Max Zinke

These sit on the landing page of EvolutionaryScale, the lab behind ESM3. The brief was to show protein-language-model-generated structures as something investigable and beautiful, not as another glowing molecule on a dark gradient. See for yourself.

A 400K-parameter SchNet graph neural network trained on 4,834 BMRB-linked experimental structures. Predicts all six backbone nuclei (H, HA, N, CA, CB, C') from any PDB, mmCIF, or AlphaFold structure — with calibrated uncertainties. Beats SPARTA+, LEGOLAS, and UCBShift2 on a 712-protein homology-separated test set. Try it in your browser.

Given assigned backbone chemical shifts, infer local dihedral geometry and the secondary structure that follows. The probabilistic framing is the interesting part — the outputs are distributions, not point estimates.

An animation of the mechanism I solved during my postdoc at Institut Pasteur: the conformational switch of ExbD that drives active nutrient uptake in Gram-negative bacteria, and the catalytic role of the peptidoglycan layer. The Ton system harnesses the proton motive force to open outer-membrane transporters — our work showed that ExbD's periplasmic domain is a dynamic dimer switching between open and closed states, and that this switch triggers a disorder-to-order transition of TonB. First-author paper in Nature Communications.

A live, shipped product in a completely unrelated field: Counter-Strike 2 case-hardened skin patterns, reverse-engineered, rendered, and searched at scale. It is on this page for one reason — it is proof I build end-to-end products that real people use.