I wanted to think about potential metrics for modelling POS - the following is a living document, open iterative work on this subject. Related posts will update on findings, failures and further ideas to be tested!
Predicting probability of success (POS) in biotech should start with biology. Evaluating the mechanism of action (MOA) of therapeutic targets, differentiating between upstream and downstream interventions, and, how the position of a target within a biological pathway influences POS. Most models overweight market size, management quality, or phase transition statistics and underweight whether the mechanism actually has durable control over the disease system. The key question is where the target sits within the network.
An upstream target may appear attractive because it regulates an entire pathway, but these targets often fail; redundant signalling pathways, feedback loops, and compensatory activation allow the phenotype to survive inhibition. KRAS in pancreatic cancer is a good example. KRAS is clearly oncogenic, yet decades of direct targeting failed because tumour signalling adapts downstream through MAPK, PI3K, and parallel survival pathways. Importance does not equal vulnerability.
Downstream targets can sometimes have higher POS because they sit closer to phenotype and clinical effect. PD-1/PD-L1 inhibitors worked partly because tumour immune dependency was measurable in humans through biomarkers, pharmacodynamic response, and observable clinical activity early in trials.
Metrics for Modeling POS
A model integrating quantitative data (biomarker validation, trial phase success rates, pathway position) with qualitative factors (regulatory sentiment, market dynamics).
Key metrics include:
MOA positioning within pathways.
Translational fidelity of preclinical models.
Regulatory success rates stratified by target class.
Combining Quantitative and Sentimental Analysis
Machine learning algorithms can incorporate structured data from past trials and unstructured sentiment analysis of regulatory trends. (sidenote: this makes biotech investing a lot more interesting and hopefully smarter too. I’m going to build a prototype and share a snippet of the results.) This hybrid approach allows real-time updates to POS predictions as new data emerges. For example, natural language processing (NLP) could analyse FDA communications to gauge regulatory sentiment for novel MOAs.
To build this, I’ll use a combination of datasets from past clinical trials: success rates, phase transitions, and MOA classifications. On the unstructured side, NLP can extract sentiment and trends from FDA communications, regulatory announcements, and scientific publications. The model will integrate these inputs to provide real-time updates on POS predictions. For instance, we could measure how changes in sentiment around novel MOAs correlate with trial outcomes, refining the model’s accuracy as more data flows in.
