Exploring a “Patient-First” Path in Drug Discovery

We read Bruce Booth’s recent commentary, “Biotech Wisdom of the Crowds: Competition and Capitalism," with great interest—as we have with all of his posts over the years. His insights into the structural forces shaping biotech R&D are timely and thought-provoking, particularly his reflections on target crowding, capital dynamics, and the comparative advantages of technology startups.

We want to offer a perspective from our work, which we see as a complementary path in

drug discovery—what we’ve considered a “Patient-First Path.” This approach doesn’t

seek to replace the current model but offers an adjacent strategy that may help unlock

additional therapeutic opportunities, particularly for patients with unmet needs.

The “Target-First Path” Works—But Also Facing Challenges

As Bruce rightly notes, biopharma has delivered incredible advances. The Target-First

Path, anchored in target-based discovery, modality specialization, and capital-driven

focus, has brought transformational therapies to market. But it is also constrained by

long timelines, high failure rates, and a tendency to converge around a relatively narrow

set of biological targets. Capital has been flowing towards comfortable ideas instead of

true unmet medical need. Ironically the comfortable bets actually offer the least

differentiation and return where a setback for one company can trigger sector-wide

corrections regardless of the individual merits of each program.

Over the past 25 years, our teams have worked at the intersection of genomics,

machine learning, and translational science. Recent advances in computation, AI,

and data infrastructure now make it possible to support the Target-First Path with a

complementary discovery engine, one that starts from clinical relevance and scales

through technology.

Patient-Signal First: Starting With Patients, Not Just Targets

In this Patient-First Path, we begin not with a predefined target but with biomarker

signatures derived from real patient cohorts. These signatures serve as

classifiers—grounded in clinical outcome data—that guide the discovery process from

the outset. Rather than search for molecules that bind a particular protein, we use large-

scale virtual screening and generative AI to identify molecules predicted to shift gene

expression patterns toward those associated with desired clinical outcomes.

This is not meant to sidestep causality. Many of us have spent years building causal

modeling platforms and understand the importance of mechanistic insight. Instead, we

see this approach as a way to broaden the search space early, prioritize relevance,

and bring patient biology into discovery much sooner.

Identifying Actionable Biology for Unmet Need

A new generation of patient classifiers is emerging that incorporates large multimodal

datasets integrated and interrogated with artificial intelligence. These classifiers serve

as biomarkers of drug response, and also explainers of non-response. They incorporate

dozens to hundreds of molecules rather than relying on single gene mutations or target

protein expression. Gene expression measured by RNA-seq is a gold standard for

building such algorithms, as it simultaneously describes disease phenotypic and genetic

background.

By training AI to learn disease biology, and by measuring a rich, deep, and

comprehensive swath of biology, the new patient classifiers generalize well across

clinical settings and histologies. Validation of these algorithms in real world and clinical

trial contexts lends confidence that they have sufficiently learned the underlying disease

biology. These algorithms serve two purposes. The first is to help identify patient

subgroups that are likely to benefit from established and investigational therapies

discovered via the “target-first path”. The second is to find patients who will not benefit

from these drugs and thus remain needful of new medicines designed in a new way.

Why It Matters Now

Several dynamics make this moment especially well-suited for such an approach:

• Relevance: Starting from real patient cohorts and disease-linked classifiers

ensures that molecules are aligned with clinical context from day one.

• Scale & Speed: With virtual catalogs of trillions of molecules and fast generative

modeling, we can screen billions of compounds in days—then synthesize promising

candidates for wet-lab validation within weeks.

• Cost Efficiency: Early-stage discovery efforts that once cost tens of millions can

now be executed for hundreds of thousands of dollars in computing, enabling broader

exploration at lower risk.

• Technology Readiness: Platforms now exist that can model likely biological

effects (e.g., gene expression shifts) across different concentrations and cell types,

offering mechanistic clues alongside predictive performance

Embracing Complexity, Not Avoiding It

We also believe this approach allows us to embrace the biological complexity that

often only emerges late in traditional development. Polypharmacy, off-target effects, and

cell-type specificity are not liabilities—they are part of the initial screen. This perspective

draws on findings like those of Smith and Sheltzer (Cell Reports, 2022), which remind

us that predictive biomarkers don’t always equate to good targets—but they remain

valuable sensors of disease biology.

The Patient-First Path doesn’t discard targets. It often suggests new uses for known

targets or generates hypotheses for new intervention points. In some cases, it highlights

alternative ways to modulate the biology downstream of challenging or crowded

targets—offering new angles on well-trodden ground.

System 1 and System 2 Thinking: Both Are Needed

Psychologist and Nobel Laureate Daniel Kahneman famously described two modes of

human thought: System 1 is fast, intuitive, and pattern-driven. System 2 is slow,

deliberate and analytical. Both systems are essential to drug discovery, with System 1

allowing for speed and breadth of insight, and System 2 bringing rigor and reductive

analysis. Using Target-First paths, we’ve relied heavily on System 2 thinking:

hypothesis-driven, target-first, step-wise validation, and careful reductionism. We are

proposing a complementary approach, a System 1-inspired discovery engine using AI

and patient-derived data to create a parallel path that can move faster, see patterns,

explore broader, and start closer to the patient. We still validate rigorously – but we start

from the patient signal, rather than prior target assumptions.

Partnership as a Foundation

We’ve seen the Patient-First Path not as a product of any single company but as a

networked effort. It combines capabilities across AI-driven screening, chemical

synthesis, clinicogenomic data, and patient-relevant models. We’ve found that

collaboration, not vertical integration, is often the key to speed and rigor.

In Closing

We’re encouraged by the progress we and others are seeing using this approach—from

rapid preclinical validation to identifying molecules with performance that matches or

exceeds traditional benchmarks. We believe the Patient-First Path can help accelerate

the delivery of relevant, effective therapies—particularly in areas where the Target First

Path has struggled to find traction.

We share Bruce’s optimism about the future of biotech and as always, value his voice in

shaping the conversation. We simply offer this other path as an additional tool—one that

may help the industry work faster, fail smarter, and stay closer to the biology that

matters most: the patient's.