Computational biology, and where it differs

A distinction worth keeping

One handles data,the other builds models

The two overlap heavily and the words are used interchangeably, but the difference is real: one processes measurements, the other predicts behaviour.

Bioinformatics

Data that exists: sequences, alignments, counts. The question is what the measurements say.

Computational biology

Models of systems: dynamics, networks, populations. The question is what the system would do.

In practice, both

A model needs data to fit against, and data needs a model to be interpreted. Most projects use both words.

Scales

From an atomto a population

Computational biology spans about twelve orders of magnitude, and the method changes completely at each step.

MoleculesMolecular dynamics and docking. Nanoseconds and nanometres, with a force field standing in for chemistry.

Pathways and networksOrdinary differential equations over reaction rates, or flux balance analysis across a whole metabolism.

Cells and tissuesAgent-based models where each cell follows rules, and the tissue behaviour is what emerges from them.

PopulationsEpidemiology and evolutionary dynamics. Compartment models, phylogenies, and selection over time.

Nothing bridges these scales directly. A molecular dynamics simulation cannot tell you about a tissue, and a population model knows nothing about a binding pocket. Multi-scale work means running separate models and passing parameters between them, which is as awkward as it sounds and where a great deal of the field's effort goes.

What changed

Structure predictionstopped being hard

For fifty years, predicting a protein's shape from its sequence was the field's defining unsolved problem. It is now largely a lookup.

AlphaFold changed the baseline

Accurate predicted structures for essentially every known protein, available to download.

Confidence is per residue

pLDDT tells you which parts to trust. Low-confidence regions are often genuinely disordered, not errors.

A structure is a starting point

Static and unbound. Questions about motion and binding still need simulation on top of it.

This did not end the field — it moved it. A predicted structure is a static snapshot with a confidence score, and most biological questions are about motion, binding and change over time. What the shift did is make molecular dynamics and docking cheaper to start, because you no longer wait on a crystal structure to have something to simulate.

Honestly

A model isa hypothesis

The failure mode in this field is treating simulation output as measurement, and it is easy to do because the output looks like data.

DataOutput looks likebut it is a prediction

Falsifiable?Always askwhat would the model get wrong

BothThe loopprediction, then measurement

A simulation produces numbers with as many decimal places as you like, regardless of whether the model behind them resembles the system. The useful discipline is to ask what the model would predict if it were wrong, and to check that against something measured. Prediction plus experiment is the loop; either one alone is not.

On a machine

SAMtools isalready built

Nothing here needs compiling. Pick a machine, and these two commands put it on there.

hub://samtools

$ dxflow workflow create --identity samtools hub://samtools

$ dxflow workflow start samtools

Start a machine What SAMtools needs

Read it, then run itPrepaid and by the hour, on a machine that is yours about a minute after you ask.

Start a machine All notes