CheMLT-F: multitask learning in biochemistry through transformer fusion.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning in drug discovery

Background:

• Drug discovery is slow and expensive due to extensive screening requirements.
• Current single-task predictors create fragmented workflows and hinder knowledge reuse.
• A unified approach is needed to integrate diverse prediction tasks.

Purpose of the Study:

• To develop a compact multitask transformer (CheMLT-F) for unified molecular and protein sequence modeling.
• To cover a broad range of over 680 endpoints, including toxicity, physicochemical properties, and drug-target interactions.
• To streamline in silico screening and facilitate data-driven decision-making in early drug discovery.

Main Methods:

• Developed CheMLT-F, a multitask transformer fusing molecular and protein sequence encoders.
• Employed a tailored training strategy with partial encoder freezing, global-local loss balancing, and weighted task sampling.
• Evaluated performance across 13 public benchmarks and an external protein-family benchmark.

Main Results:

• CheMLT-F matches state-of-the-art toxicity prediction and sets new benchmarks for physicochemical property prediction.
• Achieved competitive performance in drug-target affinity prediction (KIBA and Davis).
• Demonstrated broad generalizability in bioactivity prediction across protein superfamilies.

Conclusions:

• CheMLT-F offers a lightweight, inference-efficient, and extensible solution for computational drug discovery.
• Multitask parameter sharing reduces complexity and operational overhead compared to single-task models.
• Establishes a scalable foundation for holistic predictive modeling, lowering barriers in early-stage drug development.