01 Motivation
Furfural is derived from agricultural residues rather than petrochemical feedstock, which makes it an attractive starting point for building heterocyclic scaffolds. Heterocycles are everywhere in medicinal chemistry: they provide the rigid cores and hydrogen-bonding patterns that let a small molecule present a defined shape to a protein. The project asked whether a biobased route could deliver a scaffold of genuine medicinal interest.
02 Reaction design and catalysis
The synthetic work compared heterogeneous catalysts — Amberlyst-26, γ-alumina and calcium oxide — for the key transformation. Heterogeneous catalysts are chosen partly for practical reasons: they can be recovered and reused, and they avoid contaminating the product with residual metal. Comparing an ion-exchange resin against two metal oxides meant comparing quite different surface chemistry, and so quite different behaviour in the same reaction.
Why this matters computationally. Catalyst comparison is an experiment with a small number of expensive runs, which is the opposite of the data regime I now work in. Having done both, I have a reasonably concrete sense of what experimental data actually costs to produce — which changes how you treat a public dataset of ten thousand samples.
03 Analytical characterisation
Thermogravimetric analysis, transmission electron microscopy and gas chromatography were used to characterise the catalysts and the reaction outputs. Each answers a different question: TGA on thermal stability and loss of volatiles, TEM on morphology and particle structure, GC on what the reaction mixture actually contains.
The habit this builds is corroboration. No single technique establishes a result on its own, and when two disagree the disagreement is the finding. That transfers directly: a differential expression result and a pathway enrichment result that point in different directions are telling you something, and the temptation to report only the one that fits the story is exactly the thing to resist.
04 Molecular docking
The computational component investigated how the synthesised scaffold might bind a protein target, examining poses and the interactions available to the ligand within the binding site.
The limitation I would state first. Docking scores rank poses within a scoring function. They are not binding affinities, they are not validated predictions of activity, and the correlation between docking score and measured affinity is frequently poor. The honest claim is that docking generated a structural hypothesis about how the scaffold could sit in the site. Confirming it would require experimental binding data.
05 What it taught me
Three things, all of which show up in the computational work on this site.
- Molecules are physical objects. When a model outputs a logP or a descriptor vector, I have some intuition for whether the number is plausible, because I have handled compounds whose behaviour those numbers are trying to describe.
- Characterisation is adversarial. You run several techniques because you are trying to catch yourself being wrong, not to accumulate supporting evidence.
- Computational output needs a stated epistemic status. A docking pose is a hypothesis. A SIFT prediction is a hypothesis. Treating either as a finding is the most common way computational work loses credibility with experimentalists.