Research map / Vision-language and multimodal / Human and embodied modeling
Uncertainty-Aware Yield Prediction with Multimodal Molecular Features
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UAM, a multimodal (SMILES transformer, molecular graph, human-defined features with MoE) yield prediction model with contrastive pretraining and uncertainty-aware heteroscedastic loss, evaluated on two HTE datasets and an Amide coupling literature dataset.
Problem
Reaction yield prediction is hurt by high sensitivity of yields to small structural changes and by noisy yield measurements in real-world literature data, and single-modality molecular representations fail to capture these effects.
Approach
Fuses SMILES, molecular graph, and expert-defined reaction features (via a sparse MoE encoder) with cross-modal contrastive pretraining, late fusion, and a learned Gaussian yield distribution with heteroscedastic loss plus a regularization term to model data and model uncertainty.
Key results
- On the ACR dataset, UAM reaches R2 0.262 vs 0.216 for the best baseline (YieldGNN), with MAE 14.76 and RMSE 19.33.
- On Buchwald-Hartwig, UAM achieves R2 0.976 ± 0.001 and MAE 2.89, slightly ahead of MPNN (R2 0.974).
- On Suzuki-Miyaura, UAM achieves R2 0.888 ± 0.009 and MAE 6.04, marginally above MPNN (R2 0.886).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UAM instead of YieldGNN | Amide coupling reaction yield prediction · Amide coupling reaction | MAE: 15.27 ± 0.18 -> 14.76 ± 0.15 | Helped |
| Use UAM instead of MPNN | Buchwald–Hartwig reaction yield prediction · Buchwald–Hartwig reactions | MAE: 2.92 ± 0.06 -> 2.89 ± 0.06 | Helped |
| Use UAM instead of MPNN | Suzuki–Miyaura reaction yield prediction · Suzuki–Miyaura reactions | MAE: 6.12 ± 0.22 -> 6.04 ± 0.18 | Helped |
| Add uncertainty quantification loss (ablation) | UAM · Amide coupling reaction | MAE: 15.08 ± 0.13 -> 14.76 ± 0.15 | Helped |
| Add regularized dropout loss (ablation) | UAM · Amide coupling reaction | MAE: 14.80 ± 0.16 -> 14.76 ± 0.15 | Helped |
| Use MoE instead of an equally layered FFN (ablation) | UAM · Amide coupling reaction | MAE: 15.12 ± 0.18 -> 14.76 ± 0.15 | Helped |
| Add SMILES sequence features (ablation) | UAM · Amide coupling reaction | MAE: 14.97 ± 0.16 -> 14.76 ± 0.15 | Helped |
| Add graph features (ablation) | UAM · Amide coupling reaction | MAE: 15.06 ± 0.15 -> 14.76 ± 0.15 | Helped |
| Add human-defined features (ablation) | UAM · Amide coupling reaction | MAE: 15.83 ± 0.20 -> 14.76 ± 0.15 | Helped |
Surprising findings
- Gains over baselines are largest on the noisy literature ACR dataset and marginal on the HTE datasets, where most baselines already exceed R2 0.85-0.95.
- Removing the SMILES or graph modality has only a small effect on ACR performance, while removing human-defined features causes the largest drop.
Limitations
- All models perform poorly on the ACR dataset (R2 below 0.5), so absolute predictive power remains limited.
- Improvements on HTE datasets are small, and the paper does not test on broader reaction types beyond these datasets.
- Does not incorporate 3D molecular structure, which the authors flag as future work.
Experimental setup
Datasets: Buchwald-Hartwig HTE dataset Suzuki-Miyaura HTE dataset Amide coupling reaction (ACR) dataset
Benchmarks: Amide coupling reaction (ACR) dataset Buchwald-Hartwig HTE dataset Suzuki-Miyaura HTE dataset
Metrics: MAE RMSE R2
Baselines: One-hot YieldBert MPNN YieldGNN Mordred
Builds on: YieldBert MPNN Mixture of Experts Kendall and Gal Bayesian uncertainty loss Wu et al. 2021 regularized dropout
Closest papers by meaning
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.89)
- Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets (NeurIPS 2025 · similarity 0.87)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.87)
- Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures (ICML 2025 · similarity 0.86)
- The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning (NeurIPS 2025 · similarity 0.85)
- ReactGPT: Understanding of Chemical Reactions via In-Context Tuning (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.