Multitask Extension of Geometrically Aligned Transfer Encoder

TL;DR

The paper addresses data scarcity in molecular property prediction by transferring information across multiple tasks. It generalizes Geometrically Aligned Transfer Encoder (GATE) to a many-task setting by aligning latent-space geometries via mappings to a locally flat universal manifold derived from SMILES, enabling mutual information flow across tasks. The method introduces per-task regression units, autoencoder-based task-to-manifold mappings, and a composite loss $l_{tot}$ that combines $l_{reg}$, $l_{auto}$, $l_{cons}$, $l_{map}$, and $l_{dis}$ to enforce local and global geometric alignment. Empirically, it yields improved or competitive performance across 10 molecular-property datasets, with clear synergy in multi-task settings and robust behavior relative to standard multitask learning, albeit with increased computational complexity and opportunities for global-geometry-based enhancements.

Abstract

Molecular datasets often suffer from a lack of data. It is well-known that gathering data is difficult due to the complexity of experimentation or simulation involved. Here, we leverage mutual information across different tasks in molecular data to address this issue. We extend an algorithm that utilizes the geometric characteristics of the encoding space, known as the Geometrically Aligned Transfer Encoder (GATE), to a multi-task setup. Thus, we connect multiple molecular tasks by aligning the curved coordinates onto locally flat coordinates, ensuring the flow of information from source tasks to support performance on target data.

Figures (3)

• Figure 1: Four different coordinate frames are demonstrated in the figure, with coordinate transformation maps to each pair of tasks. One can interpret each coordinate frame as task-specific coordinates and map them with transformation models. An arbitrary point in the overlapping region of the manifold can be transformed from one task coordinate to another by combining mapping functions $\phi$. Moreover, by introducing perturbation points, as demonstrated in the figure, one can define the distance between points to match the geometrical shape in the overlapping region.
• Figure 2: Schematic diagram for the Extended GATE algorithm. The algorithm consists of a number of Regression Units. Each Regression Unit corresponds to an individual task. The universal manifold covers the entire coordinate space of RU's, and the transformation model maps a vector from each RU to a locally flat frame on the universal manifold. One can take the reverse path from the manifold to reconstruct the original vector. Furthermore, one can also transfer a vector to another RU coordinate by utilizing a different task's inverse transformation module.
• Figure 3: Regression performance of three-task GATE and two-task GATE in root mean square error (RMSE). For evaluating regression performance of two-task GATE, all three possible pairs of three tasks were trained separately and averaged.