ConTP uses contrastive learning to predict transporter substrate specificity
Researchers from King Abdullah University of Science and Technology (KAUST) and Ningxia Medical University introduced ConTP, a new framework for annotating membrane transporters. Traditional methods rely on evolutionary similarity to guess which molecules a transporter moves, but this approach often fails because functional specificity does not always align with phylogenetic distance. ConTP…
Key points
- ConTP realigns protein embeddings around substrate semantics instead of sequence similarity.
- Benchmarks span 70 substrate types and 1,352 Transporter Classification families.
- The model recovers cross-family convergence and multi-substrate specificity in NRAMPs.
The team built benchmarks covering 70 fine-grained substrate types and 1,352 Transporter Classification (TC) families to test the model. In these tests, ConTP successfully identified cross-family convergence, such as sodium transport across different superfamilies, and accurately captured the multi-substrate specificity of NRAMP transporters. The study also demonstrated that projecting generated sequences through the model can reveal substrate-fidelity errors in existing protein design tools.
The work was supported by KAUST’s Office of Research Administration and other grants. The authors argue that this geometry-aware approach offers a more accurate view of transporter function than homology-centric methods, addressing systematic blind spots in current substrate-level inference.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Nvidia pledges $250M in compute for UK AI agent safety research · 8 src
- OpenAI offers U.S. university labs free ChatGPT Business seats with Rosalind access · 2 src
- Anthropic reports small AI models may self-replicate on IoT devices within months · 29 src
- Researchers propose LLM-referenced C-test for evaluating supra-intelligence · 8 src
- Researchers study permutation diffusion LLMs for recommendation reranking · 1 src
Comments
via GitHub Discussions