Key Development
AI drug development is moving beyond molecule design toward a more expensive bottleneck: predicting whether an experimental medicine is likely to work in humans before enormous capital is committed to clinical trials.
Reuters reports that companies including BioinvestGPT and QuantHealth are constructing virtual patients and simulated trials to forecast efficacy, safety signals and trial outcomes. BioinvestGPT predicted that Novartis’ Phase 3 HARBOR study of del-desiran would fail to generate meaningful clinical benefit before the result was known; the study subsequently missed its primary endpoint. Of six major trial readouts the platform shared prospectively with Reuters, five predictions were correct and one was wrong. That is intriguing, but far too small a sample to establish general predictive performance.
The economic incentive is substantial. The biopharmaceutical industry spends roughly $140 billion annually on human clinical testing, Reuters reports, while only about 12% of drug candidates ultimately reach regulatory approval. Some virtual simulations can be completed within a month or less, compared with years for conventional clinical development.
Why It Matters
- The largest ROI may come from avoiding the wrong trial, not merely shortening the right one.
- Simulation can become a capital-allocation tool across go/no-go decisions, indication selection, protocol optimization and asset licensing.
- Regulatory infrastructure is emerging. FDA finalized ICH M15 in 2026 and explicitly prioritizes clinical-trial simulation, outcome prediction and mechanistic modelling within its MIDD program.
- Human randomized trials remain the evidentiary gold standard. AI is better positioned today to augment design and decision-making than replace efficacy and safety evidence.
Healthcare Insight Analysis
Healthcare Insight believes the framing that “AI will replace clinical trials” misses the more economically consequential opportunity.
The better question is:
How many doomed trials could AI prevent from ever starting?
Traditional drug development often discovers biological weaknesses late:
target → Phase 1 → Phase 2 → Phase 3 → failure.
Virtual-trial technology aims to insert another gate:
target → virtual stress test → protocol/cohort optimization → kill or redesign → human trial.
That changes the economics of R&D.
A failed Phase 3 program destroys more than trial expenditure. It consumes years of patent life, manufacturing preparation, executive attention and investor confidence. Reuters noted that Novartis shares fell 11% after del-desiran’s failed readout.
However, reliable simulation requires much more than finding statistical patterns.
A useful model must capture:
disease biology + mechanism of action + patient heterogeneity + dose + comparator + endpoint + protocol design.
The limits are already visible.
BioinvestGPT incorrectly forecast success for Novartis’ pelacarsen cardiovascular study. Its founder later said the model had misunderstood part of the mechanism involving genetically determined variation in lipoprotein(a) particle size.
That failure is strategically important.
An AI model can process enormous quantities of data and still be confidently wrong when its causal representation of biology is incomplete.
QuantHealth uses a different architecture built around real-world patient data and predicted patient-level treatment responses. The company says it has simulated more than 600 trials and achieved up to 90% endpoint-prediction accuracy. Those figures are company-reported, however, and should not be interpreted as an independently established industry-wide accuracy rate.
This makes model credibility the critical competitive moat.
A serious virtual-trial platform will increasingly need to show:
prospective prediction → timestamp before readout → blinded validation → cross-disease performance → population transportability → lifecycle monitoring.
Retrospectively reproducing known outcomes is not enough.
There is also a hidden portfolio risk.
The diseases with the weakest historical datasets — rare diseases, new mechanisms and first-in-class therapies — are often exactly where innovation is most valuable. Over-reliance on historical AI could systematically reject unfamiliar biology and produce false-negative innovation filtering.
The optimal future model is therefore not:
AI replaces scientific judgment.
It is:
AI attacks the investment thesis before companies expose patients and capital to it.
Market Implications
Big Pharma: virtual simulation could become standard portfolio governance and licensing due diligence.
Biotech: prospective simulation may become supportive evidence for financing, but cannot substitute for human data.
CROs and clinical-tech companies: value creation could shift from executing more trials toward designing fewer, better ones.
Regulators: context-of-use, data provenance, validation, auditability and model-version control will become increasingly important regulatory capabilities.
ANALYTICAL MATRIX
| Use Case | Potential Value | Main Risk | KPI |
| Go/no-go | Avoid failed trials | False negatives | Prospective accuracy |
| Trial design | Better cohort/endpoints | Model bias | Power/sample-size gain |
| Licensing | Better asset diligence | Missing proprietary data | Calibration |
| Digital twins | Patient-level prediction | Generalizability | Predicted vs observed |
| Regulatory evidence | More efficient development | Model credibility | Agreed context of use |
The industry’s most important metric therefore should not be:
“How accurately can AI explain trials whose results we already know?”
It should be:
“How accurately can it predict blinded future trials — and do decisions based on those predictions improve portfolio returns?”
If that standard is eventually met, virtual trials will not eliminate human clinical studies.
They could do something more economically powerful: ensure that only programs with sufficiently resilient scientific theses proceed into the industry’s most expensive stages — while reducing the number of patients exposed to trials whose failure could have been identified earlier.