It's time for decision sciences to transform early drug discovery
AI use disclosure
All of the prose below was hand-written by myself unless otherwise stated. I used Claude Code (Sonnet 5 / Opus 4.8 / Opus 5) to support and pressure test my research, identify opportunities to improve my arguments, and to format and render the text. I conceived of the visuals myself, and used Claude to render them.
I recently had the pleasure of joining the folks at convoke.bio to participate in their lunch-and-learn session. This is the team that released the unmet needs index back in May, and more recently released an MCP for agents to query their global drug-development pipeline database.
We had a great discussion about target prioritization, and how that work is evolving in the AI era. A highlight was when I unwittingly quoted their co-founder Alex Telford’s own blog post back to him in the conversation:
“If you automate 5 processes with AI, and do the remaining 995 the old way — you’re still pretty much a regular biotech”
In my view, the field of decision sciences provides the unifying framework to automate the remaining 995 processes, and target prioritization is a good place to start, since it’s a massive search space, analysis intensive, and typical tools used to simplify the process leave a trail of unpicked opportunity ripe for AI to surface.
We talked about the difference between automating existing workflows and analyses using AI vs transforming workflows entirely in the AI age. This was the theme of my talk on Target ID, which forms the basis of this post.
TL;DR:
- Target prioritization approaches today use simplifying assumptions to reduce the analytical burden needed to drive decisions
- AI tools relieve the burden, allowing greater breadth of search, and greater depth in analysis before reaching a functional expert for manual diligence
- Moreover, AI enables decision sciences tools typically reserved for later-stage programs due to their analytical intensity
- These techniques center the next best action—rather than the target’s current value—to unearth latent opportunities where an inexpensive action can lead to outsized payoffs
- Beyond using AI to accelerate and expand target ID workflows, early discovery teams should collaborate with their portfolio analytics teams to transform them entirely with decision sciences
Why target prioritization is hard
Target prioritization is the very first decision on the long, $2B+ road to bringing a drug into the clinic, and it’s a challenging one for several reasons:
- It’s a massive search space: ~20,000 genes in the human genome, only 854 of them have FDA-approved drugs to date.
- The decision criteria span disparate and deeply technical domains.
- The known-unknowns and unknown-unknowns are enormous across all of them.
In the process, you’re generally trying to answer these questions simultaneously across 20,000 genes, multiplied by each of their associated diseases:
- How do you predict the commercial opportunity of a drug 12–15+ years in advance?
- How well can you truly evaluate the strength of a biological hypothesis before lifting a pipette?
- How easily will you be able to actually identify a molecule that is suitable for clinical development?
- Amidst all this, where is your competitive edge?
Each of these questions requires different expertise, and desk research can only take you so far on any one of them. Before you have enough confidence to make a decision, you need to pick up the phone to call an expert, replicate a biological experiment, synthesize or purchase a tool molecule, and so on.
Normally, we think about those next steps at the very end of the process. With AI, I believe we should be thinking about the “next steps” from the very beginning. It all boils down to unearthing pockets of opportunity that are overlooked through typical prioritization approaches.
Target prioritization: BCE (Before the Claude Era)
In my experience, this question typically plays out with a common playbook:
- Cut the universe down to a tractable size with sweeping assumptions.
If you’re lucky, you can get the list down to 50–100; bioinformaticians are your friends here.
Some sweeping assumptions are necessary. If you’re focused on mAbs, it makes sense to eliminate intracellular targets. If your focus is on oncology, it’s safe to eliminate metabolic targets with no known oncology linkage.
But some cuts primarily exist to simplify the search space while accepting some false negatives. For instance, screening for ‘first-in-class’ opportunities is more straightforward than teasing apart the literature for signals from which ‘best-in-class’ drugs can emerge. Or sometimes, you might cut anything with an aggregate ‘genetic validation’ score below some threshold to enrich for targets with higher biological evidence.
I’ll elaborate in more detail in the next section, but AI can unlock new avenues to minimize those false-negatives.
Score the remaining targets against a defined rubric that reflects your organization’s priorities and capabilities.
Here, we’re trying to build a red/yellow/green scorecard as efficiently as possible to illustrate trade-offs between the remaining targets and surface the best opportunities of each trade-off archetype. Maybe you land at 5–10 targets.
- Run deeper, manual diligence on a shortlist.
This is the bottleneck: assembling the cross-functional experts to profile each shortlisted target, validate initial findings, articulate a clear value proposition, surface key risks, recommend mitigation plans, and lay out a clear path to value inflection. On closer inspection, some targets might get ruled out entirely.
Select a promising subset (1–3) to begin exploratory experimentation, or to discuss with external experts in order to de-risk areas of concern or validate assumptions that have high sensitivity/uncertainty.
What does this approach cost?
- Functional experts have finite time. This generally isn’t their day job, and these evaluations can be time-intensive. You can reduce the burden by reducing false positives.
- Missed opportunities. Coarse assumptions in the screening stages can miss pockets of opportunity. For instance, a rudimentary red/yellow/green scoresheet might not be designed to capture a poor-scoring target that, with an inexpensive experiment, might be turned green.
Target prioritization, accelerated with AI
The 80/20 rule goes out the window.
AI enables scoring a larger universe of targets
AI allows automated scoring at a more granular level
Functional expert review is focused on validating and refining the scores—and learnings from the corrections should be embedded in the workflow for the next pass.
When done right, AI can improve this process in two ways:
“Boil the ocean”: scoring analyses can go deeper before they reach functional experts’ eyes. AI tools are getting better at domain expertise and analysis, and companies can further encode company-specific lenses into these analyses. As a result, experts’ roles transition from doing the last mile of analysis to reviewing the AI’s work.
“Boil the universe”: by making fewer simplifying assumptions at the top of the funnel, teams can increase the breadth of opportunities assessed and accelerate the speed of the process.
Altogether, you can score every relevant target/disease permutation, instead of filtering down to 50–100. You can analyze unmet need at the patient segment level rather than relying on disease area averages. You can surface competitor liabilities to identify best-in-class opportunities, rather than setting a threshold competition level. Set things up right, and you can find a needle in the haystack.
Set it up wrong, and your AI tool will find you 20 pieces of hay, call them needles, and when you mash your keyboard in frustration, Claude will respond unabashedly, “You’re right to push back—…”
The risk: systematic biases are amplified and distort the field
- Previously, systematic biases would get pruned at the last mile. At scale, those systemic biases can saturate the targets that make it to the end of the funnel, and can deprioritize good targets en masse.
- These distortions are visible to the trained eye by looking at the types of targets that rise to the top and what they have in common.
Mitigation:
- Involve functional experts throughout the process, not just at the end.
- Encode experts’ feedback with each iteration to improve the scoring mechanisms with every turn.
Target prioritization, transformed by AI
“Acceleration” is using AI to get through the existing process faster. “Transformation” means rethinking the process from the ground up to maximize the impact of AI on the speed of the process and quality of the outputs. In addition, an AI transformation should consider how expert insights and decisions are durably encoded in the process.
At their core, these prioritization workflows are ultimately driving toward decisions on where to spend the next dollar or hour of diligence: maybe it’s a biology or chemistry experiment, a virtual screen, a call with an expert, a proprietary dataset purchase.
But often, those activities are given short shrift in the early stages of the funnel, where the focus almost entirely on the target itself. Understandably so: it generally aligns with how we think as scientists: you start with a hypothesis (“drugging this target will modulate this disease”), then you pick the right set (and sequence) of tools to interrogate the hypothesis.
Not to mention that it substantially simplifies the analysis. Putting the experimental cart before the hypothesis horse would multiply the analytical effort by the multitude of possible actions/experiments one could run.
The problem with typical target-centric scoring approaches is that they can emphasize a target’s current value, and may be blind to those targets where a cheap action could substantially increase the value of the target.
Rather than a target-centric approach, discovery teams should consider an action-centric one instead.
The simplified framework is shown here:
More math in the appendix.
To be sure, we’re still scoring targets, but (a) we layer on top of each target additional scores for every action we could take for that target and (b) rank the top actions across the universe of targets. Furthermore, we need to factor in downstream decision gates, otherwise such a system tells you to spend all of your capital on cheap experiments that otherwise don’t advance a program to a real value inflection point.
This thinking also steers toward panels and screens where adding additional targets has a low incremental cost. For instance, I worked with a colleague to design a CRISPR panel across a set of genes using this approach.
The panel contained a mix of ‘long-shot’ targets, where the value of the experiment was derived from the potential upside, and ‘safer-bet’ targets, where the biology was better validated, and the value was as a quick-kill: avoiding higher downstream spend with a cheap experiment. In the end, by changing to an action-oriented approach, we were able to get a lot more bang for our experimental buck, rather than treating each target as its own isolated entity.
A key feature of any of this is that while a well-designed system will recommend top targets based on their value, the final decisions rest with the experts. They need to be able to deconstruct how the value of the experiment emerges from the system – especially when non-intuitive recommendations emerge.
This is where discovery organizations should be talking to their companies’ portfolio teams. They are versed in the field of decision sciences, which provides frameworks to quantify the value of new information, and assess tradeoffs across a portfolio of exploratory discovery activities.
There’s no escaping the need to make simplifying assumptions in this new regime, and by design, the relevant parameters balloon quickly, so it’s important to decide upfront how ambitious your first pass will be, which assumptions should be hard coded upfront vs left to the LLM’s judgment and reviewed downstream, and so on.
Having a solid framework design is the most critical place to invest time, and even with LLMs available to grind through the analyses, it’s best to pilot these efforts on a narrowly scoped question, and build out from there.
Finally, even if the workflow looks meaningfully different, it should land you in the same place. Your functional experts are still going to refine the coarse AI-enabled analysis and validate or refine the recommended actions. Your leadership team will still be presented with the cross functional team’s recommended path forward for investments.
But, if you’ve done things right, you’ll have arrived at a better answer, you’ll potentially have arrived there faster, and you’ll have built a reusable foundation for repeatable analyses moving forward.
The outlook
Target prioritization is just one of many places that I think decision sciences can have its moment across the biopharma value chain—not just for individual high stakes decisions, but where high-volume, and additively high-value decisions can be supported by the analytical speed of AI.
Combined with thoughtful and durable infrastructure investments which capture institutional knowledge, I believe AI can move beyond accelerating insights toward driving decisions—and even executing some of those decisions with meaningful autonomy.
Further reading
- Matt Shlosberg (COO, Deep Origin), writes about using AI to run quick kill experiments: “automated skepticism” is what he calls it.
- A team at a recent hackathon in London designed a tool to identify liabilities for given targets, and recommend high-impact experiments to resolve them.
- Derek Lowe wrote up a recent article in Nature Reviews Drug Discovery about the impact of AI on drug discovery and recommendations for the field. I’ll likely be sharing my own POV on this in my next post.
- An interview on Mike Rea’s blog from last September: Alex Telford (convoke.bio co-founder) shared a similar line of thinking about rethinking biopharma decision-making with AI.
EVOI detailed explanation
The full step-by-step derivation behind the two formulas above — worked through term by term, then plugged in with numbers:
Case study contrived by Alex, primarily assembled by Claude.