Mol-JEPA: A multimodal Joint Embedding Predictive Architecture for Molecules

Carsten Eickhoff; Florian Rottach; Randall Balestriero; Sebastian Schieferdecker; William Rudman

REVIEW

Mol-JEPA: A multimodal Joint Embedding Predictive Architecture for Molecules

Florian Rottach, Sebastian Schieferdecker, William Rudman, Randall Balestriero, Carsten Eickhoff

Machine Learning · Submitted 23 Aug 2026

Pith's verdict: Mol-JEPA is a solid, unusually transparent multimodal pretraining contribution with real ablations and open code, but its headline numbers are inflated by oracle probe selection and the paper never checks whether its 4.7M-molecule corpus contains the benchmark molecules it later evaluates on.

T0 · AI review

How this review was made: An AI referee (deepseek-v4-flash) read the paper on August 27, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 3 major objections · 6 minor · Ask about this paper · The compiled record · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2608.22642 (version 4)

Pith IDOC55K2EJ, Pith's permanent record of this paper and its review

CitationCarsten Eickhoff, Florian Rottach, Randall Balestriero, Sebastian Schieferdecker, William Rudman. "Mol-JEPA: A multimodal Joint Embedding Predictive Architecture for Molecules." arXiv:2608.22642 (2026). https://pith.science/paper/OC55K2EJ

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

Dear colleague,

Punchline: Mol-JEPA is worth your time. The paper builds a genuinely new object: a JEPA-style model pretrained on 14 molecular modalities, from graph structure and DFT descriptors to cell-painting embeddings and sparse ADMET label vectors, on a 4.69M-molecule corpus. That scale and breadth is new for the molecular-JEPA line, and the authors ship code, data, and Hugging Face integration, which makes the empirical claims checkable. They also do the homework most papers skip: leave-one-modality-out, data-scaling analysis, training-dynamics plots, and a candid limitations section that admits the ablations ran on 100k molecules, not the full corpus. Credit where it is due: this is a serious empirical effort, not a demonstration.

The main soft spot is also the main threat to the headline claims. Section 3.1 builds the pretraining corpus from ChEMBL, TDC, PCBA, and other public sets. The benchmarks are ExpansionRx, ASAP-Polaris, PXR, and Biogen ADME, and TDC alone contains overlapping ADMET endpoints (Caco-2, solubility, clearance, hERG). Nowhere in the paper is an exclusion step reported: no statement that benchmark molecules were removed from pretraining, no overlap analysis. The cluster splits in Section 3.5 reduce structural similarity within each benchmark, but they do nothing about cross-corpus contamination. If a molecule appears in pretraining with its endpoint label, the low-data and out-of-distribution gains could be memorization. This is a load-bearing concern, though not necessarily fatal: because the data and code are public, a referee can run the overlap check directly. The authors should be required to do it.

A second, smaller issue: the 'Mol-JEPA (Best)' column in Table 3 selects the best probe per dataset on the test set. That inflates the averages. The 'Transformer' variant alone is still competitive but less dominant, and the paper's own Wilcoxon analysis shows TabICLv2 with AlvaDesc wins more pairwise comparisons (47% vs 38%). The abstract says 'strong performance,' which is fair; it does not say 'state of the art everywhere,' which is also fair. But the oracle selection should be disclosed in the main text, not just in the table caption.

Minor: the reliance on frozen embeddings from UMA, ChemGPT, CLOOME, BioXMol, Boltz-2, MOE, xTB, and DFT means the pretraining objective can absorb upstream biases. The authors acknowledge this indirectly, and the leave-one-out analysis shows redundancy, so I would not call it a flaw so much as an assumption to flag.

Overall: this paper deserves a serious referee. The contribution is real, the execution is unusually honest, and the contamination question is answerable with the released artifacts. I would send it to peer review with the request that the authors add a contamination check and commit to a single downstream probe protocol before acceptance.

Best, [Your name]

Referee report

Significance

If the empirical claims hold, Mol-JEPA is a meaningful step toward multimodal molecular representation learning: it avoids chemically invalid augmentations by masking entire modalities, provides a public pretraining corpus of nearly five million molecules with 14 diverse modalities, and releases code and checkpoints. The paper also includes unusually extensive analyses: leave-one-modality-out experiments, CKA representational similarity, GTM-based coverage analysis, and paired Wilcoxon significance testing. These strengths make the contribution potentially valuable to the molecular machine learning community. However, the headline performance claims are weakened by two load-bearing issues: the lack of any reported decontamination between the pretraining corpus and the benchmark sets, and the selection of the 'Best' probe variant on the test set. The paper's own Wilcoxon analysis also shows that TabICLv2 with AlvaDesc descriptors wins more pairwise comparisons than Mol-JEPA, which is in tension with the abstract's 'outperforming baselines' wording.

Major objections (3)

[Section 3.1 and Section 3.5] The pretraining corpus contains molecules and labels that directly overlap topically with the evaluation benchmarks: TDC contributes ADMET endpoints including Caco-2 permeability, solubility, clearance, and hERG, and ChEMBL contributes bioactivity data, while the benchmarks include Caco-2, LogD, kinetic solubility, HLM/MLM clearance, hERG-related endpoints, and PXR activity. The cluster splits described in Section 3.5 only reduce structural similarity within each benchmark dataset; they do not remove benchmark molecules from the pretraining set, nor do they prevent exact or near-identical molecules from appearing in both. No deduplication or overlap analysis between the pretraining corpus and the benchmark molecules is reported. Consequently, the low-data and out-of-distribution gains claimed in Section 4 and Figure 2b could be driven by the model having memorized molecules and their endpoint values during pretraining rather than by transferable structure-property understanding. The authors should either perform an InChI-based decontamination step and re-run the evaluation, or demonstrate that the overlap is negligible (for example, by reporting exact-match counts and nearest-neighbor similarity distributions for each benchmark against the pretraining corpus).
[Table 3 and Section 3.7] The 'Mol-JEPA (Best)' column is defined as the best performance achieved across all probe and embedding variants, but the text does not state on which data the selection was made. If the choice of probe (linear, nonlinear, transformer, TabICLv2, LoRA) and embedding variant is made using the test set, the reported average MAE of 0.402 is an optimistically biased estimate and the comparison against baselines is not apples-to-apples, since the baselines are selected via cross-validation on the training split. The 'Mol-JEPA (Transformer)' column (average MAE 0.427) is a more honest fixed-probe comparison, and it still beats most baselines but by a smaller margin. The authors should clarify the model-selection protocol, and either report a validation-selected fixed configuration as the primary result or treat 'Best' explicitly as an upper-bound analysis. If the 'Best' numbers were selected on the test set, they should not be the headline result.
[Section 4, Figure 3a, and Appendix D] The paper's own statistical analysis shows that TabICLv2 with AlvaDesc descriptors wins 47% of pairwise Wilcoxon comparisons versus Mol-JEPA's 38%, and Appendix D reports that on the public competition splits TabICLv2 AlvaDesc achieves higher MAE and Wilcoxon win rates as well. This directly qualifies the abstract and conclusion statements that Mol-JEPA 'outperforms our baselines' and 'delivers strong performance' across the board. The authors should temper the central claim and provide a more nuanced discussion: Mol-JEPA appears to be particularly competitive on smaller datasets (ASAP, Biogen) and in terms of R2, while TabICLv2 retains advantages on larger datasets and in MAE/Wilcoxon terms. As written, the discrepancy between the summary claims and the detailed statistical results is too large to ignore.

Minor objections (6)

[Section 2] In the paragraph on multimodal architectures, the phrase 'using contrastive learning learning' contains a duplicated word; it should read 'using contrastive learning.'
[Table 2] The Boltz-2 embedding dimension is 6912, which is substantially larger than the other modalities (e.g., ChemGPT 2048, BioXMol 1024); please clarify whether this dimension is used directly or projected before entering the shared latent space, since this affects the interpretation of the modality dimensions.
[Section 3.4, Eq. (1)] The notation in Eq. (1) is confusing because the same symbol s-hat is used for the masked target embedding and for the predicted embedding; please introduce separate notation, for example s_m_i for the target and p_m_i for the prediction, to make the objective unambiguous.
[Figure 2b] The caption states that the lines show the average MAE for molecules across all datasets binned by Tanimoto similarity, but it is not clear how the aggregation is performed across datasets with very different label scales and units; please specify whether each dataset is normalized before averaging or whether the raw values are pooled.
[Appendix A, Table 6] In Table 6, the PXR (pEC50) task is listed under the OpenADMET grouping, whereas the main text describes PXR as a separate blind challenge; please adjust the table grouping or the text to be consistent.
[Section 4, 'Modality scaling'] The claim that 'we observe a better performance when increasing the number of modalities' is supported only by an annotated curve in Figure 4a; please state explicitly which downstream metric is being reported (e.g., average probe MAE on which benchmark family) and report the corresponding error bars.

How Pith checks papers → · Full audit of this paper → · Review history (2 revisions) →

The paper

The paper claims

Predicting masked biochemical views beats structure-only molecular models

The reading

This paper tries to establish that a molecular representation is stronger when it is trained to predict the latent embeddings of many biochemical modalities—structure, cell-painting phenotype, binding affinity, ADMET measurements, quantum-chemical descriptors—than when it is trained from molecular structure alone. The proposed model, Mol-JEPA, masks whole modalities during pretraining and reconstructs their embeddings in latent space, sidestepping chemically invalid augmentations and contrastive negative sampling. On open benchmark suites, the learned representations beat standard baselines on average and show their largest edge on small datasets and on molecules far from the training distribution. That matters for drug discovery, where labeled data are scarce, noisy, and expensive to collect.

What carries the argument

The load-bearing mechanism is modality masking inside a Joint Embedding Predictive Architecture. Rather than perturbing a molecule's graph or SMILES string, Mol-JEPA randomly deactivates entire modalities—each modality is one of 14 views, such as a cell-painting embedding, a binding-affinity vector, or an ADMET label vector—and trains a Transformer encoder to predict the masked latent embeddings from the remaining ones. Three learnable encoders (atom, graph, vector) project all modalities into a shared space, and SIGReg applies an Epps-Pulley isotropy test on random 1D projections of the embeddings to prevent representation collapse. The prediction error is the mean squared error $\|\hat{s}^m_i - \hat{s}^m_{i,\mathrm{pred}}\|_2^2$ summed over masked modalities.

Extended reading notes

Core claim

On its own terms, the paper's central claim is that modality-masked latent prediction creates molecular world models. Mol-JEPA combines 14 views of a molecule, from frozen embeddings of upstream models (atomistic, SMILES, cell-painting, binding-affinity predictors), computed descriptors (MOE, ECFP, GFN2-xTB, DFT), and experimental label vectors (ChEMBL, PCBA, TDC), across 4.69 million curated compounds. During training a Transformer predictor sees a random subset of encoded modalities and predicts the mean-squared error of the masked ones; Sketched Isotropic Gaussian Regularization (SIGReg) keeps the latent space from collapsing. The authors report that the best Mol-JEPA variant reaches an average MAE of 0.402 across 22 benchmark tasks, below every baseline (best baseline, TabICLv2 with AlvaDesc descriptors, 0.453), and that the advantage grows on small benchmark families and with Tanimoto distance from the training set. They also report that adding modalities consistently improves downstream performance and that removing the experimental modalities degrades it.

Load-bearing premise

The claim rests on the premise that the frozen embeddings and experimental label vectors used as 'views' of a molecule are trustworthy enough that predicting them teaches real chemistry instead of reproducing upstream artifacts, noise, or redundant biases.

Editorial extensions

If this is right

Adding modalities improves downstream performance; training time grows roughly linearly with modality count, so the framework can absorb new data sources predictably.
Mol-JEPA's edge is largest on low-data benchmarks (ASAP/Polaris, Biogen ADME) and on molecules with low Tanimoto similarity to the training set, which is the regime drug-discovery teams actually face.
Full multimodal pretraining cuts downstream MAE by about 14% (linear probe) and 13% (nonlinear probe) relative to training on molecular graph and ECFP4 fingerprints alone.
In leave-one-modality-out training, the graph modality contributes most, but experimental modalities still add value, supporting the paper's thesis that biochemical context is informative beyond structure.
A lightweight Transformer probe over modality-specific embeddings and TabICLv2 over CLS tokens are the strongest downstream uses, while LoRA fine-tuning underperforms.

Reading between the lines

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

Our inference: because the framework never invents molecular edits, it should transfer to other data-scarce scientific domains where multiple noisy, partially paired readouts of the same object exist, such as materials synthesis or cellular screening.
Our inference: Mol-JEPA's ceiling is set by the quality of its frozen upstream embeddings; replacing the atomistic, cell-painting, or binding-affinity encoders with newer models should improve the learned representations with no architectural change.
Our inference: the reconstruction objective may act as a denoiser that is robust to label noise, but a direct test—scrambling one modality's embeddings across molecules and measuring downstream degradation—would separate genuine cross-modal signal from mere pretraining regularization.
Our inference: the reported gain on small datasets suggests Mol-JEPA embeddings could serve as a fixed feature extractor for active-learning loops, where assay budget, not model capacity, is the constraint.

Original abstract

Despite recent advances in molecular foundation models, several limitations remain, such as chemically invalid augmentations, modality collapse, and incomplete representation of biochemical environments. To address these challenges, we present \textbf{Mol-JEPA}, a scalable framework for learning molecular world models. Rather than relying on suboptimal molecular perturbations, our model uses modality masking to exploit information from molecular structures, cellular phenotypes, binding affinities, ADMET profiles, quantum chemistry simulations and other drug discovery data. Across various benchmarks, we show that the representations learned by Mol-JEPA deliver strong performance, demonstrating the value of incorporating biochemical context through latent space prediction.

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 3,638 input tokens.

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.12 · human participants 0.05 · controlled comparison 0.14 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.34 · mentions code or data availability 0.03.

The abstract, sentence by sentence

arXiv listing abstract v4. Results and the main claim are underlined; show every role.

Despite recent advances in molecular foundation models, several limitations remain, such as chemically invalid augmentations, modality collapse, and incomplete representation of biochemical environments. To address these challenges, we present \textbf{Mol-JEPA}, a scalable framework for learning molecular world models. Rather than relying on suboptimal molecular perturbations, our model uses modality masking to exploit information from molecular structures, cellular phenotypes, binding affinities, ADMET profiles, quantum chemistry simulations and other drug discovery data. Across various benchmarks, we show that the representations learned by Mol-JEPA deliver strong performance, demonstrating the value of incorporating biochemical context through latent space prediction.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	Despite recent advances in molecular foundation models, several limitations remain, such as chemically invalid augmentations, modality collapse, and incomplete representation of biochemical environments.	background 0.61	0.07	0.03	asserted 1.00
s1	To address these challenges, we present \textbf{Mol-JEPA}, a scalable framework for learning molecular world models.	contribution 1.00	0.78	0.07	asserted 0.96
s2	Rather than relying on suboptimal molecular perturbations, our model uses modality masking to exploit information from molecular structures, cellular phenotypes, binding affinities, ADMET profiles, quantum chemistry simulations and other drug discovery data.	method 0.98	0.83	0.08	asserted 1.00
s3	Across various benchmarks, we show that the representations learned by Mol-JEPA deliver strong performance, demonstrating the value of incorporating biochemical context through latent space prediction.	result 0.94	0.96	0.78	asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.56.

What was examined, and what was not

sources read

arXiv listing abstract v4

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

Figures

**Figure 1.** Mol-JEPA: Rather than representing molecules solely through their structure, our model learns from a spectrum of biochemical and physical modalities. During training, modalities are randomly masked and predicted from the remaining inputs using a Transformer-based predictor. To prevent representation collapse, we employ isotropic regularization through SIGReg. 1 arXiv:2608.22642v1 [cs.LG] 23 Aug 2026

**Figure 2.** Mol-JEPA evaluation results. 3.7 MOL-JEPA DOWNSTREAM PREDICTION There exist multiple strategies for using Mol-JEPA embeddings for downstream tasks. First, we use the CLS token to extract a global summary of all modalities. We feed this representation of size 512 into a linear and non-linear probe and experiment with TabICLv2 as a predictor. We use default parameters and do not perform any hyperparameter optimization. Secondly, we use the modality-specific predicted embeddings in an aggregative probe. We choose a 2 layer transformer encoder that receives tokens of size 512 for each…

**Figure 3.** Mol-JEPA training results. (a) Performance increases with more modalities. The lines represent the training time and epochs on a 100k subset dataset, annotated with the average downstream MAE. (b) Leave-one-modality-out analysis. We analyze the effect of removing modalities during training. The reported scores are the average downstream probe MAEs

**Figure 4.** Modality analysis nation is that it is the only modality with a trainable GNN backbone, whereas the remaining modalities rely on frozen representations. We also find that the molecular fingerprint modalities MOE and ECFP4 rank among the most important modalities, indicating that their expert-designed features contain valuable information for downstream tasks. Lastly, we find that modalities capturing chemical knowledge through binding affinities and cellular interaction profiles also contribute to the learned representations, highlighting the value of incorporating biological…

**Figure 5.** Multimodal training dynamics. Effective rank, prediction loss and SIGReg loss for different modalities during training. While effective rank is continuously increasing, the losses exhibit instabilities for some modalities. works Kamai et al. (2026), showing that different modalities might require different approaches and not all might be suitable for joint embedding predictive architectures

**Figure 6.** Losses and downstream error. 5 CONCLUSION In this work, we presented Mol-JEPA, a multimodal foundation model that enables stable learning from many molecular modalities by predicting their latent representations. Using various high quality benchmarks, we show that Mol-JEPA achieves strong performance across different downstream tasks. Furthermore, we statistically compare the predictions and find that our model is particularly better on smaller datasets and in out-of-distribution settings. Additionally, we analyze the impact of individual modalities on training dynamics and…

**Figure 7.** Cluster splits t-SNE visualization for ExpansionRx LogD dataset using ECFP4 fingerprints.

**Figure 8.** Loss parameters performance

**Figure 9.** Mean absolute error of different probe variants across datasets.

**Figure 10.** Mean absolute error of different feature variants across datasets.

**Figure 11.** Win rates on public competition splits light two representative examples on the cluster splits, demonstrating Mol-JEPA’s ability to significantly outperform the evaluated baseline methods in predictive accuracy. We further evaluate model win rates on the public splits of ExpansionRX, ASAP, and PXR. These splits are generated using a temporal strategy, which, according to the competition organizers, better represents realistic drug discovery settings by preserving the chronological order of data generation. As shown in

**Figure 13.** Pretraining data analysis Molecular space coverage To understand the role of the pretraining data for the downstream tasks, we analyze the molecular similarity using generative topographic mappings (GTM) Bishop et al. (1998). We randomly sample 5,000 molecules from the pretraining corpus and 2,000 molecules from the benchmark datasets, convert them into ECFP4 fingerprints, and project them into a shared GTM space. The resulting visualization, shown in Figure 13a, reveals that while some benchmark datasets are well represented within the pretraining distribution, others occupy more…

**Figure 14.** Training configurations E PRETRAINING DATA SCALING We analyze the effect of pretraining dataset size on downstream performance. As shown in Figure 14a, performance steadily improves with increasing pretraining data for both linear and nonlinear probes across all downstream datasets (reflected by the confidence bands). To better highlight this trend, performance is reported as 1-MAE, where higher values indicate better performance. The observed scaling behavior suggests that Mol-JEPA benefits from additional pretraining data and further increases in dataset size are expected to…

**Figure 15.** Multimodal performance gains. We re-train the model on the full dataset using only Graph and ECFP4 modalities. We find that the downstream performance is significantly better when using all modalities

**Figure 16.** Centered Kernel Alignment specific representations according to their underlying biological and physicochemical relationships while preserving information unique to individual modalities. H IN-CONTEXT LEARNING COMPARISON We assess the effectiveness of Mol-JEPA representations for in-context learning by comparing them with widely used molecular representations. As illustrated in Figures 17-19, Mol-JEPA CLS-token embeddings achieve superior performance across most tasks for different TabICLv2 hyperparameters. This finding indicates that the multimodal architecture effectively…

**Figure 17.** OpenADMET PXR TabICLv2 comparison

**Figure 18.** OpenADMET Caco2 Efflux TabICLv2 comparison.

**Figure 19.** OpenADMET LogD TabICLv2 comparison. 26

Showing 18 of the paper's 19 figures. See all figures in the PDF ↗ Figures reproduced from arXiv:2608.22642 by the authors.

References

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Can phenotypic activity be predicted without experimental readouts?
cs.LG 2026-10 conditional novelty 6.0 of 10

Pretrained molecule-morphology encoders (CLOOME, CellCLIP) show no advantage over physicochemical descriptors once pretraining leakage and the activity/toxicity confound are controlled.
How 'Foundational' Are Current Molecular Foundation Models?
q-bio.BM 2026-09 conditional novelty 4.0 of 10

None of today's molecular foundation models satisfies all three proposed tests of generality, transferability, and out-of-distribution generalization, and simple fingerprint baselines often keep pace with the large models.

How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 8 sentences without a confident label

Uses the method or tool of 3 papers

To avoid mode collapse, we apply Sketched Isotropic Gaussian Regularization (SIGReg) [ref] .

LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics · Model Architecture · 0.99
Mol-JEPA is implemented using the stable-pretraining framework [ref] and hyperparameter-tuned with Optuna [ref] over the search space specified in Appendix [ref], resulting in a final model with around 50 million parameters.

2511.19484 · Model Architecture · 1.00
For in-context predictions, we use the same feature sets along with different variants of TabICLv2 [ref] , a tabular foundation model that doesn't require any significant hyperparameter tuning.

TabICL: A Tabular Foundation Model for In-Context Learning on Large Data · Baselines · 0.97

Uses data or values from 2 papers

In addition, we integrate 21 ADME, 11 toxicity and 12 high-throughput screening endpoints from Therapeutic Data Commons (TDC) [ref] .

Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development · Pretraining Data · 1.00
We also use PCBA-1328 [ref] , a subset of PubChem [ref] , containing 1,328 bioassays with "Active" or "Inactive" flags with at least 10 labels per assay.

Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets · Pretraining Data · 1.00

Cites as support 2 papers

Joint embedding approaches have been shown to be more robust to variations in the input and target spaces [ref] , making them particularly interesting for molecular data.

2505.12477 · Introduction · 0.98
The experienced complexity of multimodal pretraining is in line with recent works [ref] , showing that different modalities might require different approaches and not all might be suitable for joint embedding predictive architectures.

When to Align, When to Predict: A Phase Diagram for Multimodal Learning · Results · 0.97

Cites as background 3 papers

For example, several methods have generated different views on a molecule by augmenting molecular graphs, which are jointly encoded or predicted [ref] .

GraphCL: Contrastive Self-Supervised Learning of Graph Representations · Related Work · 0.89
Early works adapted popular neural network architectures to molecular data, yielding models such as ChemBERTa [ref] , Chemical VAE [ref] or Chemformer [ref] .

ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction · Related Work · 1.00
Graph Neural Networks (GNNs) [ref] have become a prominent approach for molecular representation learning, as they integrate inductive priors that allow them to operate directly on molecular graphs.

Semi-Supervised Classification with Graph Convolutional Networks · Related Work · 0.96

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.