liveBook · Manning

Problem definition

As explained in Build AI Drug Discovery Pipelines

Problem definition is the specification of a predictive or analytical task, its success criteria, and whether a chosen method is appropriate.

Problem definition specifies the predictive task, success criteria, constraints, and whether machine learning is appropriate. It anchors every later stage, from data acquisition through evaluation and deployment. section 3.0

In ligand-based virtual screening, the task is to identify promising compounds that may block the human hERG potassium channel, an antitarget associated with drug-induced cardiotoxicity. The model predicts either hERG non-blocker (0) or hERG blocker (1); a non-blocker prediction does not establish safety. section 3.1

A useful definition makes the following decisions explicit:

  • State the input compounds, target label, intended use case, regulatory or timeline constraints, and business goal. section 3.1
  • Distinguish important errors: missing a blocker can create clinical cost and reputational damage, while a false alarm can remove a valuable candidate. section 3.1
  • Separate uncertainty from classification so low-confidence compounds can receive experimental follow-up instead of confident labels. section 3.1
  • Reassess the business objective after evaluation and reconsider project scope or alternatives if the model does not meet it. section 3.0 section 3.5
Figure 3.1 Key stages in our ML workflow.

Use the workflow overview to place problem definition before data acquisition, exploration, preprocessing, featurization, training, evaluation, deployment, and maintenance. section 3.0

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines requirements, labels, failure costs, and uncertainty, then stops short of production deployment and maintenance.
Examples
Uses Malaria Box similarity hits, hERG blocker labels, experimental follow-up, and two contrasting screening failure cases.
What it assumes
Assumes a candidate-compound set and a screening objective, while defining antitargets and hERG cardiotoxicity within the workflow.
What you can do after
Evaluate/understand whether a machine-learning screening task has defined labels, risks, constraints, and useful success criteria.
How it connects
Connects virtual screening with data acquisition, featurization, model evaluation, uncertainty, deployment, and business objectives.

Also explained in

Compare how 8 books explain problem definition →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.