Natalija Stepurko

Out-of-training properties in frozen interatomic-potential embeddings are decodable and ordered by physical scale

Universal interatomic potentials such as ORB-v3 and UMA-S are trained on DFT energies, forces and stresses. I probed its layer-wise embeddings across 154,875 Materials Project structures and nine properties to ask what else it had to learn to get the energies right. Five properties absent from the training objective, including band gap, bulk modulus and magnetic character, are decodable, and the information is organised by physical scale: composition-linked properties are readable at the input, while properties that depend on long-range atomic arrangement are assembled through message passing. The equivariant model's angular channels give it a clear edge on crystal-system classification.

Do a sequence model and a structure model learn the same biology?

ESM-2 reads only the amino-acid sequence; ProteinMPNN reads only backbone coordinates. If sequence and shape are two views of one biology, the two should converge. Across 4,898 proteins and 1,374,765 residues I compared their layer-wise representations with CKA, SVCCA and mutual k-NN, bounded by an untrained floor and a same-model-different-seed ceiling. The strongest raw "agreement" sits at the sequence model's embedding layer, which is simply the amino-acid identity both models are trained to output. Subtracting the per-residue-type mean removes that shared answer key and reveals what remains: the two families line up along some of the same directions but do not arrange proteins the same way, and neither model's representation can be mapped into the other's and still do its job: the two are not interchangeable in function.

Choosing a train/test regime when data is limited

With a few hundred lab measurements and a foundation model's embeddings to test or fine-tune on, how many should be trained on and how many held back? I measured this on frozen ORB-v3 and UMA-S embeddings over six Materials Project properties. Training size sets how far the score sits below the full-data value (shortfall); test size sets how far it moves when the split is redrawn (spread). Both are knowable here, which is unusual: the full-data score comes from the probing study above, and a 20,000-material reference held out from all training gives every estimate a true score to be checked against, so whether an error bar keeps its promise can simply be counted. The shortfall is not one number: at the same 5,000 training materials it is ten times larger for crystal system than for formation energy, and the gap between hard and easy properties widens as data is added. Combining the two into a total error puts the best split at 80/20 from 500 materials upward, drifting toward training data as the dataset grows. Five-fold cross-validation, scored against the same reference, lands closer to the true score than repeated holdout in most cells and keeps its interval coverage, but changes total error by about 2%, because it leaves training size unchanged.

Molecular insights into signalling mechanisms within the TGF-β superfamily

Experimental data-assisted protein modelling: I collected small-angle X-ray scattering data at synchrotron beamlines to characterise the solution structure, flexibility and oligomerisation of TGF-β superfamily proteins, and wrote custom Python pipelines for automated SAXS processing and downstream statistics.

• Ab initio shape reconstruction, rigid-body modelling, homology modelling and MD simulation, tied back to biophysical binding assays.
• Showed that follistatin's C-terminal tail folds back onto its FS1 domain and that all three isoforms alternate between two solution conformations; designed a bacterial assay for transmembrane-helix dimerisation of ALK receptors.
• Engineered and purified recombinant proteins from bacterial and mammalian systems; talks at the S4SAS Diamond Light Source conference and the Herchel Smith symposia.
• Supervised Cambridge undergraduates for four years (250+ hours) and designed a final-year research project that ranked top of its cohort.

Cyclin druggability assessed by fragment binding

Final-year project on structure-based drug discovery at the Northern Institute for Cancer Research: cell-cycle and transcriptional cyclins expressed, purified and screened for crystallisation; CycA2 crystals soaked with FragLite fragments and analysed by X-ray crystallography, alongside computational druggability assessment and docking. The 2.0 Å CycA2 structure showed a DMSO molecule bound in the pocket with the highest druggability score.

• Amgen Scholar at the University of Cambridge (2017) in Marko Hyvönen's lab on TGF-β structural biology, later the foundation of the PhD; NCL+ Advanced Award in Research.
• Earlier placements in structural biology of meiosis, bacterial carbohydrate metabolism and protein-purification scale-up.

Two years of medicine before moving to research

The preclinical half of a medical degree: how the body is built and how it works, from anatomy and physiology to genetics, microbiology, research methodology with biostatistics, and the first clinical training in examining patients.