Natalija Stepurko

Where Material-Property Information Resides in Frozen Interatomic Potentials

Machine-learning interatomic potentials are trained only on energies, forces and stresses. I probed the frozen, layer-wise embeddings of two architecturally opposite models, ORB-v3 and UMA-S, across 154,875 Materials Project structures and nine properties. Properties absent from the training objective are recoverable, with substantial improvements over composition-only descriptors for several targets, e.g. band-gap R² rises from 0.698 with Magpie to 0.890 with ORB-v3 and 0.907 with UMA-S. Local and additive properties, including density and bulk modulus, tend to be most decodable in early representations, whereas band gap and metal/insulator character benefit from deeper layers. The equivariant UMA-S angular channels give it a clear edge on crystal-system classification.

Where does BioEmu-1 fail against solution scattering, and what should be measured next?

I benchmarked raw BioEmu-1 conformational ensembles against 399 experimental SAXS profiles, comparing predicted and measured protein sizes, using maximum-entropy reweighting to characterise discrepancies, and testing six alternative protein models. Only 37% of raw ensembles met the SAXS fit threshold. For disordered proteins, predicted radii of gyration were 21% too large at the median, a discrepancy confirmed directly in the sampled conformer coordinates.

I then tested whether sequence and ensemble features could predict these errors before measurement. Under grouped cross-validation, the best model explained only 5% of their variation, and an experimental acquisition policy performed no clearly better than random selection. The resulting two-arm campaign prioritises controlled replication of known failures and an independent test of the disordered-protein extension bias.

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 the structure model’s representation does not survive the sequence model’s layers intact, most clearly when it enters early: the two are not interchangeable in function.

Do foundation-model representations uncover clinically useful biology that a linear model misses?

I defined cell states in a 50,002-cell breast-cancer atlas using scGPT, Geneformer and an HVG-PCA baseline, turned each state into a marker-gene signature, and asked whether those signatures predict survival in 1,094 TCGA breast tumours. Every representation went through the same clustering, marker selection and bulk-scoring pipeline; winner selection was tested against a family-wise permutation null and random gene sets matched for size and expression.

Neither foundation model cleared the pre-specified family-wise null, neither gained more prognostic signal above its random-gene floor than HVG-PCA, and adding the selected states to age and stage changed the cross-validated C-index by essentially zero. With the signatures fixed before METABRIC outcomes were examined, all eight foundation-model-minus-baseline margins were negative in 1,979 METABRIC patients.

When is a probe score trustworthy under limited data?

With limited data, a probe score can mislead in two separate ways. Too little training data leaves it below what the same probe reaches with more data (the shortfall); too little test data makes it move when the split is redrawn (the spread). I measured the two separately on frozen ORB-v3 and UMA-S embeddings over nine Materials Project properties, with composition-only Magpie features run on the same draws as a baseline. Every probe was scored on one frozen 20,000-material test set and compared with the full-data score from the probing study above.

Stability and sufficiency are different: band gap gives a reproducible score from 500 training materials but keeps rising until 7,000, where it comes within 0.05 of the full-data score. From 500 materials up the shortfall costs more than the spread, so the best split sits on the training side, and the usual 80/20 is within 0.004 of the best. Five-fold cross-validation gives the most accurate estimate of the score, but it cannot remove the shortfall, because each of its models still trains on 80% of the data.

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.
• Designed a bacterial assay for transmembrane-helix dimerisation of ALK receptors.
• 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) 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.

BravelyCultured

A VC-backed microbial biotechnology company harnessing novel marine microorganisms to produce proteins and biosurfactants that replace petroleum- and animal-derived ingredients. We raised dilutive and non-dilutive funding, established a laboratory facility and a team, coordinated microbial sample collection across multiple geographies (UK, Korea and French Polynesia), collaborated with companies and organisations on joint research projects, and licensed our microbial strains.