Applying Data Science and Artificial Intelligence to Managing Biomedical Portfolios

A promising biological discovery is not yet a medicine, and a promising medicine is not yet a financeable investment. Biomedical portfolio management connects scientific evidence, experimental decisions, patient needs, and capital that can remain available through years of uncertainty.

The central question is how to fund a sequence of useful experiments without letting a single failure—or a shortage of cash—destroy the whole enterprise. Data science can improve estimates and decisions. Financial engineering can distribute and finance risks. Neither replaces reliable scientific evidence.

These notes expand the supplied lecture outline associated with Andrew W. Lo’s biomedical-finance work. Its numbered reference fragments lack a bibliography and are replaced with linked primary research and company or regulatory records. Historical examples are dated, hypothetical calculations are identified, and the notes do not claim to reproduce an investment backtest or clinical prediction model.

Learning goals

Explain the translational funding gap and distinguish scientific from financing risk.
Model single-project outcomes and portfolio diversification without assuming independence.
Connect staged experiments, AI predictions, and liquidity-aware capital allocation.
Interpret the BridgeBio case without treating one company’s experience as proof of universal success.
Extend the framework cautiously to fusion and other deep technologies.

For quantitative background, see Portfolio Management, Regression Analysis, and Counterparty Risk Optimization.

1. Why scientific progress can outpace investment

Modern biology produces many kinds of information:

Field	What is measured	Possible development question
Genomics	DNA variation and sequence	Is there evidence linking a target to disease?
Epigenomics	Regulatory modifications and chromatin state	In which cells and conditions might a pathway be active?
Transcriptomics	RNA expression	Does the proposed mechanism change relevant cell activity?
Metabolomics	Small molecules and metabolic patterns	Does treatment alter a disease-associated biochemical pathway?
Microbiomics	Microbial communities and their functions	Could microbial composition relate to response or disease?

These measurements can generate hypotheses and help prioritize targets. Association, prediction, mechanism, and clinical benefit remain different claims. A strong omics signal may fail to translate because of delivery, toxicity, human biology, manufacturing, trial design, or an incorrect causal hypothesis.

The valley of death describes a gap between early discoveries and the evidence or development progress needed to attract later-stage funding. It is not always one fixed point. A company can also face a financing gap after favorable trials but before commercial revenue supports continued operations.

Scientific risk asks whether a treatment works safely for its intended use. Financing risk asks whether enough suitable capital remains available to reach the next informative milestone. They interact: a delay can consume cash, force a poorly timed financing, or stop a scientifically worthwhile program.

2. Read single-drug statistics with their assumptions attached

The supplied outline combines roughly 200 million of development cost, ten years, a 3.4–5% success probability, approximately 12% expected annual return, and approximately 423.5% standard deviation. These are not interchangeable current industry constants.

The 2012 Fernandez–Stein–Lo paper uses a stylized 200-million, ten-year oncology program with 5% success, and illustrates pooling 150 programs into a 30-billion vehicle. Its roughly 11.9% figure annualizes expected terminal value under specified commercial-payoff assumptions; it is not a guaranteed annual return. The paper reports a roughly 423% single-draw risk statistic and discusses dependence as a limit to diversification.

Separately, Wong, Siah, and Lo’s historical study reports an oncology success estimate of 3.4% for its sample and method. That historical probability is not automatically appropriate for a specific current candidate, stage, modality, or rare disease.

Why annualization can mislead

For a terminal payoff V_T and initial cost C, (E[V_T]/C)^(1/T)−1 is generally different from E[(V_T/C)^(1/T)−1]. Failure paths can lose all invested capital, and funding arrives or is spent at different dates. A Sharpe calculation must match the definitions and horizon of its expected excess return and standard deviation.

Even if the outline’s 12% and 423.5% were directly comparable, the zero-cash-rate ratio would be only 0.12/4.235≈0.0283. A positive risk-free rate lowers the excess-return numerator. That arithmetic illustrates extreme dispersion; it does not independently validate those inputs or make Sharpe the complete risk measure for illiquid, binary projects.

3. A transparent single-project model

For a deliberately simplified example, invest C=10 million now. At a specified future date, the project pays V=100 million with probability p=0.2, and zero otherwise. Ignore interim spending, taxes, financing, recoveries, and time discounting.

Let I be a Bernoulli success indicator. The terminal net return is:

R=(V/C)I−1=10I−1.

Thus E[R]=10p−1=100%, and SD(R)=10√[p(1−p)]=400% over the full project horizon. The large upside creates a positive mean even though failure occurs 80% of the time. These intentionally simple numbers are not biomedical return forecasts.

For n equal-cost independent projects, the success count S is binomial and portfolio return is R_p=(10/n)S−1. Its mean remains 100%, while standard deviation is 400%/√n.

At n=20, standard deviation is about 89.44%, and the probability of at least one success is 1−0.8²⁰≈98.85%. But total cost is 200 million: one 100-million success still loses half the capital. Two successes break even before costs; at least three are needed for a strictly positive return. The probability of a strictly positive return is only about 79.39%.

At least one success, preservation of capital, debt repayment, and an attractive equity return are different events. A portfolio model must calculate the event relevant to its investors and cash obligations.

4. Diversification depends on joint risk, not project count alone

For equal-weight projects with common variance σ² and average pairwise correlation ρ, the familiar expression is:

Var(R_p)=σ²[ρ+(1−ρ)/n].

With ρ=0, volatility falls as 1/√n. With positive common correlation, residual risk remains as n grows. The formula describes return covariance; a biologically plausible joint model is still needed for success counts, timing, and tail events.

Rare-disease programs with different causal mechanisms may diversify some scientific risks. They can still share delivery technology, manufacturing capacity, regulators, trial sites, reimbursement assumptions, or the same financing window. A portfolio of different disease names can be concentrated in one platform failure mode.

What 10–20 projects can and cannot promise

For independent projects with success probability 5%, twenty projects give 1−0.95²⁰≈64.15% probability of at least one success. With 150, it is approximately 99.95%. Neither probability alone establishes enough revenue to repay lenders or fund remaining trials.

Under an idealized matching-horizon equal-mean, independent-return model, Sharpe scales as √n if the excess-return mean is unchanged. Starting at 0.0283, twenty projects would produce approximately 0.127, not above one. Correlation, expenses, and financing can reduce that idealized improvement. The outline’s Sharpe-above-one claim therefore needs a different specified economic model and evidence; it does not follow from adding 10–20 low-correlated projects.

Portfolio aggregation can create a more financeable risk distribution, but a megafund is not defined simply by crossing a universal minimum project count. Capital requirements depend on development stage, costs, cash-flow timing, dependence, and investor obligations.

5. Financial structure: who absorbs losses, and when?

A portfolio vehicle can raise equity and, where supportable, different debt or royalty-linked claims. A financing waterfall allocates cash flows in a specified order. Senior debt has priority, while equity or subordinated capital absorbs losses first. This reallocates risk; it does not manufacture additional drug efficacy or eliminate aggregate economic losses.

Capital source	Potential role	Main modeling issue
Common equity	Absorb uncertain early development outcomes	Dilution, governance, and large loss probability
Long-dated debt	Finance sufficiently supported future cash flows	Default probability, cash coverage, covenants, and timing
Royalty financing	Exchange part of future product revenue for present funding	Commercial concentration and long-run retained economics
Licensing or asset sale	Realize value at intermediate milestones	Negotiating position, timing, and foregone upside
Grants or philanthropic first-loss capital	Support socially valuable or otherwise underfunded work	Mandate, transparency, and who ultimately bears the risk

The related orphan-disease financing research studies how scientific and economic features influence financing feasibility. A debt tranche should be evaluated on its actual projected payment and recovery distribution, not on a portfolio-wide average success count alone.

Liquidity can fail before terminal value is realized

Consider a vehicle with 200 million available and planned spending of 60, 70, and 50 million over three years. It appears to have a 20-million buffer. A 25% cost overrun makes those cash needs 225 million, creating a 25-million gap unless spending, funding, or realizations change.

Expected valuable assets in year ten do not pay trial invoices in year two. Model cash balances through time, refinancing access, milestone receipts, delays, and contingency reserves. New debt may reduce equity dilution while increasing the cost of a delay or failed readout.

6. Development is a sequence of decisions, not one coin toss

Drug development can be represented as states—discovery, preclinical, clinical phases, submission, approval, and commercialization—with possible failure, pause, sale, or redesign at each stage. A transition probability should be conditional on the stage reached and the evidence available.

If successive conditional success probabilities are 0.5, 0.4, and 0.7, the probability of completing all three stages is 0.5×0.4×0.7=14% under those stated conditional probabilities. Using a late-stage success rate for a discovery-stage candidate overstates its prospects.

At each milestone, update the evidence and decide whether to continue. A schematic decision rule compares:

continue value = −next-stage cost + discounted E[best subsequent action value | current evidence]

against the value of stopping, selling, or redirecting the project. The option to stop can preserve capital, but stopping rules must be scientifically and ethically sound. Patient safety, protocol obligations, and trial integrity cannot be reduced to an investor payoff threshold.

Value of information

Suppose the best current action has expected value 8 million. A proposed experiment would produce information allowing an expected best decision worth 11 million before paying for the experiment. If it costs 2 million and timing is otherwise comparable, its modeled net value of information is 11−8−2=1 million.

That result depends on how informative the experiment is and what decisions it can actually change. Collecting more data is not automatically valuable if no feasible decision responds to it.

7. Where data science and AI can help

Task	Possible method	Decision it may improve
Target and mechanism prioritization	Multimodal omics models and causal-evidence review	Which hypotheses merit an experiment?
Trial-outcome estimation	Calibrated classification or survival models	How much capital and time should be reserved?
Patient and site feasibility	Recruitment and event-rate models	Can a trial complete on a realistic timeline?
Evidence synthesis	Retrieval and language models with source verification	What is known, missing, or contradictory?
Commercial scenarios	Epidemiology, uptake, payer, and competition models	What revenue paths are plausible after approval?
Portfolio simulation	Bayesian models and correlated Monte Carlo	What combinations survive delays and failures?
Capital allocation	Constrained optimization or dynamic programming	Which feasible set of next experiments offers the best trade-off?

A useful system links predictions to decisions. A model that ranks promising targets well may still produce badly calibrated probabilities for a financing model. An accurate approval classifier does not directly predict patient recruitment, required cash, reimbursement, or product sales.

Build the data around the decision date

Define whether the model predicts phase advancement, regulatory approval, time to readout, or commercial success. Retain provenance, timestamps, program identity, stage, modality, endpoint, and sponsor information. Distinguish failure from discontinuation for commercial reasons and from a project whose outcome is still unknown.

Avoid leakage from later publications, retrospective registry updates, post-outcome press releases, duplicate compounds, or related trials split between training and test sets. Temporal splits and grouping by compound, target family, or platform can expose weaknesses that a random row split hides. Model current uncompleted programs with appropriate censoring rather than treating them all as failures or successes.

Evaluate probabilities and uncertainty

Discrimination asks whether successes rank above failures. Calibration asks whether, for example, projects assigned 20% success actually succeed around 20% of the time in a suitable evaluation cohort. Both matter, but financing is especially sensitive to probability calibration and shared error.

Use held-out calibration curves and proper scores such as Brier score=(1/N)Σ(pᵢ−yᵢ)² or log loss, alongside time-to-event evaluation when relevant. Report uncertainty and subgroup performance, not just a single accuracy or AUC. Rare diseases can have small samples, and a high overall accuracy can simply reflect predicting failure for nearly everything.

An AI estimate is not clinical evidence that a treatment works. For AI-derived information supporting regulatory decisions, the FDA’s January 2025 draft guidance describes context-of-use and risk-based credibility assessment; it is explicitly draft, nonbinding guidance. The FDA–EMA January 2026 guiding principles provide a further framework for good AI practice across the medicines lifecycle. These documents do not mean every discovery model is validated or approved by a regulator.

8. An integrated portfolio model

For each project, track stage-dependent probabilities, conditional costs, timing, possible milestone or licensing proceeds, commercial cash flows, and uncertainty about every major input. Include common shocks rather than drawing every project independently by default.

For simulated scenario s and decision policy π, let CF_s,t(π) be cash flow at date t and Cash_s,t(π) the resulting available cash. One illustrative decision formulation is:

maximize E[NPV(π)]−λ ES_α(loss(π)),

subject to stage budgets, scientific capacity, ethical and protocol requirements, diversification limits, and liquidity conditions such as P(min_t Cash_s,t(π)<0)≤ε.

This is a schematic objective, not a claim that real biomedical optimization is automatically convex or exactly solvable. Stage-gated binary decisions, chance constraints, dependence, and uncertain probabilities can make computation difficult. The trade-off parameter λ and acceptable shortfall ε reflect a stated mandate, not values chosen merely to make a presentation attractive.

Stress at least: a shared-platform safety issue, trial delays, simultaneous funding closure, manufacturing problems, lower net product prices, slower uptake, and several failures before any saleable asset emerges. Portfolio safety depends on the cash-flow path and financing waterfall as well as expected terminal NPV.

9. BridgeBio: a documented operating example

BridgeBio is relevant because it combines multiple programs with asset-focused teams and shared organizational resources. The 2024 case study by Kumar, Lo, Shukla, and Stephenson describes both its portfolio approach and the practical work of obtaining suitable capital. It is an operating company case, not a literal implementation of every securitized megafund design.

Verified milestones and their boundaries

BridgeBio’s 2019 annual filing records its 2015 founding and 2019 IPO history.
The 2024 case study reports approximately 401 million of gross IPO proceeds including the underwriters’ overallotment. This is a dated financing event, not the initial founding round or a current market capitalization.
The company’s 2024 annual filing records FDA approval of Attruby on 22 November 2024 and describes earlier approvals and financing sources.
Its 2025 annual filing distinguishes products it markets from approved products marketed by others, and identifies Truseltiq’s approval as withdrawn. Historical approval, current availability, and current ownership are different statuses.

The supplied outline’s 400–500-million early funding description, 15-project snapshot, roughly 4.6-billion market capitalization, and “over 20 projects in clinical trials” lack matching dates and definitions. They should not be presented as today’s capitalization or pipeline. Market value changes, and a broad development portfolio is not the same count as active clinical-stage programs.

Neil Kumar is a co-founder and CEO, and the portfolio-finance connection is documented by the coauthored case study. The supplied “former student” description is not needed to establish the economic argument and is not independently verified here.

What the case demonstrates

It demonstrates that a portfolio-oriented biomedical company can raise multiple types of financing and bring products through development. It does not prove diversification caused every approval, guarantee debt safety, establish an investment return for all shareholders, or show that every proposed megafund will work. Scientific selection, execution, financing terms, and realized setbacks all affect the result.

10. Extending the framework to fusion and deep technology

Fusion shares some financing difficulties with drug development: substantial upfront capital, long time to meaningful cash flows, technical milestones, and uncertain commercial success. A portfolio can include alternative scientific approaches and enabling technologies with potentially separate customers.

The 2023 Financing Fusion Energy research proposes a portfolio and financing framework, while MIT’s commercial-fusion ecosystem report examines opportunities across materials, components, systems, software, and services. These are research and ecosystem arguments, not guarantees of a commercially viable power plant.

Illustrative component opportunities include superconducting magnets, vacuum systems, diagnostics, and control software. Their value outside fusion must be established through technical requirements, customers, unit economics, and intellectual-property rights. A technology useful in a fusion system is not automatically qualified for pharmaceutical manufacturing or another regulated market.

Intermediate licensing or sales can reduce dependence on one final plant outcome. Yet component companies may still share the same capital cycle, supply chain, specialist workforce, or fusion customers. Dual-use possibilities create options, not automatically uncorrelated revenue streams.

The same discipline applies to biomedical platforms: identify which intermediate assets have real buyers, how long qualification takes, and whether selling them compromises the primary development program.

11. Careers and “doing well by doing good”

The valuable combination is quantitative skill plus enough domain understanding to know what the numbers mean. Biology informs target credibility; clinical expertise informs endpoints and feasible experiments; statistics handles uncertainty and bias; finance translates cash-flow risks into sustainable funding.

A useful portfolio-data-science project would build a time-stamped development dataset, estimate and calibrate stage transitions, simulate correlated costs and outcomes, and compare financing policies under stress. Its deliverable should explain which decisions improve and how much confidence the evidence supports—not simply display an impressive model score.

Scientific impact and financial return can align, but they are not identical objectives. Include patient benefit, access, affordability, and safety explicitly where they belong in the mandate. A compelling business narrative should communicate evidence and cash needs clearly rather than replace them. These are interdisciplinary opportunities, not advantages exclusive to one university.

12. Practice questions

If independent project success probability is 5%, what is the probability of at least one success among twenty projects?
In the 10-million-cost, 100-million-payoff toy model, do one or two successes guarantee a positive return for a twenty-project portfolio before costs?
Why can rare-disease projects remain correlated despite treating different conditions?
What happens to equally weighted portfolio variance as project count grows when common positive correlation remains?
Why can strong AUC still be insufficient for a financing model?
A program must pass three stages with conditional success probabilities 0.5, 0.4, and 0.7. What is the overall completion probability under those inputs?
Does an IPO or an approved drug establish that a portfolio financing model is risk-free or profitable for every investor?
Why is a fusion component’s possible use in another industry an option rather than an assured independent cash flow?

Answers

1−0.95²⁰≈64.15%, assuming independence.
No. Total cost is 200 million: one success loses money, two break even before costs, and three or more yield a positive return in this simplified model.
They can share platforms, manufacturing, regulators, trial infrastructure, reimbursement assumptions, and financing conditions.
It approaches a positive common-risk floor rather than zero under the equal-variance formula.
Ranking quality does not ensure calibrated probabilities, reliable dependence estimates, or adequate performance on genuinely new projects.
14%.
No. Financing terms, investment dates, ownership, costs, and subsequent outcomes still matter.
Qualification, demand, pricing, rights, and timing must be demonstrated; potential applicability alone does not create revenue or remove shared risks.