Product Finder: expiring patents for a generics manufacturer

A generics manufacturer's business model rests on one question: which active substance becomes available when, and is it worth pursuing? The answer sits in regulatory databases and patent records – in different formats, at different levels of granularity, spread across several agencies.

Miss an expiring patent term and you lose a market window. Misjudge market attractiveness and you commit development capacity to a product that carries no margin against competitors. Monitoring both by hand is not feasible at this data volume.

European Medicines Agency database. Around 1,250 entries, spread across numerous subpages and not available as a download. We built a web crawler to collect the records and turn them into a structured dataset.

U.S. Food and Drug Administration database. Around 95,000 entries, available as CSV.

A keyword dataset from text mining. From the drug descriptions in both sources we extracted the central keywords through text mining – the basis for matching substances and indications across the two registers.

Data quality in the FDA source. The larger of the two datasets was the more problematic one. Data cleaning and wrangling were the precondition for the records being reliably usable in a dashboard application at all.

Differing granularity between the registers. EMA and FDA describe the same matter at different levels of detail. Using filters, sorting logic and mapping algorithms we consolidated both sources into a single dataset that could be analysed in Power BI. This is the core of the project: without that consolidation they remain two lists that cannot be read against each other.

Operationalising market attractiveness. Attractiveness is not a quantity present in the data. We modelled it through a deliberately simple, inspectable ratio: the fewer generics that exist in an indication, the more attractive the substance. The advantage of a transparent measure over a complex score is that specialist functions can check the assessment and correct it where they disagree.

The Product Finder surfaces expiring drug patents and identifies active substances and generics that are candidates for the development pipeline. Analysis runs through dynamic dashboards selected by use case:

• Dynamic filters produce a current list of patents about to expire
• Drilldowns show detail on an individual substance beneath the overview
• Manufacturer-level views open up all indications and products of a competitor – competitive monitoring as a by-product of the same data foundation

The architecture is deliberately extensible: further databases can be connected and additional dashboards added without disturbing the existing analysis.

The pattern – opening up scattered regulatory sources, consolidating registers of differing granularity, placing a transparent assessment measure on top, and making the result operable for non-technical users – carries over wherever decisions depend on knowledge that is publicly available but poorly accessible: portfolio analysis, competitive monitoring, evaluation of regulatory registers.

The task is the same; the tooling has improved.

For patent data we put structured interfaces such as EPO OPS ahead of crawling; crawling remains for sources like the EMA register, which offers no download. Classical text mining gives way to two-stage extraction: specialised models for the bulk, a language model with an enforced output schema for the hard cases.

Matching across sources no longer runs on mapping rules but on normalisation against INN, ATC and RxNorm – which eliminates the most common source of error, the same substance under three spellings. And instead of a dashboard alone, natural-language querying is possible, with a mandatory link back to the register entry.