Safe Multitask Molecular Graph Networks for Vapor Pressure and Odor Threshold Prediction

TL;DR

This work tackles joint vapor pressure (VP) and odor threshold (OP) prediction under scaffold-based out-of-distribution evaluation. It combines chemistry-aware graph representations (A20/E17) with a degree-aware backbone (PNA) and late temperature conditioning to establish a strong graph-centric baseline that generalizes across novel scaffolds. A safe multitask training regimen treats OP as a noisy auxiliary signal, introducing delayed activation, a small auxiliary weight, and optional gradient detaching to avoid negative transfer while preserving or improving VP performance. The results show VP accuracy improving over single-task baselines, OP remaining stable as a regularizer, and robust behavior across OOD regimes, with comprehensive ablations and diagnostics guiding practical deployment. The work provides reproducible data, splits, and baselines that support future development of exposure-aware molecular design and risk assessment models.

Abstract

We investigate two important tasks in odor-related property modeling: Vapor Pressure (VP) and Odor Threshold (OP). To evaluate the model's out-of-distribution (OOD) capability, we adopt the Bemis-Murcko scaffold split. In terms of features, we introduce the rich A20/E17 molecular graph features (20-dimensional atom features + 17-dimensional bond features) and systematically compare GINE and PNA backbones. The results show: for VP, PNA with a simple regression head achieves Val MSE $\approx$ 0.21 (normalized space); for the OP single task under the same scaffold split, using A20/E17 with robust training (Huber/winsor) achieves Val MSE $\approx$ 0.60-0.61. For multitask training, we propose a **"safe multitask"** approach: VP as the primary task and OP as the auxiliary task, using delayed activation + gradient clipping + small weight, which avoids harming the primary task and simultaneously yields the best VP generalization performance. This paper provides complete reproducible experiments, ablation studies, and error-similarity analysis while discussing the impact of data noise and method limitations.

Figures (7)

• Figure 1: Validation molecules' maximum Tanimoto similarity to the training set (ECFP4). The mass is centered around 0.35--0.65 and a visible low-similarity tail confirms the out-of-scaffold setting. This motivates reporting OOD robustness rather than random-split performance.
• Figure 2: Training-split node-degree histogram. Most atoms have degree 1--2 with a long 4$+$ tail, motivating degree-aware scaling (amplification/attenuation) in PNA.
• Figure 3: Model architecture. A graph encoder (GINE/PNA) produces a molecule representation $\mathbf{h}$. VP is predicted via a late-fused temperature head using $[\mathbf{h};t]$, while OP uses masked heads for air/water. Safe-MT controls the auxiliary OP contribution with a delayed, low-weight schedule to reduce negative transfer under scaffold OOD.
• Figure 4: Validation molecules' maximum Tanimoto similarity to training (ECFP4). A broad mass around 0.35--0.65 with a low-similarity tail confirms the out-of-scaffold regime and rules out trivial near-duplicate leakage.
• Figure 5: Parity plots for vapor pressure (VP) on the scaffold-split test set in the normalized log$_{10}P$ space. Each point corresponds to a (molecule, temperature) pair. Left: single-task PNA + A20/E17 (ST-VP); right: safe multitask (VP+OP, PNA). Both models produce well-calibrated predictions close to the $y{=}x$ line, with safe-MT showing a visibly tighter spread of residuals, consistent with the small but systematic improvements in VP MSE/MAE reported in Table \ref{['tab:main']}.