provide the detailed per-metric results behind [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: General properties of Molexar on MolGenBench. Evaluation results of basic molecular properties, including validity, QED, SA score, uniqueness, diversity, and motif distributions (atom types, ring types, functional groups) with reference actives. Values represent the mean from three independent replicates; all metrics are normalized to a 0–1 scale, where higher values indicate better performance. F. Exampl… view at source ↗
Figure 13
Figure 13. Figure 13: Examples of generated molecules on de novo generation. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Examples of generated molecules on fragment-constrained generation of Eliglustat. Fragment WT=250, LogP=1.5, HBDC=2 Lead-like WT=350, LogP=2.5, HBDC=2 Drug-like WT=450, LogP=3.5, HBDC=2 bRo5 WT=600, LogP=4.5, HBDC=3 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Examples of molecules generated under the realistic drug-regime multi-property settings. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
read the original abstract

Molecular generation is a central challenge in drug discovery, requiring models that explore vast chemical space while satisfying diverse design constraints. We present Molexar, a unified multimodal molecular foundation model built on Fragment-SELFIES, a robust, fragment-aware molecular language with validity-preserving decoding and explicit fragment structure. A pretrained autoregressive decoder learns the Fragment-SELFIES syntax and molecular distribution; supervised fine-tuning (SFT) then trains the same decoder on condition-molecule pairs spanning scalar molecular properties, pharmacophore fingerprints, protein sequences, and binding pockets, injecting each condition by in-place replacement of value-token embeddings so that all generation modes share one autoregressive path. Molexar achieves strong efficiency at a small parameter count while matching or exceeding larger models. The pretrained model reaches 100% validity and high drug-likeness in unconditional and fragment-constrained generation; the SFT model follows single- and multi-property instructions and remains competitive on target-conditioned generation on the CrossDocked2020 test set. On MolGenBench, Molexar further generates molecules with favorable safety and potency. These results establish Molexar as a practical unified foundation for computational chemistry and drug-design workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents Molexar, a unified multimodal molecular foundation model using Fragment-SELFIES as input representation. A single autoregressive decoder is first pretrained on molecular strings for unconditional and fragment-constrained generation, then supervised fine-tuned on condition-molecule pairs (scalar properties, pharmacophore fingerprints, protein sequences, binding pockets) by injecting conditions via in-place replacement of value-token embeddings. The model is claimed to achieve 100% validity, high drug-likeness, competitive target-conditioned generation on CrossDocked2020, and favorable safety/potency on MolGenBench, all at small parameter count while matching or exceeding larger models.

Significance. If the central claims hold after detailed validation, the work would demonstrate a practical, parameter-efficient unified foundation model for multiple molecular generation modes in drug design. The Fragment-SELFIES representation and single-decoder conditioning strategy could simplify workflows; the reported 100% validity and competitive CrossDocked2020 results would be notable strengths if supported by reproducible code or data splits.

major comments (2)
  1. [Abstract / Methods] Abstract and methods description: the central conditioning mechanism (in-place replacement of value-token embeddings during SFT to handle scalar properties, pharmacophore fingerprints, protein sequences, and binding pockets within one autoregressive path) is described at high level only, with no equations, pseudocode, or ablation showing that this replacement preserves validity and performance across modes without interference or mode collapse. This is load-bearing for the unified multimodal claim.
  2. [Results] Results section: performance numbers (100% validity, competitive CrossDocked2020 scores, MolGenBench safety/potency) are stated without data splits, training hyperparameters, baseline details, error bars, or statistical tests, making it impossible to assess whether the small-parameter efficiency claim is supported or if results are reproducible.
minor comments (1)
  1. [Abstract] The abstract would benefit from explicit mention of model size (parameter count) and comparison models to ground the 'small parameter count' and 'matching or exceeding larger models' claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate the revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and methods description: the central conditioning mechanism (in-place replacement of value-token embeddings during SFT to handle scalar properties, pharmacophore fingerprints, protein sequences, and binding pockets within one autoregressive path) is described at high level only, with no equations, pseudocode, or ablation showing that this replacement preserves validity and performance across modes without interference or mode collapse. This is load-bearing for the unified multimodal claim.

    Authors: We agree that the current description of the in-place embedding replacement is high-level. In the revised manuscript we will expand the Methods section to include formal equations for the embedding replacement operation, pseudocode for the conditioned autoregressive generation procedure, and an ablation study that quantifies validity preservation and absence of mode collapse or cross-mode interference when conditioning on scalar properties, pharmacophores, proteins, and pockets. revision: yes

  2. Referee: [Results] Results section: performance numbers (100% validity, competitive CrossDocked2020 scores, MolGenBench safety/potency) are stated without data splits, training hyperparameters, baseline details, error bars, or statistical tests, making it impossible to assess whether the small-parameter efficiency claim is supported or if results are reproducible.

    Authors: We concur that the reported performance figures require additional supporting information for reproducibility and evaluation of the efficiency claims. The revised Results section will explicitly list the data splits employed for CrossDocked2020 and MolGenBench, all training hyperparameters, baseline configurations, error bars from repeated runs, and statistical significance tests. We will also release code and data splits to enable independent verification. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an empirical architecture (Fragment-SELFIES language, pretrained autoregressive decoder, SFT via in-place value-token embedding replacement) and reports benchmark performance. No equations, fitted parameters renamed as predictions, self-citation load-bearing premises, uniqueness theorems, or ansatzes appear in the provided text. All claims reduce to standard training and evaluation procedures without internal reduction to the authors' prior definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no concrete information on free parameters, background axioms, or newly postulated entities.

pith-pipeline@v0.9.1-grok · 5752 in / 1187 out tokens · 31428 ms · 2026-06-25T19:27:44.506543+00:00 · methodology

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

13 extracted references · 11 canonical work pages

  1. [1]

    Baell, J

    URL https://openreview.net/forum? id=KSLkFYHlYg. Baell, J. B. and Holloway, G. A. New substructure fil- ters for removal of pan assay interference compounds (PAINS) from screening libraries and for their exclu- sion in bioassays.Journal of Medicinal Chemistry, 53 (7):2719–2740, 2010. doi: 10.1021/jm901137j. URL https://doi.org/10.1021/jm901137j. Bagal, V ...

    work page doi:10.1021/jm901137j 2010

  2. [2]

    URL https: //doi.org/10.1002/cmdc.200700139

    doi: 10.1002/cmdc.200700139. URL https: //doi.org/10.1002/cmdc.200700139. Brown, N., Fiscato, M., Segler, M. H. S., and Vaucher, A. C. GuacaMol: Benchmarking models for de novo molecular design.Journal of Chemical Information and Modeling, 59(3):1096–1108, 2019. doi: 10.1021/acs.jcim. 8b00839. URL https://doi.org/10.1021/acs. jcim.8b00839. Bruns, R. F. an...

    work page doi:10.1002/cmdc.200700139 2019

  3. [3]

    URL https: //doi.org/10.1039/D1SC02436A

    doi: 10.1039/D1SC02436A. URL https: //doi.org/10.1039/D1SC02436A. Jadhav, A., Ferreira, R. S., Klumpp, C., Mott, B. T., Austin, C. P., Inglese, J., Thomas, C. J., Maloney, D. J., Shoichet, B. K., and Simeonov, A. Quantitative analyses of aggrega- tion, autofluorescence, and reactivity artifacts in a screen for inhibitors of a thiol protease.Journal of Med...

    work page doi:10.1039/d1sc02436a 2010

  4. [4]

    Jing, B., Eismann, S., Suriana, P., Townshend, R

    URL https://proceedings.mlr.press/ v119/jin20b.html. Jing, B., Eismann, S., Suriana, P., Townshend, R. J. L., and Dror, R. O. Learning from protein structure with geometric vector perceptrons. InInternational Confer- ence on Learning Representations, 2021. URL https: //openreview.net/forum?id=1YLJDvSx6J4. Kallenborn, F., Chacon, A., Hundt, C., Sirelkhatim...

    work page doi:10.1038/s41592-025-02819-8 2021

  5. [5]

    Lemos, P., Beckwith, Z., Bandi, S., van Damme, M., Crivelli-Decker, J., Shields, B

    URL https://proceedings.mlr.press/ v267/lee25o.html. Lemos, P., Beckwith, Z., Bandi, S., van Damme, M., Crivelli-Decker, J., Shields, B. J., Merth, T., Jha, P. K., 14 Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design De Mitri, N., Callahan, T. J., Nish, A., Abruzzo, P., Salomon-Ferrer, R., and Ganahl, M. SAIR: Enabling deep learning...

    work page doi:10.1101/2025.06.17.660168 2025

  6. [6]

    2017.114

    URL https://doi.org/10.1038/nprot. 2017.114. Li, Y ., Pei, J., and Lai, L. Structure-based de novo drug design using 3D deep generative models.Chemical Science, 12(41):13664–13675, 2021. doi: 10.1039/ D1SC04444C. URL https://doi.org/10.1039/ D1SC04444C. Lin, H., Huang, Y ., Zhang, O., Ma, S., Liu, M., Li, X., Wu, L., Wang, J., Hou, T., and Li, S. Z. DiffB...

    work page doi:10.1038/nprot 2017

  7. [7]

    Liu, Z., Li, Y ., Han, L., Li, J., Liu, J., Zhao, Z., Nie, W., Liu, Y ., and Wang, R

    URL https://proceedings.mlr.press/ v162/liu22m.html. Liu, Z., Li, Y ., Han, L., Li, J., Liu, J., Zhao, Z., Nie, W., Liu, Y ., and Wang, R. PDB-wide collection of binding data: Current status of the PDBbind database. Bioinformatics, 31(3):405–412, 2015. doi: 10.1093/ bioinformatics/btu626. URL https://doi.org/10. 1093/bioinformatics/btu626. Luo, S., Guan, ...

    2015

  8. [8]

    cc/paper_files/paper/2021/file/ 314450613369e0ee72d0da7f6fee773c-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2021/file/ 314450613369e0ee72d0da7f6fee773c-Paper. pdf. Luo, Y ., Yang, K., Hong, M., Liu, X. Y ., and Nie, Z. MolFM: A multimodal molecular foundation model, 2023. URL https://arxiv.org/abs/2307.09484. arXiv preprint arXiv:2307.09484. Mysinger, M. M., Carchia, M., Irwin, J. J., and Shoichet, B. K. Dir...

    work page doi:10.1021/jm300687e 2021

  9. [9]

    Ragoza, M., Masuda, T., and Koes, D

    URL https://proceedings.mlr.press/ v235/qu24a.html. Ragoza, M., Masuda, T., and Koes, D. R. Generating 3D molecules conditional on receptor binding sites with deep generative models.Chemical Science, 13(9):2701–2713,

  10. [10]

    URL https:// doi.org/10.1039/D1SC05976A

    doi: 10.1039/D1SC05976A. URL https:// doi.org/10.1039/D1SC05976A. Schneuing, A., Harris, C., Du, Y ., Didi, K., Jamasb, A., Igashov, I., Du, W., Gomes, C., Blundell, T. L., Lio, P., Welling, M., Bronstein, M., and Correia, B. Structure- based drug design with equivariant diffusion models.Na- ture Computational Science, 4(12):899–909, 2024. doi: 10.1038/s4...

    work page doi:10.1039/d1sc05976a 2024

  11. [11]

    URL https://openreview.net/forum? id=g3VCIM94ke. Schuffenhauer, A., Schneider, N., Hintermann, S., Auld, D., Blank, J., Cotesta, S., Engeloch, C., Fechner, N., Gaul, C., Giovannoni, J., Jansen, J., Joslin, J., Krastel, P., Lounkine, E., Manchester, J., Monovich, L. G., Pellicci- oli, A. P., Schwarze, M., Shultz, M. D., Stiefl, N., and Baeschlin, D. K. Evo...

    work page doi:10.1021/acs.jmedchem 2020

  12. [12]

    URL https: //doi.org/10.1038/s41467-025-56349-0

    doi: 10.1038/s41467-025-56349-0. URL https: //doi.org/10.1038/s41467-025-56349-0. Zhang, O., Zhang, J., Jin, J., Zhang, X., Hu, R., Shen, C., Cao, H., Du, H., Kang, Y ., Deng, Y ., Liu, F., Chen, G., Hsieh, C.-Y ., and Hou, T. ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling.Nature Machine Intelligence, 5(9):1020–1030,

    work page doi:10.1038/s41467-025-56349-0

  13. [13]

    URL https: //doi.org/10.1038/s42256-023-00712-7

    doi: 10.1038/s42256-023-00712-7. URL https: //doi.org/10.1038/s42256-023-00712-7. Zhu, H., Zhou, R., Cao, D., Tang, J., and Li, M. A pharmacophore-guided deep learning ap- proach for bioactive molecular generation.Nature Communications, 14(1):6234, 2023. doi: 10.1038/ s41467-023-41454-9. URL https://doi.org/10. 1038/s41467-023-41454-9. Zhung, W., Kim, H.,...

    work page doi:10.1038/s42256-023-00712-7 2023