Molecular Property Prediction under Structural Shift with Tabular Foundation Models

Jinmo Lee; Dooho Lee; Minho Jeong; Jaemin Yoo

cs.LGSep 30, 2026

Molecular Property Prediction under Structural Shift with Tabular Foundation Models

Authors: Jinmo Lee, Dooho Lee, Minho Jeong, Jaemin Yoo

Organizations: Nums AI

Abstract

Predicting molecular properties for compounds that differ structurally from labeled training molecules is important for drug discovery and materials design. Tabular foundation models (TFMs) offer a promising approach through in-context learning, but their performance under structural shifts and the value of molecular comparisons in this setting remain underexplored. We study structural generalization in molecular property prediction and introduce MolPAIR (Molecular Pair-Augmented In-context Refinement), a framework that combines molecule-level and molecular-pair contexts without task-specific parameter updates. A global tabular foundation model (TFM) first predicts a query's property from labeled molecular examples. A second frozen TFM predicts differences in prediction errors between the query and labeled reference molecules, using these comparisons to refine the initial prediction. Across 58 MoleculeACE and Polaris tasks, CheMeleon representations combined with TabPFN-3 already outperform each evaluated baseline on a majority of tasks. MOLPAIR further improves this predictor on 46 of 58 tasks, with gains across four molecular representations and three TFM backbones. These results show that explicit molecular comparisons can strengthen tabular in-context learning for structural generalization while keeping the molecular encoder and pretrained model weights fixed. The code and datasets are available at https://github.com/nums-ai/MolPAIR.

Figures & tables

Figure 1: Pairwise task-win matrix. Each cell counts the tasks on which the row method outperforms the column method based on seed-averaged primary metrics. * indicates significance (two-sided exact sign test, Holm-adjusted $p<0.05$ ).

Figure 2: MolPAIR framework. The blue path predicts the query’s property while the purple path uses the relation context to predict error differences. Snowflakes mark frozen components.

Method	MACE RMSE $\downarrow$	MACE rank $\downarrow$	Polaris cls. rank $\downarrow$	Polaris reg. rank $\downarrow$	Top one $\uparrow$	Elo $\uparrow$
MolPAIR	0.885 $\pm$ 0.100	2.00	2.18	2.59	29	1360 [1289, 1455]
Global baseline	0.893 $\pm$ 0.097	2.93	2.36	4.00	7	1246 [1196, 1310]
MiniMol (tuned)	0.931 $\pm$ 0.107	5.30	3.91	3.12	14	1122 [1060, 1190]
ExtraTrees (tuned)	0.921 $\pm$ 0.102	4.90	5.36	5.12	2	1062 [1007, 1123]
CheMeleon (fine-tuned)	0.930 $\pm$ 0.089	5.40	6.91	4.24	3	1035 [976, 1097]
XGBoost (tuned)	0.924 $\pm$ 0.101	5.03	6.36	5.59	2	1026 [969, 1088]

Table 1: Structural generalization across 58 tasks. MACE RMSE is mean

\pm

SD across tasks after averaging seeds; ranks are task means and Top one counts first places. Brackets give 95% Elo intervals; bold marks the best values.

Cutoff	Global RMSE	MolPAIR RMSE	W/L
0.40	0.9939 $\pm$ 0.1745	0.9926 $\pm$ 0.1792	38/14
0.50	0.9570 $\pm$ 0.1318	0.9521 $\pm$ 0.1302	41/13
0.60	0.8932 $\pm$ 0.0966	0.8849 $\pm$ 0.1004	46/12
0.70	0.8357 $\pm$ 0.0791	0.8327 $\pm$ 0.0790	42/16

Table 2: Sensitivity to the training-similarity cutoff. RMSE is task mean

\pm

SD on MoleculeACE; W/L counts MolPAIR wins/losses across retained tasks. Bold marks lower RMSE.

Figure 3: Elo across representations and backbones. Labels give MolPAIR -minus-global

\Delta

Elo; error bars show 95% intervals. Absolute ratings are comparable only within pairs.

Control	Pooled $\Delta$ Elo [95% CI]	W/L	RMSE reduction [95% CI]
MolPAIR (full)	Reference	–	Reference
Global baseline	295.7 [209.6, 393.9]	46/12	0.00819 [0.00235, 0.01388]
Two-global-TFM ensemble	209.6 [111.1, 321.3]	44/14	0.00547 [ $-$ 0.00032, 0.01093]
$k$ NN residual	112.0 [37.9, 191.7]	41/17	0.00402 [ $-$ 0.00110, 0.00930]
Single-molecule residual	160.9 [78.7, 253.9]	43/15	0.00768 [0.00251, 0.01295]
Shuffled relation labels	88.9 [14.0, 168.8]	41/17	0.00405 [ $-$ 0.00100, 0.00947]

Table 3: Prediction and correction controls. Pooled

\Delta

Elo and W/L compare full MolPAIR with each control on 58 tasks; RMSE reductions use MoleculeACE. Positive values favor MolPAIR ; brackets give 95% endpoint-bootstrap intervals.

Figure 4: Elo versus computational cost. Higher Elo and lower wall-clock time are better (upper left).

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Global representation	CheMeleon, 2,048 input features before filtering
Relation representation	RDKit2D, up to 996 selected features
Global and relation backbone	TabPFN-3, eight estimators
Cross-fitting	Five folds within the outer training split
Training anchors	Up to eight non-self anchors within the same OOF fold
Relation context budget	At most 8,192 rows, deterministic subsampling

Appendix

Table 4: Default context and inference settings.

Suite	Endpoints	Planned seeds	Task types	Primary metrics
MoleculeACE	30	5	regression	RMSE
Polaris	28	5	mixed	MAE , MSE , Pearson, Spearman, PR-AUC , ROC-AUC
Total	58	5	mixed	released metric

Appendix

Table 5: Benchmark tasks and metrics. Planned seeds are per endpoint.

Max. similarity	Valid splits	Query–seed predictions	Endpoints	Direct $\Delta$ Elo
0.40	219 / 290	26,488	52	171.9
0.50	268 / 290	42,159	54	197.7
0.60	286 / 290	55,733	58	231.1
0.70	289 / 290	72,474	58	166.3

Appendix

Table 6: Retained populations by similarity cutoff. Valid splits are out of 290 planned task–seed combinations. Prediction counts include repeats across seeds. Positive direct

\Delta

Elo favors MolPAIR over the global predictor.

Representation	Dimension	Source or use
RDKit2D	data dependent	deterministic descriptors
Morgan	1,024 to 4,096	radius two or three, bit or count
CheMeleon	2,048	frozen directed message-passing fingerprint
MiniMol	512	frozen pretrained GINE fingerprint
Monroe	released dimension	frozen pretrained molecular fingerprint

Appendix

Table 7: Molecular representations before train-only filtering.

Model and trials	Search space
Random Forest, 30	max_features in $\{\sqrt{d},\log_{2}d,0.2,0.5,0.8,1.0\}$ ; max_depth in $\{\texttt{None},8,16,32,64\}$ ; min_samples_leaf from 1 to 16 (log); min_samples_split from 2 to 32 (log); bootstrap on or off; and, for classification, class weight in {none, balanced, balanced subsample}.
ExtraTrees, 30	The same space as Random Forest.
XGBoost, 50	Learning rate from 0.005 to 0.2 (log), depth 3–12, child weight 0.01–20 (log), row subsampling 0.5–1.0, column subsampling 0.4–1.0, gamma 0–5, L1 $10^{-8}$ –10 (log), L2 $10^{-3}$ –100 (log), and maximum bins in $\{128,256,512\}$ .
LightGBM, 50	Learning rate 0.005–0.2 (log), leaves 15–255 (log), depth in $\{-1,4,6,8,12,16\}$ , minimum child rows 5–100 (log), row subsampling 0.5–1.0, column subsampling 0.4–1.0, L1 $10^{-8}$ –10 (log), L2 $10^{-3}$ –100 (log), and maximum bins in $\{127,255,511\}$ .

Appendix

Table 8: Tree-baseline search spaces. Log denotes logarithmic sampling.

Model and trials

Search space

Chemprop, 20

Message width in

\{256,512,768,1024\}

, message-passing depth 3–6, mean or sum aggregation, feed-forward width in

\{128,256,512,1024\}

, one to three feed-forward layers, dropout 0–0.4 in steps of 0.1, batch size in

\{32,64,128\}

, one to five warmup epochs, maximum learning rate

3\times 10^{-5}

–

3\times 10^{-3}

(log), initial-rate ratio 0.02–0.2 (log), final-rate ratio 0.001–0.05 (log), and class balancing on or off for classification.

MiniMol head, 20

Hidden-width multiplier in

\{0.5,1,2\}

, one to three hidden layers, dropout 0–0.3 in steps of 0.1, learning rate

3\times 10^{-5}

–

2\times 10^{-3}

(log), weight decay

10^{-7}

–

10^{-3}

(log), batch size in

\{16,32,64,128\}

, epochs in

\{15,25,40\}

, and warmup epochs in

\{2,5,8\}

Appendix

Table 9: Molecular-baseline search spaces. Log denotes logarithmic sampling.

Regression on 47 tasks
Method	MACE RMSE	RMSE reduction [95% CI]	W/L	Direct $\Delta$ Elo [95% CI]
MolPAIR	0.8849	Reference	–	–
Global baseline	0.8931	0.00828 [0.00246, 0.01427]	40/7	298.1 [184.0, 451.9]
PADRE (RF)	1.0012	0.11632 [0.08687, 0.14614]	44/3	451.9 [298.1, 809.9]
DeepDelta	1.0588	0.17393 [0.14058, 0.20757]	44/3	451.9 [298.1, 809.9]
iMoLD	1.0796	0.19478 [0.15370, 0.24499]	47/0	809.9

Appendix

Table 10: Comparison with specialized molecular methods. W/L and direct

\Delta

Elo compare MolPAIR with each method. RMSE uses MoleculeACE; positive reductions and Elo favor MolPAIR . Brackets give 95% intervals; – denotes unavailable or inapplicable values.

Endpoint	Metric	MolPAIR	Global	MiniMol	XGBoost	Chemprop
1862 Ki	RMSE $\downarrow$	$\mathbf{0.741\pm 0.104}$	$0.747\pm 0.102$	$0.844\pm 0.116$	$0.751\pm 0.058$	$1.028\pm 0.125$
1871 Ki	RMSE $\downarrow$	$\mathbf{0.770\pm 0.102}$	$0.774\pm 0.099$	$0.802\pm 0.105$	$0.815\pm 0.093$	$0.806\pm 0.075$
2034 Ki	RMSE $\downarrow$	$\mathbf{0.940\pm 0.066}$	$0.963\pm 0.068$	$0.996\pm 0.038$	$1.067\pm 0.040$	$1.188\pm 0.102$
2047 EC50	RMSE $\downarrow$	$0.934\pm 0.135$	$0.915\pm 0.109$	$\mathbf{0.881\pm 0.130}$	$0.960\pm 0.118$	$0.912\pm 0.066$
204 Ki	RMSE $\downarrow$	$\mathbf{0.904\pm 0.174}$	$0.911\pm 0.169$	$1.022\pm 0.103$	$0.981\pm 0.092$	$1.077\pm 0.100$
2147 Ki	RMSE $\downarrow$	$\mathbf{0.695\pm 0.066}$	$0.704\pm 0.077$	$0.791\pm 0.052$	$0.827\pm 0.121$	$0.944\pm 0.111$

Appendix

Table 11: MoleculeACE endpoint RMSE. Values are mean

\pm

SD across available seeds. No SD is available for the single-seed endpoint CHEMBL2835_Ki .

Endpoint	Metric	MolPAIR	Global	MiniMol	XGBoost	Chemprop
adme fang hclint 1	Pearson $\uparrow$	$0.696\pm 0.010$	$0.673\pm 0.012$	$\mathbf{0.734\pm 0.029}$	$0.689\pm 0.023$	$0.630\pm 0.039$
adme fang hppb 1	Pearson $\uparrow$	$0.605\pm 0.109$	$0.587\pm 0.106$	$\mathbf{0.723\pm 0.118}$	$0.614\pm 0.122$	$0.226\pm 0.228$
adme fang perm 1	Pearson $\uparrow$	$0.700\pm 0.035$	$0.674\pm 0.038$	$\mathbf{0.762\pm 0.040}$	$0.713\pm 0.038$	$0.641\pm 0.033$
adme fang rclint 1	Pearson $\uparrow$	$0.717\pm 0.008$	$0.703\pm 0.008$	$\mathbf{0.737\pm 0.012}$	$0.711\pm 0.019$	$0.676\pm 0.031$
adme fang rppb 1	Pearson $\uparrow$	$0.594\pm 0.072$	$0.550\pm 0.096$	$\mathbf{0.665\pm 0.108}$	$0.513\pm 0.080$	$0.246\pm 0.369$
adme fang solu 1	Pearson $\uparrow$	$0.582\pm 0.052$	$0.565\pm 0.047$	$\mathbf{0.679\pm 0.024}$	$0.519\pm 0.060$	$0.538\pm 0.072$

Appendix

Table 12: Polaris endpoint metrics. Values are mean

\pm

SD across available seeds; arrows indicate the preferred direction.

Backbone	Representation	Global RMSE	MolPAIR RMSE	$\Delta$ Elo
TabPFN-3	RDKit2D	$0.9298\pm 0.1211$	$\mathbf{0.9227\pm 0.1228}$	+197.2
TabPFN-3	CheMeleon	$0.8932\pm 0.0966$	$\mathbf{0.8849\pm 0.1004}$	+231.1
TabPFN-3	MiniMol	$0.9106\pm 0.1120$	$\mathbf{0.9047\pm 0.1114}$	+290.8
TabPFN-3	Monroe	$0.8918\pm 0.0917$	$\mathbf{0.8861\pm 0.0943}$	+166.3
TabICLv2	RDKit2D	$0.9310\pm 0.1163$	$\mathbf{0.9226\pm 0.1151}$	+166.3
TabICLv2	CheMeleon	$0.9076\pm 0.1070$	$\mathbf{0.8979\pm 0.1054}$	+269.4

Appendix

Table 13: Representation and backbone robustness. RMSE is task mean

\pm

SD on MoleculeACE. Direct

\Delta

Elo uses all 58 tasks; positive values favor MolPAIR . Bold marks lower RMSE within each pair.

Control	Regression W/L	Classification W/L	RMSE reduction [95% CI]	Direct $\Delta$ Elo [95% CI]
Gated single residual	37/10	6/5	0.00768 [0.00260, 0.01292]	181.4 [85.0, 290.8]
Row-matched single residual	41/6	6/5	0.00479 [0.00110, 0.00826]	249.7 [151.7, 368.8]

Appendix

Table 14: Molecular-pair versus single-molecule correction. W/L counts MolPAIR wins/losses. RMSE reductions use MoleculeACE; direct

\Delta

Elo uses all 58 tasks. Brackets give 95% intervals.

Control	W/L	RMSE reduction [95% CI]
In-sample residuals	44/14	0.01547 [0.00839, 0.02274]
Random anchors	41/17	0.00373 [-0.00015, 0.00763]
No bidirectional symmetry	33/25	0.00042 [-0.00015, 0.00105]
No uncertainty weighting	28/30	0.00003 [-0.00038, 0.00045]

Appendix

Table 15: Correction controls on released standard splits. W/L counts MolPAIR wins/losses across 58 tasks. Positive MoleculeACE RMSE reductions favor MolPAIR ; brackets give 95% intervals.

Variant	MACE RMSE	Full W/T/L	Full $-$ variant $\Delta$ Elo [95% CI]
Full reference	0.88496	–	–
No global-output features	0.88577	35/0/23	+72.4 [ $-$ 11.9, 166.3]
Pair predictor, similarity-only weights	0.88505	24/0/34	$-$ 60.1 [ $-$ 151.7, 23.8]
$k$ NN residual, similarity-only weights	0.88899	41/0/17	+151.7 [60.1, 269.4]
Uniform weights	0.88639	41/0/17	+151.7 [60.1, 249.7]
8 query anchors	0.88436	27/0/31	$-$ 23.8 [ $-$ 110.7, 60.1]

Appendix

Table 16: Relation-design sensitivity. RMSE averages 30 MoleculeACE tasks; W/T/L compares the full reference with each variant on 58 tasks. Positive direct

\Delta

Elo favors the reference; brackets give 95% endpoint-bootstrap intervals.

Weighting	Pair RMSE	$k$ NN RMSE	W/L	$\Delta$ Elo [95% CI]
Similarity-only	0.88505	0.88899	40/18	+137.7 [47.9, 249.7]

Appendix

Table 17: Pair prediction versus neighbor-error transfer. RMSE averages 30 MoleculeACE tasks. W/L and positive direct

\Delta

Elo favor pair prediction across 58 tasks; brackets give its 95% endpoint-bootstrap interval.

Gate	MACE RMSE	Adaptive W/L	Direct $\Delta$ Elo [95% CI]
Adaptive	0.885112	–	–
Fixed 0.10	0.889325	44/14	$+197.2$ [97.7, 314.1]
Fixed 0.20	0.886889	32/26	$+35.8$ [ $-47.9$ , 124.0]
Fixed 0.25	0.886215	27/31	$-23.8$ [ $-110.7$ , 60.1]
Fixed 0.2674	0.886066	26/32	$-35.8$ [ $-124.0$ , 47.9]
Fixed 0.30	0.885905	24/34	$-60.1$ [ $-151.7$ , 23.8]

Appendix

Table 18: Adaptive versus fixed reliability-gate coefficients. W/L counts adaptive-gate wins/losses on 58 tasks; positive direct

\Delta

Elo favors the adaptive gate. Brackets give 95% confidence intervals. Bold marks the lowest mean MoleculeACE RMSE.

Method	Hit identification	Lead optimization
	PR-AUC	Cluster Spearman
$k$ NN with ECFP4	0.4606	0.1579
Gradient boosting with ECFP4	0.4594	0.2140
SVM with ECFP4	0.4065	0.3124
Chemprop	0.5040	0.2172
CheMeleon with TabPFN-3	0.5551	0.3004

Appendix

Table 19: Official Lo-Hi results. Values are means across released folds; higher is better.

Method	MoleOOD	ERM	MixUp	Global TFM	MolPAIR	GroupDRO
ROC-AUC	0.6946	0.6739	0.6714	0.6704	0.6687	0.6423

Appendix

Table 20: Official DrugOOD results. Values are mean ROC-AUC across six tasks; higher is better.

	Regression			Classification
Similarity	$N$	Improved (%)	Worsened (%)	$N$	Improved (%)	Worsened (%)
$[0.20,0.30)$	6,824	57.1	41.9	2,494	64.8	33.0
$[0.30,0.40)$	12,415	57.2	42.8	3,624	67.4	32.6
$[0.40,0.50)$	10,596	56.5	43.5	3,757	70.0	30.0
$[0.50,0.60)$	11,298	54.8	45.2	3,071	74.0	26.0
$=0.6$	419	53.2	46.8	104	71.2	28.8

Appendix

Table 21: Correction outcomes by anchor similarity.

N

counts query–seed predictions. Improved/worsened denotes lower/higher absolute error (regression) or Brier loss (classification).

Diagnostic	Regression	Classification	All
Endpoints	47	11	58
Task–seed runs	231	55	286
Query–seed predictions	42,026	13,707	55,733
No-anchor predictions	474	657	1,131
No-anchor fallback (%)	1.13	4.79	2.03
Active correction (%)	98.71	94.81	97.75

Appendix

Table 22: Anchor availability and correction activity. Fallback means no accepted anchor; active correction means a nonzero change. Percentages and means weight query–seed predictions equally and include fallback cases.

Scenario	Global	MolPAIR	MiniMol head
Cold end to end	5.17	46.11	13.91
Warm task adaptation	0.17	13.41	8.60
Update with 64 labels	0.15	4.49	2.09
Warm query with 1 molecule	1.10	1.64	0.002
Warm query with 128 molecules	1.14	3.91	0.003
Warm query with 10,000 molecules	7.16	228.95	0.004

Appendix

Table 23: Setup and query runtime. Values are median seconds across six endpoints.

Explore similar work

Apr 17, 2026cs.LG

Tabular foundation models for in-context prediction of molecular properties

Accurate molecular property prediction is central to drug discovery, catalysis, and process design, yet real-world applications are often limited by small datasets. Molecular foundation models provide a promising direction by learning transferable molecular representations; however, they typically involve task-specific fine-tuning, require machine learning expertise, and often fail to outperform classical baselines. Tabular foundation models (TFMs) offer a fundamentally different paradigm: they perform predictions through in-context learning, enabling inference without task-specific training. Here, we evaluate TFMs in the low- to medium-data regime across both standardized pharmaceutical benchmarks and chemical engineering datasets. We evaluate both frozen molecular foundation model representations, as well as classical descriptors and fingerprints. Across the benchmarks, the approach shows excellent predictive performance while reducing computational cost, compared to fine-tuning, with these advantages also transferring to practical engineering data settings. In particular, combining TFMs with CheMeleon embeddings yields up to 100% win rates on 30 MoleculeACE tasks, while compact RDKit2d and Mordred descriptors provide strong descriptor-based alternatives. Molecular representation emerges as a key determinant in TFM performance, with molecular foundation model embeddings and 2D descriptor sets both providing substantial gains over classic molecular fingerprints on many tasks. These results suggest that in-context learning with TFMs provides a highly accurate and cost-efficient alternative for property prediction in practical applications.

Karim K. Ben Hicham, Jan G. Rittig, Martin Grohe +1

RWTH Aachen University, Process Systems Engineering (AVT.SVT), Aachen, Germany. · Lehrstuhl Informatik 7, RWTH Aachen University, Aachen, Germany. · JARA-CSD, Aachen, Germany. +1

Molecular Property Prediction Molecular Representation Learning Tabular Foundation Models Graph Foundation Model In-Context Learning

Ask PDF

Jun 30, 2026cs.LG

Can Tabular In-Context Learners Generalize to Biomolecular Property Prediction?

Predicting biomolecular properties from limited labeled data is a central bottleneck in protein engineering and small-molecule design. As strong pretrained encoders now supply rich fixed-length representations, the difficulty has shifted from representation learning to building a data-efficient predictor for the few-shot regime. Tabular foundation models such as TabPFN and TabICL are unlikely candidates for this role: they are in-context learners pretrained on synthetic tables drawn from random causal graphs, a generative prior with no obvious correspondence to the processes that produce protein sequences or molecular graphs. That this tabular, causal inductive bias should transfer to biomolecular data at all is counter-intuitive, yet we find it does. Treating each method as a predictor-representation pair, we evaluate across two domains. We find that on protein fitness regression tasks these in-context learning models coupled with ESM Cambrian representations achieve or exceed state-of-the-art results on ProteinGym, and outperform task-specific supervised regressors on a diverse esterase catalytic activity dataset. For small-molecule classification with ECFP/RDKit descriptors, no single predictor-representation pairing dominates across TDC ADMET, MoleculeNet, FS-Mol, and DrugOOD, but they are competitive with the existing task-specific state-of-the-art. Crucially, on both protein and small-molecule few-shot tasks, these predictor-representation pairs offer strong performance. We conclude that tabular foundation models can be strong biomolecular predictors, but only when coupled with expressive representations.

Davy Guan, Lu Zhang, Asiri Wijesinghe +7

Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia

Molecular Property Prediction Tabular Foundation Models Moleculenet Biomedical Knowledge Graph Biomarker Representation Learning+2

Ask PDF

Apr 17, 2026cs.LG

A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the Foundation Model Era

Molecular property prediction integrates quantum chemistry, cheminformatics, and deep learning to connect molecular structure with physicochemical and biological behavior. This survey traces four complementary paradigms, including Quantum, Descriptor Machine Learning, Geometric Deep Learning, and Foundation Models, and outlines a unified taxonomy linking molecular representations, model architectures, and interdisciplinary applications. Benchmark analyses integrate evidence from both widely used datasets and datasets reflecting industry perspectives, encompassing quantum, physicochemical, physiological, and biophysical domains. The survey examines current standards in data curation, splitting strategies, and evaluation protocols, highlighting challenges including inconsistent stereochemistry, heterogeneous assay sources, and reproducibility limitations under random or poorly defined splits. These observations motivate the modernization of benchmark design toward more transparent, time- and scaffold-aware methodologies. We further propose three forward-looking directions: (i) physics-aware learning embedding quantum consistency, (ii) uncertainty-calibrated foundation models for trustworthy inference, and (iii) realistic multimodal benchmark ecosystems integrating computational and experimental data. Repository: https://github.com/Zongru-Li/Survey-and-Benchmarks-of-DL-for-Molecular-Property-Prediction-in-the-Foundation-Model-Era.

Zongru Li, Xingsheng Chen, Honggang Wen +8

The University of Hong Kong, Hong Kong SAR · Nanyang Technological University, Singapore · University of Cambridge, United Kingdom +4

Molecular Property Prediction Quantum Chemistry Foundation Model Deep Learning

Ask PDF