Pith. sign in

REVIEW 4 major objections 5 minor 65 references

MultiPUFFIN claims that a physics-aware multimodal model can outperform a SMILES-only pretrained baseline on all nine thermophysical properties despite using roughly 2,000× fewer labelled molecules.

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

2026-08-02 19:45 UTC pith:RDYHTNIP

load-bearing objection Real system paper with the strongest ablation work in the PUFFIN line, but the 'wins all nine' headline is not supported: the logP win is contaminated by a descriptor the paper itself admits leaks, and the pre-registered scope filters rest on assertion. the 4 major comments →

arxiv 2603.00857 v2 pith:RDYHTNIP submitted 2026-03-01 cs.LG cs.AI

MultiPUFFIN: A Multimodal Domain-Constrained Foundation Model for Molecular Property Prediction of Small Molecules

Idelfonso B. R. Nogueira , Carine M. Rebello , Mumin Enis Leblebici , Erick Giovani Sperandio Nascimento This is my paper
classification cs.LG cs.AI
keywords molecular property predictionfoundation modelmultimodal learningthermophysical propertiesphysics-informed machine learningself-supervised pretrainingmulti-task learninginductive bias
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 argues that baking thermodynamic equations into a molecular model's output is a more data-efficient route to accurate property prediction than scaling up unlabelled pretraining. It introduces MultiPUFFIN, which fuses SMILES sequences, 2D molecular graphs, and 3D conformer geometries, pretrains on 500,000 unlabelled molecules, and routes each of nine thermophysical properties to a physically motivated head (Antoine, Andrade, van 't Hoff, Born, Shomate, group contribution, or a direct net). On a scaffold-split test set, the model reports an in-scope mean R² of 0.784 and beats a fine-tuned SMILES-only transformer baseline on all nine properties while using roughly 2,000× fewer labelled molecules. If this holds, chemistry-domain inductive bias is a practical substitute for brute-force data scaling in process engineering and drug discovery.

Core claim

The paper's central claim is that a single multi-task model with domain-informed output heads and condition-aware refinement can outperform a separately fine-tuned SMILES-only foundation model on every one of nine thermophysical targets. The advantage is largest for temperature-dependent properties (vapor pressure, viscosity), where the SMILES-only model has no way to encode measurement temperature and its R² collapses, while MultiPUFFIN's Antoine and Andrade heads guarantee monotonic temperature behavior by construction. On the six temperature-independent properties the margin is positive but modest (about 0.04–0.16 in R²). The paper attributes the overall result to three mechanisms acting

What carries the argument

The load-bearing mechanism is the per-property head tournament. For each of the nine properties, four candidate output heads are trained jointly on the shared backbone: a primary thermophysical equation (e.g., Antoine for vapor pressure, Andrade for viscosity), a Joback group-contribution head, a fragment-count head, and a direct or alternative-physics head. Each domain-informed head predicts equation parameters from the fused embedding and evaluates the equation at the measurement temperature, making the correct temperature dependence a structural guarantee rather than a learned behavior; the winner per property is chosen on validation RMSE. Above the heads, five identity-initialized condit

Load-bearing premise

The headline number depends on the per-property G+ applicability filters having been fixed before the test set was ever scored; the paper asserts this pre-registration but provides no external proof, so the reader cannot exclude that the filters were selected with knowledge of the test composition.

What would settle it

Inspect the creation history of the per-property G+ filter definitions — a dated registry, code commit, or pre-registration document. Then recompute the reported R² on the full 8,877-molecule test set without any class exclusions. If the filters cannot be shown to predate test-set access, or if removing them drops the mean R² by more than the paper's own seed-to-seed uncertainty band (±0.03–0.05), the 0.784 headline is a selection-biased in-scope metric rather than a deployment metric.

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

If this is right

  • Accurate multi-property prediction becomes possible with tens of thousands of labelled molecules rather than tens of millions, lowering the entry cost for chemical-engineering screening.
  • Physically consistent temperature dependence (monotonic vapor pressure, monotonic viscosity) is guaranteed by construction, so the predicted curves can be handed to process simulators without post-hoc fitting or rejection of unphysical outputs.
  • One shared backbone replaces nine separately fine-tuned models, and data-rich properties measurably improve data-scarce ones such as hydration free energy and heat capacity.
  • The condition-module design is ready to absorb pH, polymorph, and test-method data when such measurements become available, so the same architecture can widen its scope without retraining from scratch.

Where Pith is reading between the lines

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

  • The paper's 'wins all nine' headline leans on the temperature-dependent properties, where the comparison is asymmetric because MultiPUFFIN receives temperature as an input while the SMILES-only baseline does not; the six temperature-independent comparisons are a cleaner test of representation quality.
  • The headline in-scope R²=0.784 assumes the per-property G+ filters were pre-registered before the test set was scored; the paper provides no externally verifiable record of that timing, so the honest deployment metric could be lower if the filters were shaped by test composition.
  • The 512-dimensional fused embedding, shown to organise molecules by thermophysical similarity, is a plausible off-the-shelf feature representation for properties outside the nine; the paper does not test this, but it would be a direct way to reuse the model without retraining.

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 presents MultiPUFFIN, a multimodal foundation model that fuses SMILES strings, 2D molecular graphs, and 3D conformer geometries with auxiliary encoders for experimental conditions and molecular descriptors. The backbone is pretrained on 500,000 unlabeled PubChem molecules with three self-supervised objectives, then trained on 37,968 molecules for nine thermophysical properties. Domain-informed output heads (Antoine, Andrade, van 't Hoff, Born, Shomate, group contribution) are organized into a per-property four-head tournament, with condition-aware refinement modules and two cross-property coupling losses. On a scaffold-split test set, the paper reports an in-scope mean R² of 0.784 and claims to outperform fine-tuned ChemBERTa-2 on all nine properties despite using roughly 2,000× fewer labeled molecules. Systematic ablations examine architectural components, head choices, and equation-level alternatives, and the paper is unusually candid about limitations, including single-seed evaluation, descriptor leakage, and deployment-frequency caveats.

Significance. If the central claims hold, the work would be significant: a single physically constrained multimodal model that achieves strong accuracy across nine diverse thermophysical properties with orders-of magnitude less labeled data than a large SMILES-pretrained baseline, while providing thermodynamic consistency by construction. The manuscript's strengths include its systematic ablation studies (Section 3.5), the explicit documentation of negative results (Sections 2.12 and 3.10), the multi-source curated dataset, and the transparent treatment of known asymmetries. However, the headline in-scope R² depends on applicability-domain filters whose pre-registration is asserted but not externally verifiable, the logP result is acknowledged to benefit from descriptor leakage, and the single-seed/single-split evaluation leaves key comparative margins without uncertainty quantification. The central claim is plausible but not yet fully supported.

major comments (4)
  1. [§2.14, Table 3, Stage 4] The headline in-scope mean R²=0.784 is computed only after applying per-property G+ applicability-domain filters, and the filters are asserted to be 'pre-registered, frozen before the test set is ever scored' without any external registry, timestamped artifact, or code. Table 3 discloses per-class test counts, but this transparency cannot rule out that the filter definitions were influenced by test composition. Because the central claim is the joint accuracy-plus-scope statement, this is load-bearing. Please provide an externally verifiable registration (e.g., timestamped hash of the filter code) or, at minimum, report full-test metrics on all 8,877 molecules, per-class metrics, and a sensitivity analysis across filter strictness. If the full-test mean is close to 0.784 and the per-class results are consistent, the deployment claim is credible; if the gap is large, the headline is select
  2. [§2.8, §3.5.1, §3.9, Table 13] Descriptor 9 in the 13-dimensional descriptor vector is explicitly stated to be the Wildman–Crippen MolLogP estimate, i.e., a precomputed approximation of the logP target. Section 2.8 calls this 'a mild form of information leakage for the logP head' and instructs downstream users to account for the asymmetry when comparing against baselines without Crippen descriptors. However, Section 3.9 and Table 13 report the logP win (R² 0.821 vs 0.687) as one of the nine wins with no such adjustment, and no ablation with MolLogP masked is reported. The concern is reinforced by Table 8, where GCN-Only achieves better logP RMSE (0.714) than full MultiPUFFIN (0.906), suggesting the descriptor input dominates the logP result. Please either remove MolLogP from the descriptor vector, or add a MolLogP-masked ablation and re-state the headline as 'eight of nine' or 'all nine except logP' if the win does no
  3. [§3.10, Tables 7 and 13] All results come from a single random seed and a single scaffold split, and the main tables contain no confidence intervals or error bars. The paper itself acknowledges in Section 3.10 that the bootstrap R² uncertainty for the smallest test sets (HFE n=269, viscosity n=331, heat capacity n=297) is ±0.03–0.05 at the 95% level, and that several equation-level ablation deltas (e.g., Table 11: Wagner vs. Antoine, ΔR²≈+0.004) fall inside this band. The largest ChemBERTa-2 margins are likely robust, but the 'all nine properties' claim mixes properties with very different test sizes and uncertainty levels. Please add multi-seed results (at least 3 seeds) or bootstrap confidence intervals for the headline in-scope mean and the Table 13 comparison, and explicitly separate in-scope metrics (Table 7) from full-test metrics (Table 13) in all summary statements.
  4. [§3.9, Table 13] For vapor pressure and viscosity, the comparison against ChemBERTa-2 is asymmetric by design: MultiPUFFIN receives temperature as an explicit input, while a SMILES-only baseline cannot distinguish measurements at different temperatures. The paper states this limitation clearly, yet still counts these two properties in the 'wins all nine' headline and in the conclusion's 'beats nine separately fine-tuned baselines on all nine targets.' For the label-efficiency conclusion, the fair head-to-head is the six temperature-independent properties (mean R² 0.769 vs 0.658) plus heat capacity; the temperature-conditioned wins should be framed as a capability of the domain-informed architecture rather than as part of the same 'outperforms on all nine' claim. Please adjust the headline claims accordingly.
minor comments (5)
  1. [§2.2, Eq. (1)] The InfoNCE loss has a missing citation marker '[?]' in the displayed equation and surrounding text. Please add the proper reference.
  2. [§2.1 and Table 7/Figure 3] Hydration free energy units are written as 'k/mol' (e.g., 'measurements in k/mol at 298.15 K' and in Table 7). This should be 'kcal/mol' throughout.
  3. [§3.9, Table 13] The text states that MultiPUFFIN achieves R²=0.921 for heat capacity, but Table 13 reports R²=0.940 for heat capacity. Please reconcile the two values.
  4. [§3.10] The 'Unweighted mean R²' caveat is important, but the abstract and conclusion still report the unweighted mean 0.784 without an accompanying distribution or weighted variant. Consider reporting both unweighted and weighted means, or at least the per-property n values next to the headline.
  5. [Code and data availability] The paper states that code, weights, and data 'will be made publicly available upon publication.' For a paper whose central claims rest on pre-registered filters and reproducibility, the artifact should be available at submission or at least under a timestamped repository DOI; promising future availability is not sufficient to verify the pre-registration claim.

Circularity Check

1 steps flagged

The 'all nine' claim is partially circular: the logP head receives a precomputed logP estimator as input, and the reported logP win over ChemBERTa-2 is counted without the asymmetry adjustment the paper itself prescribes.

specific steps
  1. fitted input called prediction [Section 2.8 (Auxiliary input encoders); Section 3.9 / Table 13]
    "One potential concern is that the inclusion of the Wildman–Crippen logP estimate (MolLogP, descriptor 9) in the descriptor vector may advantage the logP prediction task by providing a pre-computed approximation of the target value directly as input. This constitutes a mild form of information leakage for the logP head: the model can learn to use the Crippen estimate as a prior and then correct the residual from the structural representations. The effect is intentional; the descriptor encoder is designed to provide the model with the same coarse physicochemical priors that an experienced chemis"

    The logP prediction is made from a descriptor vector containing MolLogP, a precomputed estimate of exactly the logP target; the paper calls this 'information leakage' and tells readers to adjust comparisons against baselines lacking Crippen descriptors. Table 13 nevertheless reports the logP R² win (0.821 vs 0.687) as one of the nine wins with no adjustment. The head-to-head margin is thus partly carried by the input estimator rather than by the multimodal/domain-informed architecture, so the 'all nine properties' claim is not established for logP.

full rationale

The main body of the derivation is self-contained: SSL pretraining on a disjoint PubChem corpus, hybrid scaffold/property-balanced splitting, the multimodal fusion, the four-head tournament, and the cross-property couplings are all evaluated with internal ablations (architectural, head, and equation-level) rather than by appeal to prior work. The PUFFIN/ExPUFFIN self-citations are architectural heritage, not load-bearing proof; the paper's own All-DirectHeads and Swapped-Equation ablations independently demonstrate the value of the domain-informed heads. The temperature-conditioning advantage over ChemBERTa-2 is openly acknowledged and is a real architectural capability, not a hidden fit. The one genuinely circular element is the logP task: the descriptor encoder injects a precomputed Wildman–Crippen logP value directly into the representation, the paper explicitly labels this 'information leakage' and demands an asymmetry correction, but the ChemBERTa-2 head-to-head table counts the logP win without applying that correction. That makes the 'wins all nine' headline partially constructed from the input feature. Separately, the G+ applicability-domain construction is a validity concern rather than a circular-derivation one: Table 3 reports test counts per excluded class while Section 2.14 claims the filters were frozen before test scoring; this is a selection-bias risk that should be weighed in deployment claims, but it is not an equation-level reduction. On balance the central architectural derivation is independent, but one of the nine claimed wins is confounded by construction, so a partial circularity score of 6 is appropriate.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The central claim depends on a small set of domain assumptions about the validity of classical thermodynamic equations and data quality, plus hand-set training constants and scope-filter thresholds. No new physical entities are postulated. The most concerning free choice is the G+ filter set, because it directly defines the evaluation regime, and the MolLogP descriptor, because it gives the logP head a precomputed target estimate as input.

free parameters (8)
  • Contrastive temperature τ = 0.07
    Hand-set in Eq. (1); no sensitivity analysis reported.
  • Coupling weights λ_FP, λ_BP = 1e-2, 5e-2
    Hand-set weights for Eqs. (19)–(20); no grid search or sensitivity study reported.
  • Flash-point target partial pressure P*_FP = 3000 Pa
    Universal constant in Eq. (19), described as within a factor of two of flammability limits; it is itself an approximate fitted value.
  • G+ applicability-domain thresholds = heavy-atom cutoffs >50 and ≤4; per-class exclusions
    Define the in-scope test subset on which the headline R² is computed; derived from a coverage audit rather than an independent benchmark.
  • Per-property tournament head selection = e.g., Wagner for VP, Born for HFE, DirectHead for logP/Tm/Tf
    Validation-based model selection; the final deployed head set is not consistently specified across Tables 5 and 10–11.
  • Uncertainty task weights s_p = learned, values not reported
    Kendall et al. weighting in Eq. (22), initialized to zero and jointly optimized; final values omitted.
  • Stage-3 learning rates α_A and α_B/C/D = 1e-6 and 5e-6
    Hand-set for the backbone-unfrozen fine tune; no sensitivity study.
  • Temperature-smoothness weight λ_smooth = 0 (disabled)
    Final configuration disables this regularizer after reporting no validation gain; a tuning decision.
axioms (6)
  • domain assumption The assigned physics equations (Antoine, Andrade, van 't Hoff, Born, Shomate, group additivity) are appropriate functional forms for their properties.
    Section 2.11/Table 4; the tournament mitigates the risk, but the final model still assumes these forms are the right class.
  • domain assumption A single MMFF94-optimized ETKDGv3 conformer adequately represents the 3D ensemble when combined with the geometry gate.
    Section 2.1 explicitly acknowledges this is imperfect for flexible molecules; the geometry gate only partially compensates.
  • domain assumption Labels from eleven merged public sources are accurate and comparable enough for multi-task training.
    Section 2.1; inter-laboratory disagreement, especially in solubility, is acknowledged as a noise source.
  • domain assumption The hybrid scaffold + coverage-balanced split produces a test set that measures structural generalization.
    Algorithm 1; this is a procedural assumption not independently benchmarked.
  • ad hoc to paper The G+ filters were fixed before any test prediction was computed.
    Section 2.14 asserts pre-registration, but no external registry or timestamp is supplied and test counts appear in Table 3.
  • domain assumption The Crippen MolLogP descriptor is a legitimate auxiliary input for the logP prediction head.
    Section 2.8 calls the leakage intentional; the logP comparison against ChemBERTa-2 depends on it.

pith-pipeline@v1.3.0-alltime-deepseek · 46256 in / 13902 out tokens · 104204 ms · 2026-08-02T19:45:27.778875+00:00 · methodology

0 comments
read the original abstract

MultiPUFFIN is a domain-informed multimodal foundation model for predicting thermophysical properties of small molecules, addressing a critical gap in chemical engineering, drug discovery, and materials science. Existing molecular foundation models pretrain on millions of molecules to learn general-purpose representations, but their standard MLP output layers impose no physical constraints, vapor pressure predictions may violate monotonic temperature dependence, and viscosity curves may lack the functional form required by process simulators. Domain-informed approaches that guarantee thermodynamic consistency have remained limited to single properties and small datasets, whereas multimodal foundation models have focused on biological activity rather than thermophysical properties. MultiPUFFIN fills this gap by fusing SMILES sequences, 2D molecular graphs, and 3D conformer geometries through bidirectional cross-modal attention and gated fusion, supplemented by auxiliary encoders for experimental conditions and molecular descriptors. The backbone is pretrained on 500,000 unlabelled PubChem molecules using three complementary self-supervised objectives. A condition-aware refinement stack of five conditioners (temperature, pH, pressure, polymorph, and measurement method) routes each property to a four-head tournament that selects the best-performing thermodynamically informed head for that property. MultiPUFFIN achieves a mean test R2 of 0.784 and outperforms fine-tuned ChemBERTa-2 on all nine properties despite training on roughly 2,000x fewer labeled molecules.

Figures

Figures reproduced from arXiv: 2603.00857 by Carine M. Rebello, Erick Giovani Sperandio Nascimento, Idelfonso B. R. Nogueira, Mumin Enis Leblebici.

Figure 1
Figure 1. Figure 1: Architecture overview of MultiPUFFIN. Three structural encoders (GCN over the 2D graph, Transformer over the SMILES string, SchNet over the 3D conformer) and two auxiliary encoders (experimental conditions; molecular descriptors) form the multimodal backbone, pretrained on 500,000 unlabeled PubChem molecules with three SSL objectives. Cross-modal attention plus gated fusion produces a 512-dimensional unifi… view at source ↗
Figure 2
Figure 2. Figure 2: Test RMSE (left) and MAE (right) across all nine properties for training, validation, and test splits (logarithmic scale). The consistent increase from training to test error reflects the generalization challenge imposed by the scaffold-based splitting strategy. Properties are ordered by decreasing test performance. Note the logarithmic y-axis: temperature-based properties (e.g., melting point, boiling poi… view at source ↗
Figure 3
Figure 3. Figure 3: Test set parity plots (predicted vs. experimental) for all nine physicochemical properties. The solid diagonal line represents perfect prediction (y = x). Each panel reports the R 2 , RMSE, MAE, and number of test samples. Properties are ordered by decreasing test R 2 . Note: per-panel metrics in this figure are drawn from the controlled Stage-2 reference run used as the architectural-ablation baseline (Se… view at source ↗
Figure 4
Figure 4. Figure 4: Test set residual distributions for all nine properties. Each panel shows the histogram of residuals (predicted − experi￾mental) with a Gaussian fit (solid line) and a zero-residual reference line (dashed red). The mean (µ) and standard deviation (σ) of the residuals are annotated. 30 [PITH_FULL_IMAGE:figures/full_fig_p031_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics of MultiPUFFIN during Stage 2 of the four-stage protocol (Section 2.13). (a) Overall training and validation loss across the joint multi-task phase and its head-fine tuning epilogue, with the best validation loss marked (⋆). The vertical dashed line indicates the transition from joint training (with all parameters trainable) to a backbone-frozen head￾consolidation phase that closes Stage … view at source ↗
Figure 6
Figure 6. Figure 6: Data availability per property and split. The highly heterogeneous sample counts across properties reflect the different availability of experimental measurements in public databases. Training counts reflect unique molecules (before SMILES augmen￾tation). The multi-task setting provides clear benefits for data-scarce properties through positive transfer from larger prop￾erty datasets. Heat capacity (861 tr… view at source ↗
Figure 7
Figure 7. Figure 7: quantifies the generalization gap (∆RMSE = RMSEtest −RMSEtrain) for each property, providing a direct measure of how much performance degrades when moving from the training distribution to structurally novel test molecules. Because properties have different units and scales, the normalized RMSE (NRMSE = RMSE/σexp) is used for cross-property comparison. The largest generalization gaps are observed for prope… view at source ↗
Figure 8
Figure 8. Figure 8: Normalized RMSE (a) and normalized MAE (b) across all nine properties for training, validation, and test splits. Metrics are normalized by the standard deviation of experimental values in the test set (σexp), enabling meaningful cross-property comparison. The dashed line at 1.0 represents the error of a mean-only predictor. 3.5.1. Contribution of architectural components Three architectural variants were t… view at source ↗
Figure 9
Figure 9. Figure 9: Architectural ablation study: per-property test RMSE for the full trimodal MultiPUFFIN architecture versus the GCN￾Only and No SchNet variants. Removing the SchNet 3D encoder disproportionately increases RMSE for geometry-sensitive properties (HFE, heat capacity), while removing the Transformer and SchNet structural encoders (GCN-Only) degrades perfor￾mance across most properties. 3.5.2. Influence of domai… view at source ↗
Figure 10
Figure 10. Figure 10: Influence of domain-informed prediction heads on the six thermodynamically-informed properties. The All Direct￾Heads variant reveals that domain equations are most beneficial for viscosity (Andrade), while hydration free energy and log P achieve lower RMSE without the thermodynamically-informed constraints. The Swapped Antoine↔Andrade variant demonstrates catastrophic vapor pressure degradation (RMSE incr… view at source ↗
Figure 11
Figure 11. Figure 11: Per-property test RMSE heatmap across all ablation variants. Green indicates low RMSE (better performance); red indicates high RMSE. The All GroupContribution variant (itself a domain-informed representation based on thermodynamic group additivity) achieves the lowest RMSE for most properties, while the full model provides the best performance for viscosity (An￾drade equation). Both domain-informed varian… view at source ↗
Figure 12
Figure 12. Figure 12: Per-component contribution to model performance, expressed as ∆RMSE = RMSEablation − RMSEfull (positive values indicate the full model achieves lower error). The most dramatic effect is the catastrophic vapor pressure RMSE increase when the Antoine and Andrade equations are swapped (∆RMSE = +0.83). The SchNet 3D encoder contributes most to reducing HFE error (∆RMSE = +0.90 kcal/mol), while several propert… view at source ↗
Figure 13
Figure 13. Figure 13: UMAP projection of the MultiPUFFIN unified embedding (u ∈ R 512) for 37 968 unique molecules, colored by each of the nine predicted properties. Gray points indicate molecules without a measurement for the given property. All properties show smooth, continuous gradients, indicating that the learned representation organizes molecules by thermophysical similarity. Temperature-based properties (boiling point,… view at source ↗
Figure 14
Figure 14. Figure 14: Cluster structure of the MultiPUFFIN embedding space. From left to right: (a) k-Means clusters (k = 5, optimal by silhouette analysis), revealing chemically interpretable groups ranging from polar/hydrophilic to hydrophobic; (b) HDBSCAN density-based clustering, identifying two major density cores with 10.2% noise (inter-cluster transitions); (c) top-10 Murcko scaf￾fold families, showing scaffold-level co… view at source ↗
Figure 15
Figure 15. Figure 15: visualises the per-property bar chart and the head-to-head ∆R 2 margin: VP and viscosity values below −0.2 are clipped with the actual value annotated, and panel (b) restricts the ∆R 2 view to the seven properties where both models produce meaningful predictions. Cp ΔGhyd logP Tf Tb Tm logη logPvap logS −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0 T est R 2 (a) Per-property R 2 — same nine targets MultiPUFFIN (… 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

65 extracted references · 1 canonical work pages

  1. [1]

    C. Pang, H. H. Tong, L. Wei, Advanced deep learning methods for molecular property prediction, Quantitative Biology 11 (4) (2023) 395–404

    2023

  2. [2]

    G. W. Thomson, The antoine equation for vapor-pressure data, Chemical Reviews 38 (1) (1946) 1–39

    1946

  3. [3]

    E. N. d. C. Andrade, The viscosity of liquids, Nature 125 (1930) 309–310

    1930

  4. [4]

    Nannoolal, J

    Y . Nannoolal, J. Rarey, D. Ramjugernath, Estimation of pure component properties: Part 3. estimation of the vapor pressure of non-electrolyte organic compounds via group contributions and group interactions, Fluid Phase Equilibria 269 (1-2) (2008) 117–133

    2008

  5. [5]

    M. Das, A. Ghosh, R. B. Sunoj, Advances in machine learning with chemical language models in molecular property and reaction outcome predictions, Journal of Computational Chemistry 45 (14) (2024) 1160–1176. 45

    2024

  6. [6]

    K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, K.-R. Müller, SchNet–a deep learning architecture for molecules and materials, The Journal of Chemical Physics 148 (24) (2018)

    2018

  7. [7]

    Gasteiger, J

    J. Gasteiger, J. Groß, S. Günnemann, Directional message passing for molecular graphs, International Confer- ence on Learning Representations (2020)

    2020

  8. [8]

    F. H. Vermeire, W. H. Green, Transfer learning for solvation free energies: from quantum chemistry to experi- ments, Chemical Engineering Journal 418 (2021) 129307

    2021

  9. [9]

    A. M. Schweidtmann, J. G. Rittig, A. König, M. Grohe, A. Mitsos, M. Dahmen, Graph neural networks for prediction of fuel ignition quality, Energy & Fuels 34 (9) (2020) 11395–11407

    2020

  10. [10]

    J. G. Rittig, K. C. Felton, A. A. Lapkin, A. Mitsos, Gibbs–Duhem-informed neural networks for binary activity coefficient prediction, Digital Discovery 2 (2023) 1752–1767

    2023

  11. [11]

    V . V . Santana, C. M. Rebello, L. P. Queiroz, A. M. Ribeiro, N. Shardt, I. B. R. Nogueira, PUFFIN: A path- unifying feed-forward interfaced network for vapor pressure prediction, Chemical Engineering Science 286 (2024) 119623

    2024

  12. [12]

    C. M. Rebello, U. Di Caprio, J. Steen-Hansen, B. Rodrigues, E. A. Costa, A. R. dos Santos, F. Esposito, M. E. Leblebici, I. B. R. Nogueira, ExPUFFIN: Thermodynamic consistent viscosity prediction in an extended path- unifying feed-forward interfaced network, arXiv preprint arXiv:2512.06927 (2025)

    arXiv 2025

  13. [13]

    J. Choi, Y . Nam, Y .-S. Choi, Y . Jung, A perspective on foundation models in chemistry, JACS Au 5 (4) (2025). doi:10.1021/jacsau.4c01160

    work page doi:10.1021/jacsau.4c01160 2025

  14. [14]

    Batatia, et al., Foundation models for atomistic simulation of chemistry and materials, Nature Reviews Chem- istry (2025).doi:10.1038/s41570-025-00793-5

    I. Batatia, et al., Foundation models for atomistic simulation of chemistry and materials, Nature Reviews Chem- istry (2025).doi:10.1038/s41570-025-00793-5

    doi:10.1038/s41570-025-00793-5 2025

  15. [15]

    S. Wang, Y . Guo, Y . Wang, H. Sun, J. Huang, SMILES-BERT: Large scale unsupervised pre-training for molec- ular property prediction, ACM International Conference on Bioinformatics, Computational Biology and Health Informatics (2019) 429–436

    2019

  16. [16]

    Ahmad, E

    W. Ahmad, E. Simon, S. Chithrananda, G. Grand, B. Ramsundar, ChemBERTa-2: Towards chemical foundation models, arXiv preprint arXiv:2209.01712 (2022)

    Pith/arXiv arXiv 2022

  17. [17]

    H. Li, D. Zhao, J. Zeng, KPGT: Knowledge-guided pre-training of graph transformer for molecular property prediction, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 857–867

    2022

  18. [18]

    Y . Wang, J. Wang, Z. Cao, A. B. Farimani, Molecular contrastive learning of representations via graph neural networks, Nature Machine Intelligence 4 (2022) 279–287

    2022

  19. [19]

    Méndez-Lucio, et al., MolE: A foundation model for molecular graphs using disentangled attention, Nature Communications 15 (2024) 9431

    O. Méndez-Lucio, et al., MolE: A foundation model for molecular graphs using disentangled attention, Nature Communications 15 (2024) 9431

    2024

  20. [20]

    G. Zhou, Z. Gao, Q. Ding, H. Zheng, H. Xu, Z. Wei, L. Zhang, G. Ke, Uni-Mol: A universal 3D molecular representation learning framework, International Conference on Learning Representations (2023)

    2023

  21. [21]

    Lu, et al., Uni-Mol2: Exploring molecular pretraining model at scale, Advances in Neural Information Pro- cessing Systems 37 (2024)

    S. Lu, et al., Uni-Mol2: Exploring molecular pretraining model at scale, Advances in Neural Information Pro- cessing Systems 37 (2024)

    2024

  22. [22]

    Uni-Mol Team, Uni-Mol3: A multi-molecular foundation model for advancing organic reaction modeling, arXiv preprint arXiv:2508.00920 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    S. Liu, W. Nie, C. Wang, J. Lu, Z. Qiao, L. Liu, J. Tang, C. Xiao, A. Anandkumar, Multi-modal molecule structure–text model for text-based retrieval and editing, Nature Machine Intelligence 5 (2023) 1447–1457. 46

    2023

  24. [24]

    B. Su, D. Du, Z. Yang, Y . Zhou, J. Li, A. Rao, H. Sun, Z. Lu, J.-R. Wen, A molecular multimodal foundation model associating molecule graphs with natural language, arXiv preprint arXiv:2209.05481 (2022)

    Pith/arXiv arXiv 2022

  25. [25]

    P. Liu, Y . Ren, J. Tao, Z. Ren, GIT-Mol: A multi-modal large language model for molecular science with graph, image, and text, Computers in Biology and Medicine 171 (2024) 108073

    2024

  26. [26]

    Chen, et al., Pretraining graph transformer for molecular representation with fusion of multimodal informa- tion, Information Fusion 115 (2025) 102784

    R. Chen, et al., Pretraining graph transformer for molecular representation with fusion of multimodal informa- tion, Information Fusion 115 (2025) 102784

    2025

  27. [27]

    Li, et al., Advancing molecular graph-text pre-training via fine-grained alignment, arXiv preprint arXiv:2409.14106 (2024)

    Y . Li, et al., Advancing molecular graph-text pre-training via fine-grained alignment, arXiv preprint arXiv:2409.14106 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Li, et al., MolPrompt: Improving multi-modal molecular pre-training with knowledge prompts, Bioinformat- ics 41 (9) (2025)

    Y . Li, et al., MolPrompt: Improving multi-modal molecular pre-training with knowledge prompts, Bioinformat- ics 41 (9) (2025)

    2025

  29. [29]

    Weininger, SMILES, a chemical language and information system

    D. Weininger, SMILES, a chemical language and information system. 1. introduction to methodology and en- coding rules, Journal of Chemical Information and Computer Sciences 28 (1) (1988) 31–36

    1988

  30. [30]

    Baltrušaitis, C

    T. Baltrušaitis, C. Ahuja, L.-P. Morency, Multimodal machine learning: A survey and taxonomy, IEEE Transac- tions on Pattern Analysis and Machine Intelligence 41 (2) (2019) 423–443

    2019

  31. [31]

    Li, et al., MoleculeFormer is a GCN-transformer architecture for molecular property prediction, Communica- tions Biology 8 (2025) 1668

    J. Li, et al., MoleculeFormer is a GCN-transformer architecture for molecular property prediction, Communica- tions Biology 8 (2025) 1668

    2025

  32. [32]

    M. C. Sorkun, A. Khetan, S. Er, AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds, Scientific Data 6 (2019) 143

    2019

  33. [33]

    Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A. S. Pappu, K. Leswing, V . Pande, MoleculeNet: A benchmark for molecular machine learning, Chemical Science 9 (2) (2018) 513–530

    2018

  34. [34]

    Gaulton, A

    A. Gaulton, A. Hersey, M. Nowotka, A. P. Bento, J. Chambers, D. Mendez, P. Mutowo, F. Atkinson, L. J. Bellis, E. Cibrián-Uhalte, et al., The ChEMBL database in 2017, Nucleic Acids Research 45 (D1) (2017) D945–D954

    2017

  35. [35]

    D. L. Mobley, J. P. Guthrie, FreeSolv: A database of experimental and calculated hydration free energies, with input files, Journal of Computer-Aided Molecular Design 28 (7) (2014) 711–720

    2014

  36. [36]

    Mansouri, C

    K. Mansouri, C. M. Grulke, R. S. Judson, A. J. Williams, OPERA models for predicting physicochemical prop- erties and environmental fate endpoints, Journal of Cheminformatics 10 (2018) 10

    2018

  37. [37]

    Frenkel, R

    M. Frenkel, R. D. Chirico, V . Diky, X. Yan, Q. Dong, C. Muzny, ThermoML: An XML-based approach for storage and exchange of experimental and critically evaluated thermophysical and thermochemical property data. 1. Experimental data, Journal of Chemical & Engineering Data 50 (1) (2005) 42–50

    2005

  38. [38]

    European Chemicals Agency, ECHA REACH registered substances database,https://echa.europa.eu/ information-on-chemicals/registered-substances, accessed: 2024 (2024)

    2024

  39. [39]

    Bradley, A

    J.-C. Bradley, A. Lang, A. Williams, Open melting point data, Figshare (2014).doi:10.6084/m9.figshare. 1031637

    doi:10.6084/m9.figshare 2014

  40. [40]

    Y . Sun, Z. Yin, L. Yin, Prediction of flash points of organic compounds using quantitative structure-property relationship (QSPR) methods, Journal of Hazardous Materials 371 (2019) 518–524

    2019

  41. [41]

    ABB-ADD consortium, ABB-ADD liquid heat capacity dataset, curated liquid heat capacity data for organic compounds (2023)

    2023

  42. [42]

    A. K. Chew, et al., Advancing material property prediction: using physics-informed machine learning models for viscosity, Journal of Cheminformatics 16 (2024) 31. 47

    2024

  43. [43]

    S. Kim, J. Chen, T. Cheng, A. Gindulyte, J. He, S. He, Q. Li, B. A. Shoemaker, P. A. Thiessen, B. Yu, et al., PubChem 2023 update, Nucleic Acids Research 51 (D1) (2023) D1373–D1380

    2023

  44. [44]

    RDKit community, RDKit: Open-source cheminformatics,https://www.rdkit.org(2024)

    2024

  45. [45]

    Riniker, G

    S. Riniker, G. A. Landrum, Better informed distance geometry: using what we know to improve conformation generation, Journal of Chemical Information and Modeling 55 (12) (2015) 2562–2574

    2015

  46. [46]

    T. A. Halgren, Merck molecular force field. I. Basis, form, scope, parameterization, and performance of MMFF94, Journal of Computational Chemistry 17 (5-6) (1996) 490–519

    1996

  47. [47]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional transformers for lan- guage understanding, North American Chapter of the Association for Computational Linguistics (2019) 4171– 4186

    2019

  48. [48]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, International Confer- ence on Learning Representations (2017)

    2017

  49. [49]

    T. Cai, S. Luo, K. Xu, D. He, T.-Y . Liu, L. Wang, GraphNorm: A principled approach to accelerating graph neural network training, in: International Conference on Machine Learning, 2021, pp. 1204–1215

    2021

  50. [50]

    Ioffe, C

    S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift, International Conference on Machine Learning (2015) 448–456

    2015

  51. [51]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  52. [52]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in Neural Information Processing Systems 30 (2017)

    2017

  53. [53]

    Hendrycks, K

    D. Hendrycks, K. Gimpel, Gaussian error linear units (GELUs), arXiv preprint arXiv:1606.08415 (2016)

    Pith/arXiv arXiv 2016

  54. [54]

    Xiong, Y

    R. Xiong, Y . Yang, D. He, K. Zheng, S. Zheng, C. Xing, H. Zhang, Y . Lan, L. Wang, T.-Y . Liu, On layer normalization in the Transformer architecture, International Conference on Machine Learning (2020) 10524– 10533

    2020

  55. [55]

    Kendall, Y

    A. Kendall, Y . Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018) 7482–7491

    2018

  56. [56]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, International Conference on Learning Repre- sentations (2019)

    2019

  57. [57]

    Kazakov, J

    A. Kazakov, J. W. Magee, R. D. Chirico, E. Paulechka, V . Diky, C. D. Muzny, K. Kroenlein, M. Frenkel, NIST/TRC web thermo tables: An online resource for critically evaluated thermodynamic property data, Journal of Research of the National Institute of Standards and Technology 119 (2014) 136–156

    2014

  58. [58]

    Loshchilov, F

    I. Loshchilov, F. Hutter, SGDR: Stochastic gradient descent with warm restarts, International Conference on Learning Representations (2017)

    2017

  59. [59]

    Paszke, et al., PyTorch: An imperative style, high-performance deep learning library, Advances in Neural Information Processing Systems 32 (2019)

    A. Paszke, et al., PyTorch: An imperative style, high-performance deep learning library, Advances in Neural Information Processing Systems 32 (2019)

    2019

  60. [60]

    M. Fey, J. E. Lenssen, Fast graph representation learning with PyTorch Geometric, ICLR Workshop on Repre- sentation Learning on Graphs and Manifolds (2019)

    2019

  61. [61]

    Karthikeyan, R

    M. Karthikeyan, R. C. Glen, A. Bender, General melting point prediction based on a diverse compound data set and artificial neural networks, Journal of Chemical Information and Modeling 45 (3) (2005) 581–590. 48

    2005

  62. [62]

    Crawshaw, Multi-task learning with deep neural networks: A survey, arXiv preprint arXiv:2009.09796 (2020)

    M. Crawshaw, Multi-task learning with deep neural networks: A survey, arXiv preprint arXiv:2009.09796 (2020)

    Pith/arXiv arXiv 2009

  63. [63]

    Q. Li, Z. Han, X.-M. Wu, Deeper insights into graph convolutional networks for semi-supervised learning, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 32, 2018, pp. 3538–3545

    2018

  64. [64]

    McInnes, J

    L. McInnes, J. Healy, J. Melville, UMAP: Uniform manifold approximation and projection for dimension re- duction, arXiv preprint arXiv:1802.03426 (2018)

    Pith/arXiv arXiv 2018

  65. [65]

    J. S. Delaney, ESOL: Estimating aqueous solubility directly from molecular structure, Journal of Chemical Information and Computer Sciences 44 (3) (2004) 1000–1005. 49

    2004