DecoderDTA: enhancing drug-target binding affinity prediction via decoding fidelity in multi-task learning.

Accurate prediction of drug–target binding affinity (DTA) is essential for in silico drug discovery. Existing deep learning approaches typically optimize solely for affinity regression, which can lead to overly compressed representations that lose partial structural and semantic information. We introduce DecoderDTA, a multi-task framework that improves established DTA backbones by incorporating an auxiliary sequence reconstruction task to enforce decoding fidelity. During training, these decoders are tasked with reconstructing the raw SMILES strings and amino-acid sequences. This auxiliary reconstruction objective acts as a decoding fidelity, which encourages the encoders to retain richer structural and semantic information without introducing additional computational overhead during inference. On the Davis and KIBA benchmarks, DecoderDTA variants consistently improve performance. For instance, DeepDTA-Dec reduces the MSE from 0.261 to 0.220 on Davis (a ≈15.7% relative improvement) and from 0.194 to 0.167 on KIBA (≈13.9%), while also increasing the CI and [Formula: see text] metrics. Similarly, GraphDTA-Dec shows consistent improvements. Ablation studies reveal an interesting phenomenon: jointly optimizing both reconstruction tasks does not consistently outperform using a single decoder. This finding highlights the complex interplay of objectives within the multi-task framework and points to an important direction for future architectural refinements, such as adaptive task weighting. Nevertheless, the consistent improvements achieved by all DecoderDTA variants over their respective baselines demonstrate that incorporating decoding fidelity is an effective and generalizable strategy for enhancing DTA prediction, offering a new perspective for representation learning in this task.