Abstract
<jats:p>Electrochemical molecular property prediction is a low-data setting in which the downstream labels are scarce, the chemical space can be structurally heterogeneous, and the value of a source-domain representation is strongly conditional on the target task. This manuscript develops a systematic source-task view of supervised molecular pretraining and extends it into an adaptive source-routing framework. QM9 quantumchemical properties are used as supervised source tasks, including HOMO, LUMO, HOMO-LUMO gap, and u0; additional source-library experiments further include dipole moment and polarizability. These source-task-specific GINE encoders are transferred to two downstream electrochemical settings, OxPot oxidation-potential prediction and RedDB reaction-energy-related prediction, under random and scaffold splits and multiple low-data budgets. The first part of the study shows that source-task pretraining is useful but not universal. OxPot tends to benefit from frontier-orbital-related source tasks, whereas RedDB exhibits a more mixed source preference. Negative or negligible transfer occurs under some data sizes and split settings, indicating that fixed source selection is insufficient. The second part introduces AdaSGP-GINE, an adaptive source-gated pretraining framework that treats supervised source-task encoders as an expandable expert library. A sample-level gate assigns source weights, sourcespecific adapters refine transferred representations, and weighted fusion produces the downstream prediction representation. Source-weight analyses, RedDB source ablations, and 4-source versus 6-source expansion experiments support the interpretation that AdaSGP provides taskdependent routing and source-library diagnosis rather than a universal best model. A release package containing cached results, figure data, plotting scripts, audit summaries, and reproducibility documentation was also organized to support transparent downstream use.</jats:p>