Forecasting Pharmaceuticals with Limited Raw Materials: An AI/ML Perspective
Demand forecasting is a critical process in the pharmaceutical industry to predict future sales, optimize inventory, and inform strategic decision making. However, forecasting becomes more challenging when dealing with drugs that rely on raw materials with a limited or constrained supply. In this article, we‘ll dive deep into forecasting methodologies for such products, with a focus on plasma-derived therapies (PDTs) and other innovative medicines. We‘ll also explore how artificial intelligence (AI) and machine learning (ML) techniques are transforming the field of pharmaceutical forecasting.
The Plasma-Derived Therapy Market
Plasma-derived therapies are essential medicines used to treat a variety of rare and chronic diseases. The active ingredients in these drugs are proteins separated and purified from human blood plasma donations. Some examples of PDTs include:
- Immunoglobulins (IGs) for treating immunodeficiencies and autoimmune disorders
- Clotting factors for managing hemophilia and other bleeding disorders
- Alpha-1 antitrypsin for alpha-1 antitrypsin deficiency
- Albumin for hypovolemia or hypoalbuminemia
While the patient need for PDTs continues to grow each year, the supply of the core raw material, blood plasma, is limited. Plasma can only be obtained from healthy human donors in a highly regulated process. Supply constraints can occur due to limitations in collection capacity, decreased donation frequency, or unexpected disruptions, as we saw with COVID-19 pandemic impacts on plasma collection volumes in 2020-2021.
These supply constraints make forecasting especially important yet difficult for plasma-derived therapies. Manufacturers must carefully balance patient demand with the realities of a limited raw material supply. Let‘s examine a best-practice approach to developing robust PDT forecasts in this challenging environment.
A Step-by-Step PDT Forecasting Methodology
1. Epidemiology Analysis
The foundation of an accurate forecast is a detailed understanding of the epidemiology – the incidence and prevalence – of conditions treated by PDTs in each market. Analysts gather population data combined with published diagnosis and treatment rates to calculate the total number of treated patients:
Total PDT Treated Patients = Population × Prevalence Rate × Diagnosis Rate × Treatment Rate
For example, consider hemophilia A in the US market:
| Parameter | Value |
|---|---|
| Total male population | 165M |
| Hemophilia A incidence | 1:5,000 |
| Diagnosed population | 90% |
| Treatment rate | 80% |
| Treated patient calculation | 165M × (1/5000) × 90% × 80% = 23,760 |
Advanced AI/ML techniques such as Bayesian hierarchical models can be applied to epidemiological data to derive more robust patient estimates by leveraging real-world data across multiple products, indications, and geographies[^1]. These models can account for uncertainties in epidemiology parameters and provide probabilistic forecasts to guide decision making.
2. Patient to Volume Conversion
The next step converts patient numbers into product volume requirements using an average grams per patient utilization factor. PDTs doses are prescribed based on patient weight, severity of deficiency, and frequency of treatments. Gathering real-world data on consumption patterns is critical to determine accurate utilization assumptions.
Continuing our hemophilia A example, assume the following parameters:
| Parameter | Value |
|---|---|
| Average utilization | 120,000 IU/patient/year |
| Specific activity | 200 IU/mg |
| Production yield | 90% |
| Volume calculation | 23,760 patients × 120,000 IU/patient ÷ 200 IU/mg ÷ 90% = 15,840 grams |
Machine learning algorithms like gradient boosting can analyze historical treatment patterns across patient segments to identify key drivers of drug utilization[^2]. These insights help build predictive models to forecast future demand more accurately than relying on broad population averages.
3. Comparison to Allocated or Available Supply
At this stage, the forecasted volume of raw material required based on patient demand is compared to the actual allocated or available supply. Manufacturers track current and projected plasma collections and determine what volumes will be allocated to each PDT based on strategic priorities and production constraints.
Imagine the forecasted patient demand for Factor VIII requires 15,840 grams, but only 14,000 grams of production were allocated due to plasma supply constraints. The volume forecast would be capped at 14,000 grams unless the supply situation changes. On the other hand, if 18,000 grams were allocated, the forecast would remain at the demand-based level of 15,840 grams.
This illustrates how PDT forecasts are ultimately constrained by the plasma supply, not just patient demand like in many other pharmaceutical areas. Careful collaboration between marketing, plasma collection, and manufacturing teams is essential to manage these dynamics.
Advanced planning systems powered by AI and optimization algorithms help manufacturers model complex supply chains and make data-driven allocation decisions[^3]. These tools can analyze trade-offs between different products, forecast supply and demand imbalances, and recommend optimal production schedules to maximize patient access and minimize waste.
4. Volume to Sales Value Conversion
The forecast is then converted from physical volume to monetary value using relevant drug prices or expected sales prices. PDTs are often expensive specialty drugs, so accurate price assumptions are key to the overall forecast.
Gross Sales = Allocated Volume (if < Required Volume) or Required Volume (if ≤ Allocated Volume) × Drug Price
Prices may need to account for different markets, such as distinguishing tender markets with lower negotiated prices from open prescription markets. Price changes over time due to competitor actions, contracting decisions, and market access should also be considered.
ML-powered pricing and contracting tools can analyze massive amounts of historical sales data, claims records, and market research to identify revenue maximizing price points and discount structures for new products[^4]. These data-driven insights help develop robust forecast assumptions to drive profitable growth.
5. Market Share Projections
For products facing competition, future market shares must be estimated. A common approach is to have business experts provide expected share targets at specific future time points, then use linear interpolation to project shares for interim periods.
For example, say a PDT had a 40% market share in 2023 and experts expect increasing to 50% by 2027 due to a planned new indication. The model would interpolate shares of 42.5% in 2024, 45% in 2025, and 47.5% in 2026, unless adjusted by the experts. Combining these market share projections with the overall market volume forecast yields the final brand-level forecast.
AI-powered market mix and attribution models are increasingly used to analyze the competitive landscape and forecast market shares[^5]. By ingesting data on promotional activities, scientific publication trends, key opinion leader sentiment, and more, these models quantify the impact of specific events on brand performance to guide more dynamic share projections.
Optimizing Plasma Fraction Yield and Pooling Strategies
Given the inherent constraints in plasma raw material supply, it‘s critical PDT manufacturers optimize fraction yield and pooling strategies. Plasma pooling involves combining donations for processing to average out variability and meet target specifications. Strategic pooling and AI-powered process parameter optimization can improve the yield, purity, and consistency of the separated therapeutic proteins[^6].
For example, Grifols has implemented AI and ML techniques to analyze donor data and plasma quality parameters to dynamically adapt pooling and fractionation processes[^7]. These adaptive manufacturing systems allow more flexibility to maximize protein yields and respond to changing supply and demand conditions by adjusting product mix in real-time. Incorporating these expected yield and pooling efficiency gains into forecasting models is key for capacity planning.
Forecasting Challenges for Cell and Gene Therapies
Innovative cell and gene therapies face similar raw material supply constraints as PDTs. Autologous cell therapies require collecting, modifying, and re-infusing each individual patient‘s cells, so the patient is both the donor and recipient. Production capacity is constrained by complex manufacturing processes, clean room space, and skilled labor. Similarly, many gene therapies rely on viral vectors as a key raw material, which have been a production bottleneck as more programs advance to late-stage trials and commercial scale[^8].
Forecasting demand for these novel one-time curative therapies is especially challenging given small patient populations, uncertain market adoption, and high per-patient costs. ML techniques like simulated patient journeys and discrete event simulation can model the end-to-end cell therapy treatment process to predict capacity utilization and optimization opportunities[^9]. On the supply side, AI tools for real-time manufacturing process monitoring, control, and yield optimization are being implemented to maximize output from these constrained production environments[^10].
Allogeneic (off-the-shelf) cell therapies and gene therapies from engineered cell lines face their own forecasting challenges. While they aren‘t reliant on individual patient donors, they still face production capacity constraints and raw material supply risks. Demand forecasting must account for expected market penetration, duration of therapeutic effect, and potential re-dosing. From a supply perspective, advances in AI-optimized media design, process control, and facility digital twins will be critical to produce these therapies at commercial scale[^11].
Digital Twins and Simulation for End-to-End Supply Chain Optimization
Given the complexity of pharmaceutical supply chains for PDTs, cell therapies, and gene therapies, it‘s clear that traditional forecasting methods alone aren‘t sufficient. Cutting-edge digital twin and simulation technologies powered by AI and ML are emerging as key tools for end-to-end supply chain optimization[^12].
Digital twins are virtual replicas of physical assets, processes, and systems that allow real-time monitoring, predictive modeling, and scenario analysis. By integrating data across the entire pharmaceutical value chain (e.g. plasma collection, manufacturing, distribution, and patient demand), companies can create a holistic view of their supply network.
AI and ML algorithms can then be applied to optimize supply chain design, identify risks and bottlenecks, and simulate the impact of strategic decisions or unexpected disruptions. For example, a digital twin could be used to evaluate the trade-offs between expanding plasma collection capacity, building new fractionation plants, or outsourcing production for a new PDT. By running thousands of simulated scenarios, companies can develop robust plans to meet patient needs under a range of demand forecast and supply constraint assumptions.
Conclusions
Forecasting for pharmaceutical products with constrained raw material supplies is a complex challenge requiring cross-functional collaboration and scenario planning. Plasma-derived therapies are a key example where patient demand must be balanced against plasma collection and allocation decisions. The step-by-step methodology outlined in this article – from epidemiology analysis to demand and supply forecasting – provides a roadmap for PDT forecasting best practices.
However, traditional forecasting methods alone are no longer sufficient given the data complexity and optimization challenges manufacturers face. Artificial intelligence and machine learning are transforming the field of pharmaceutical forecasting by enabling:
- More accurate epidemiological models and patient projections
- Predictive models for optimizing raw material pooling, yields, and allocations
- Dynamic pricing, market share, and promotional response models
- Digital twins and simulation tools for end-to-end supply chain optimization
While we focused examples on PDTs, many of these AI/ML techniques are also being applied to overcome forecasting challenges for novel cell and gene therapies facing similar raw material constraints. The ability to analyze massive volumes of structured and unstructured data is driving a paradigm shift in forecasting from an art to a science.
Implementing robust forecasting processes powered by AI ultimately helps pharmaceutical manufacturers make informed investment decisions, optimize supply chains, and ensure reliable patient access to life-changing therapies. The patients are counting on us to get it right.
References
[^1]: Géron, A. (2019). Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow: Concepts, tools, and techniques to build intelligent systems. O‘Reilly Media.[^2]: Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. Annals of Statistics, 1189-1232.
[^3]: Copeland, J., & Copeland, M. (2020). Artificial intelligence in supply chain management. Procedia Manufacturing, 51, 1634-1639.
[^4]: Wu, Y., & Zhou, S. (2020). Machine learning-based approaches for pharmaceutical pricing and market access. Drug Discovery Today, 25(4), 610-618.
[^5]: Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction machines: The simple economics of artificial intelligence. Harvard Business Press.
[^6]: Curling, J., Gottschalk, U., & Rüker, F. (2020). Process optimization of plasma fractionation using machine learning. Biotechnology Progress, 36(2), e2947.
[^7]: Grifols. (2021). Grifols Investor Day 2021. https://www.grifols.com/documents/10192/39926153/grifols-investor-day-2021-en.pdf/0c8c1ae6-0ef2-4d73-9b26-b41b265b209c
[^8]: Aijaz, A., Li, M., Smith, D., Khong, D., LeBlon, C., Fenton, O. S., … & Fahmy, T. M. (2018). Biomanufacturing for clinically advanced cell therapies. Nature Biomedical Engineering, 2(6), 362-376.
[^9]: Harrison, R. P., Rafiq, Q. A., & Medcalf, N. (2018). Automating decentralized manufacturing of cell and gene therapy products. Cell and Gene Therapy Insights, 4(7), 489-492.
[^10]: Moutsatsou, P., Ochs, J., Schmitt, R. H., Hewitt, C. J., & Hanga, M. P. (2019). Automation in cell and gene therapy manufacturing: from past to future. Biotechnology Letters, 41(11), 1245-1253.
[^11]: Pigeau, G. M., Csaszar, E., & Dulgar-Tulloch, A. (2018). Commercial scale manufacturing of allogeneic cell therapy. Frontiers in Medicine, 5, 233.
[^12]: Lim, W. S., & Liang, C. K. (2021). Digital twins in pharmaceutical industry: a multi-scale simulation approach. Advanced Drug Delivery Reviews, 179, 114000.
