Genomic sequence analysis, protein folding visualisation and simulation frameworks, held so a result can be opened again. Same inputs. Same parameters. Same environment.
State of the field
Some problems moved from research to routine in a few years. The edges are where results get over-claimed, so that is where the checks go.
Profile-HMM and fast k-mer search tools scale homology search to very large databases. Clustering by identity is a standard first step.
Remote homology below detectable similarity. Protein language model embeddings are being used for search and compare well in published studies, but they inherit what the model has seen.
Database release, aligner version and sensitivity settings recorded with every search.
Single-domain folds. The CASP16 monomer assessment described single-domain fold prediction as nearly solved.
Ranking complexes, stoichiometry, change on binding. Diffusion-based predictors can draw spurious order in disordered regions. A prediction is one static state, and more seeds do not recover an ensemble.
Confidence kept beside the coordinates, with disordered regions marked as such.
Generate a backbone, assign a sequence, filter in silico on self-consistency and predicted error.
Reported experimental success rates sit near one in five on curated targets and are far lower in some independent applications. The filter is not the assay.
Filter thresholds and every rejected design logged, so a hit rate has a denominator.
Classical force-field dynamics with enhanced sampling. Machine-learned potentials are narrowing the gap to quantum accuracy.
Femtosecond timesteps against events that take microseconds to days. Sampling, not compute, is usually what limits a conclusion.
Topology, timestep, thermostat and seed travel with the trajectory. A run that cannot restart is incomplete.
Quality control, normalisation, variable-gene selection, embedding and clustering follow community best-practice guides.
Separating biology from technical variation. Integration methods can over-correct or add artefacts, and benchmarks do not always agree.
Batch structure declared up front. Integration tuned on one split and scored on another.
Held-out test sets and cross-validation.
Random splits leak through homology. Pretrained models can leak through their corpus: one study scored higher on a split that ignored pretraining clusters than on a harder pretraining-aware one.
Cluster-aware, pretraining-aware splits, with the split itself part of the record.
Sequence analysis
Homology search lines related sequences up column by column. What stays fixed across the family is what the family cannot afford to change, and that is the signal structure and function predictions lean on. Conservation, information content and occupancy below are computed from the rows on screen.
A constructed family of 14 sequences derived from the chain in the viewer above with a simple substitution model, in which buried positions change least. It stands in for the real alignment a search would return, and the tracks are computed the same way.
Hover a column to read its composition.
14 sequences, 89 columns, 74 strongly conserved columns.
Structure analysis
Every pair of residues has a distance. Laid out as a matrix, helices show as bands along the diagonal, a hairpin of two strands as a line crossing it, and tertiary contacts as isolated off-diagonal spots. Predicted structures are routinely scored by comparing exactly this kind of matrix.
This one is computed from the packed coordinates of the generic chain in the viewer. Switch to the change view to see which pairs moved when the chain packed.
Hover the matrix to read a pair.
Simulation and sampling
A simulation takes femtosecond steps, and the events that matter take microseconds to days. A walker on a funnelled surface with decoy basins shows why that gap decides what you can conclude. At low temperature it settles in the first well it finds and stays there.
Raise the temperature, or switch on a history-dependent bias that fills visited wells as it goes, the idea behind metadynamics, and it climbs out. This is a Metropolis walker on a constructed two-dimensional surface in reduced units, computed live.
Epigenetic age and reprogramming
DNA sequence does not age in this sense. What is measured, and what is claimed to be reversed, is the methylation pattern at specific CpG sites, read through a statistical model. That makes the question computational: how a clock is built, what it responds to besides age, and what evidence would separate a real change in biology from a shifted proxy.
A research-methodology demonstration. The cohort below is constructed: 360 synthetic samples and 300 synthetic CpGs generated in your browser from a seeded process, and the clock is fitted in your browser. Nothing here is patient data, a result, a treatment or a product, and Forticia offers none.
95% interval −4.4 to −4.1. Chronological age is unchanged.
Mean absolute error on 108 held-out samples, r = 0.982.
53% of the weight sits on cell-type-sensitive CpGs.
No age-associated methylation was changed, yet the clock moved by −4.2 years. Everything in that movement comes from the confounders.
| Source of change | Applied | Mean change, years | 95% interval | Spread between samples |
|---|---|---|---|---|
| Real methylation change | 0 years-equivalent | 0.0 | 0.0 to 0.0 | 0.0 |
| Cell-composition shift | 0.50 | −4.2 | −4.3 to −4.1 | 0.6 |
| Batch offset | 0.00 | 0.0 | 0.0 to 0.0 | 0.0 |
| Technical noise | 0.30 | 0.0 | −0.1 to +0.1 | 0.5 |
| All together | −4.2 | −4.4 to −4.1 | 0.8 |
A real applied shift reads lower than its true size. Regularised regression shrinks coefficients toward zero, so a fitted clock attenuates true effects. This is a known property of such models, and one more reason a clock reading is not a measurement of biological change.
A penalised regression, usually an elastic net, over methylation levels at a few hundred CpG sites, trained to predict an age-related target. Horvath’s 2013 multi-tissue clock uses 353 CpGs and Hannum’s 2013 whole-blood clock uses 71. Second-generation clocks change the target: PhenoAge (513 CpGs) is trained on a phenotypic age measure, GrimAge (1,030 CpGs) on methylation surrogates for smoking and plasma proteins that predict mortality, and DunedinPACE (173 CpGs) on the pace of ageing, with CpGs chosen for test-retest reliability.
Brief, repeated expression of the Yamanaka factors short of full reprogramming. In 2016, cycles of two days on and five days off extended lifespan in a progeria mouse model; results in normally ageing mice have been more modest. The factors are pro-oncogenic, full reprogramming produced teratomas, and every programme described in the review literature is preclinical.
Clocks lose accuracy at older ages and differ between tissues. Technical replicates of the same sample can differ by a median of about one to two and a half years across six widely used clocks, and by 4.5 to 8.6 years at the extreme. Cell-type composition and batch move readings. A falling clock reading and a functional improvement are different claims, and the literature reports cases where they did not track each other.
| Claim to support | What would count | Failure it rules out |
|---|---|---|
| The reading moved | The shift is larger than replicate noise measured on the same platform, and is reported with an interval. | Noise read as effect |
| It is not the layout | Treated and control samples are randomised across arrays and batches, and batch is in the model. | Batch offsets |
| It is not composition | Cell-type fractions are measured or populations are sorted, and the shift survives adjustment. | Composition shifts |
| It is not the clock | A clock fixed in advance, plus others trained differently, point the same way. The favourable one is not picked afterwards. | Clock shopping |
| The biology changed | Orthogonal readouts move with it, in function, expression and other marks, and persist after the intervention stops. | A shifted proxy only |
| It repeats | An independent cohort or laboratory reproduces it from the recorded pipeline. | One-off or cohort-specific results |
Pipelines and provenance
Every task is identified by its tool version, parameters, environment and the digests of its inputs. Change one thing and only what depends on it is invalidated; the rest is replayed from the store. Pin the environment and the seed, and a repeat reproduces the bytes. Float either, and watch what survives.
A mechanics demonstration: the digests are real SHA-256 values computed in your browser from the parameters and upstream digests shown, and the statistic each task emits is constructed, not a scientific result.
| Run | Specification | Executed | Cached | Conclusion | Result digest | Time | Compare |
|---|
Choose a run as A and another as B in the ledger, or use Run twice without cache.
What we mean by reproducible
Each question holds some parts of an experiment fixed and varies one. The dimensions follow the PRIMAD model from the reproducibility literature: platform, research objective, implementation, method, actor and data. Choose a question to see what has to be in the record for it to be answerable.
| Question | What varies | What it tests |
|---|---|---|
| Nothing | Does the record alone reproduce the result? | |
| Actor | Is the record complete without a verbal hand-over? | |
| Platform | Does the result depend on the hardware it happened to run on? | |
| Implementation | Is the finding a property of the method, or of one piece of code? | |
| Data | Does it generalise, or was it only describing what it had seen? |
| Reduction order | Result | Relative difference | Verdict |
|---|---|---|---|
| Sequential, reference | -13169.016263874486 | 0 | Reference |
| Sequential, run again | -13169.016263874486 | 0 | Within tolerance |
| Reversed order | -13169.016263874846 | 2.73e-14 | Within tolerance |
| Pairwise tree | -13169.016263874852 | 2.78e-14 | Within tolerance |
| 8 parallel lanes | -13169.016263875226 | 5.62e-14 | Within tolerance |
| 64 parallel lanes | -13169.016263874973 | 3.70e-14 | Within tolerance |
Where a tool is not deterministic, the tolerance is declared before the run. A tolerance chosen after seeing the difference is not a tolerance.
Evaluation
Biological sequences are related by descent, so a random split puts near-copies of training examples in the test set and rewards recall. Clustering by similarity before splitting is the standard remedy. Models pretrained on large sequence corpora add a second route: a test family can be absent from your training set and still be inside the pretraining data.
Each point in the plot is one sequence in an abstract similarity map, grouped into families. Change the split in the first cell and the two measured quantities in the second move with it. Both are computed from the points shown.
data, corpus = load_similarity_map(families=8)
train, test = random_split(data, test_fraction=0.22)spacing = median(nearest(train, train))
ratio = median(nearest(test, train)) / spacing
exposed = share(nearest(test, corpus) < 1.6 * spacing)ratio is how far test points sit from training data, relative to how far training points sit from each other. Near 1 means the test set looks like the training set.
exposed is the share of test points with a pretraining-corpus neighbour inside typical spacing.
Every family appears on both sides. Most test points have a close relative in training.
Computational Biology
Further reading
Internal researchers, contractors and academic collaborators can be provisioned without inheriting the rest of Forticia. Writes that change the record wait for a lead.