Protein–ligand binding affinity prediction plays a pivotal role in the field of drug discovery, with multimodalbased methods standing out. Existing approaches based on a multimodal framework typically rely on simple concatenation or pooling, which fail to identify and emphasize key features across heterogeneous sources. To tackle this bottleneck, we propose an adaptive fusion module, which injects sequence–structure embeddings into a functional annotation stream while incorporating functional annotation embeddings into the sequence–structure stream. Subsequently, learnable adaptive weights are employed to combine the outputs of these two pathways in a fully data-driven manner. Extensive experiments on the PDBBind benchmark demonstrate that our method achieves the best performance compared with state-of-the-art methods. Ablation study and hyperparameter analysis confirm that sequence, structure, and functional annotation each provide complementary information, and the joint optimization of these modalities via our adaptive fusion strategy yields the highest overall predictive accuracy. Our code is available at GitHub link https://github.com/Jessez2/AdaMM.

Current multimodal learning strategies primarily optimize in the original token space. Such a framework is easy to incorporate with the backbone of pretrained language model, but might result in modality collapse. To alleviate such issues, we leverage the Joint-Embedding Predictive Architecture (JEPA) on the multimodal tasks, which converts the input embedding into the output embedding space by a predictor and then conducts the cross-modal alignment on the latent space. We implement this predictor by a Multi-Gate Mixture of Experts (MMoE) and name the framework as M3-JEPA, accordingly. The gating function disentangles the modality-specific and shared information and derives information-theoretic optimality. The framework is implemented with both contrastive and regularization loss, and solved by alternative gradient descent (AGD) between different multimodal tasks. By thoroughly designed experiments, we show that M3-JEPA can obtain state-of-the-art performance on different modalities and tasks, generalize to unseen datasets and domains, and is computationally efficient in both training and inference. Our observation suggests that M3-JEPA might become a new basis to self-supervised learning in the open world.

Self-supervised pre-trained language models have recently risen as a powerful approach in learning protein representations, showing exceptional effectiveness in various biological tasks, such as drug discovery. Amidst the evolving trend in protein language model development, there is an observable shift towards employing large-scale multimodal and multitask models. However, the predominant reliance on empirical assessments using specific benchmark datasets for evaluating these models raises concerns about the comprehensiveness and efficiency of current evaluation methods. Addressing this gap, our study introduces a novel quantitative approach for estimating the performance of transferring multi-task pre-trained protein representations to downstream tasks. This transferability-based method is designed to quantify the similarities in latent space distributions between pre-trained features and those fine-tuned for downstream tasks. It encompasses a broad spectrum, covering multiple domains and a variety of heterogeneous tasks. To validate this method, we constructed a diverse set of protein-specific pre-training tasks. The resulting protein representations were then evaluated across several downstream biological tasks. Our experimental results demonstrate a robust correlation between the transferability scores obtained using our method and the actual transfer performance observed. This significant correlation highlights the potential of our method as a more comprehensive and efficient tool for evaluating protein representation learning.