The Alpha Fold – MMM Parallel
I finally managed to catch up on the famous documentary “The Thinking Game”. What a story !!
I noticed many parallels between protein folding and Marketing Mix Modeling (MMM).
📌The Hidden Structure
AlphaFold
In AlphaFold, we observe an amino acid sequence and want the hidden 3D protein structure.
MMM
We observe marketing/media spends and the corresponding KPI (Sales, net user addition, Awareness). We want to know the causal structure – which variables resulted in the KPI, by how much, and the interaction effects between them (if any).
This causal discovery is not as difficult as protein structure but it is not easy either.
📌 Direct Measurement is Difficult
AlphaFold
Protein structures can be measured using X-ray crystallography, but I am given to understand that it is slow and expensive.
MMM
In MMM too, the real causal structure can’t be unraveled unless one does controlled experiments. But experiments in marketing are rarely truly controlled. One can’t really run clean RCTs in marketing (check link in comments to know why).
Even if attempted, experiments rarely answer what MMM asks “How all variables together affected the KPI”
Hence the indirect approach: MMM modeling followed by confirmatory causal tests like DiD.
📌 The Feature Selection Process
We choose variables that best ‘explain’ or ‘predict’ the outcome. Before modeling, we don’t know which variables matter so we look for clues.
In AlphaFold, this clue comes from evolutionary correlations.
If two amino acids mutate together across species, they are likely spatially close. This dramatically reduces the search space.
In MMM
Many vendors use correlations to identify variables. At Aryma Labs we instead rely on Information Theory. The question being – “How much information each variable carries about changes in the KPI”. This also helps mitigate multicollinearity.
📌 The Identifiability Issue
Both MMM and protein folding face identifiability challenges.
In MMM (from a Frequentist lens), sales are realized and fixed. Multiple combinations of variables could theoretically produce the same outcome. The task is to identify the correct combination.
In AlphaFold too, many structures could fit the same sequence. The challenge is narrowing the possibilities.
📌 The Differences
Prediction vs Causal Inference
AlphaFold succeeded because it had massive datasets and strong constraints (evolutionary correlations and physical/chemical structure constraints).
The task was to ‘predict’ the structure, not to explain ‘why’ the protein folds a certain way.
In MMM, the task is causal. We care more about which variables actually caused sales, not just any random list of variables that predicts them.
📌Can there be a MMMFold?
There are many inspirations one could take from AlphaFold for marketing measurement. Our AI team is exploring some ideas but can’t reveal much yet 😎.
The real question though is: would that be overkill 🤔?
