Function (biology)×Bioinformatics

Bioinformatics has learned to predict structure and variant effects (AlphaFold, VEP, ColabFold), but the harder frontier is predicting function — what a protein, cell, or organoid actually does. Fusing functional biology's experimental readouts (CRISPR perturbation, organoids, stem-cell assays) with bioinformatics' predictive engines creates a closed loop where computation proposes function and wet-lab validates it, collapsing the annotation gap for the human genome's 'dark' proteome.

The two sides share unusually dense bridges — CRISPR, organoids, embryonic stem cells, and extracellular vesicles all appear on both sides, and 43 authors already publish separately in each field with 36 common neighbours and a high Adamic-Adar affinity (8.6). Bioinformatics has strong recent momentum (recent-share 0.129 vs Function's 0.03), meaning the computational side is accelerating toward a mature but slower functional-biology field ripe for automation. The signature papers reveal the pull: structure prediction and variant-effect scoring are solved enough that 'what does it do' is the next natural question.

Groups that own both a high-throughput functional-perturbation platform (pooled CRISPR screens, organoid phenotyping) and a modern ML-genomics stack will win — because the moat is proprietary function-labeled training data, not the model architecture. Expect the winners to be integrated 'lab-in-the-loop' teams rather than pure dry-lab or pure wet-lab shops.

A closed-loop 'function foundation model' benchmark: pool CRISPR-perturbation phenotypes across matched organoid/stem-cell lines, then train and prospectively test whether sequence+structure models can predict the measured functional phenotype of held-out genes — with the model's uncertainty actively selecting the next wet-lab screen.