There’s a particular kind of meeting I’ve learned to recognize within the first five minutes. An Agentic-AI capability is being demonstrated. It’s genuinely impressive: the system takes a question in plain English, pulls from documents, reasons across a few steps, and produces an answer that reads like a competent colleague wrote it. The room is sold. A pilot gets funded. And then, months later, the same capability is still a pilot, not because it failed loudly, but because it’s wrong just often enough, and in ways subtle enough, that no one in a regulated function will sign their name under its output.

Nothing was wrong with the model. The model was the most capable part of the system. What failed was everything underneath it, the part nobody puts in the demo.

I want to make an argument that runs against the current of nearly every AI roadmap I see: the foundational work in AI is not the AI. It is the semantic and content layer beneath it. In life sciences especially, that layer is the single largest determinant of whether an AI program produces durable value or an expensive graveyard of pilots. And it is exactly the work that gets deferred, because it is slower, less photogenic, and harder to put in a board deck than an agent that talks.

The failure is not where the budget goes

Start with the numbers, because they are blunt. MIT’s Project NANDA, in its 2025 State of AI in Business study, found that 95% of enterprise generative-AI pilots produced no measurable return, a small fraction crossed into real P&L impact, and the rest stalled. RAND, in a study dedicated to why AI initiatives fail, found that more than 80% fail, roughly twice the failure rate of IT projects that don’t involve AI and traced a large share of those failures to a single root cause: organizations underestimate the data and knowledge groundwork AI requires, and discover the gap only after they’ve already committed. Gartner arrives at the same place from a different direction, reporting that most organizations don’t have or aren’t sure they have data management practices ready for AI, and predicting that through 2026 organizations will abandon 60% of AI projects that aren’t supported by AI-ready data.

Read those findings together and a pattern emerges that has almost nothing to do with model quality. The spending goes to models, platforms, and agents. The failures trace to meaning, structure, and connection. We are investing at the top of the stack and dying at the bottom of it.

AI amplifies meaning; it does not manufacture it

Here is the principle I’d ask you to hold onto, because everything else follows from it: an AI system does not create understanding. It amplifies whatever understanding or ambiguity, already exists in the content you point it at.

Give a capable model clean, well-defined, richly connected knowledge and it does things that genuinely feel like magic. Give that same model the reality of most enterprises, thousands of PDFs, inconsistent terminology, undocumented tables, decades of context that lives only in people’s heads, and it does not become a librarian who imposes order. It becomes a mirror. It reflects your ambiguity back at you, fluently, at machine speed, with machine confidence. The output is grammatically perfect and epistemically unreliable, which is the most dangerous combination there is, because it defeats the ordinary human instinct to distrust something that sounds unsure.

“Garbage in, garbage out was always true, what’s new is that the garbage now comes back articulate.”

The two layers that never make the slide

When I say “foundation,” I mean two specific layers that sit below any model or agent.

1. The first is the content layer, the unglamorous pipeline that turns raw source material into something a machine can use precisely. Ingestion from internal and external sources. Decomposition of documents into clean, structured units. Careful chunking. Extraction of tables and figures. And, non-negotiable in our industry, provenance on every piece, so the system can always answer “where did this come from?” A great deal of failed retrieval is born right here: if a document was chunked badly, the smartest model in the world will confidently retrieve the wrong passage and reason beautifully from it.

2. The second is the semantic layer, and this is the one almost everyone skips. It is the set of shared, machine-readable definitions and relationships that give content meaning: controlled vocabularies so terms are normalized; ontologies that define what a gene, a protein, a disease, a pathway, a phenotype, or an adverse event actually is and how it relates to everything else; and a knowledge representation, typically a knowledge graph, where those entities and relationships are stored so they can be reasoned over. Embeddings and vector search live near here too, but embeddings capture statistical similarity, not formal meaning. On their own they’ll cheerfully tell you that two things are “similar” without knowing that one causes the other. The semantic layer is what supplies the causes, the inhibits, the is-a-type-of, the structure that turns a pile of similar documents into knowledge you can actually traverse.

The model and the agent sit on top of these two layers. They are the roof. We keep trying to install the roof before the foundation has cured, and then act surprised when the building won’t stand.

Life sciences proved this twenty-five years ago

Here is what should make this argument easy to accept in our field: we have already done this work once, deliberately, for human consumption. The Gene Ontology launched in 2000 to give biology a shared, controlled vocabulary across organisms. The OBO Foundry coordinated dozens of biomedical ontologies so they would interoperate rather than fragment. The UMLS integrated more than a hundred vocabularies into a common semantic fabric. SNOMED CT did it for clinical terminology. And the FAIR principles: Findable, Accessible, Interoperable, Reusable, exist because we recognized, as a field, that data without shared semantics and provenance is data you cannot safely combine.

We built all of that so scientists and systems could integrate evidence across sources without guessing what a term meant. An autonomous agent is precisely such a “system”, one with far less common sense than a human expert and far more willingness to paper over a gap with a confident fabrication. It needs that semantic backbone more than we ever did, not less. The irony of this moment is watching organizations that helped build these standards now bypass them in the rush to put agents on top of raw, un-modeled content.

What actually breaks without the foundation

Remove the foundation and the failures are specific and predictable.

Retrieval degrades from correct to merely plausible. Without entity normalization, the system doesn’t know that “NSCLC” and “non-small-cell lung cancer” are the same thing, or that two papers describe the same target under different names. It retrieves something that looks right, and reasons from it with total confidence.

Multi-hop reasoning becomes impossible. The questions that create real value in research are connective: which of our programs share a pathway with this target and also carry a known safety signal? Which biomarkers link this patient subgroup to a mechanism we already understand? Those are graph traversals, they require following relationships across many entities. A knowledge graph makes them answerable; this is exactly what was demonstrated when Himmelstein and colleagues integrated biomedical knowledge into a graph and used it to systematically prioritize drugs for repurposing. Without that connected structure, an agent can summarize a single document but cannot take the second hop.

Hallucination stops looking like a model bug and reveals itself as a meaning problem. The most reliable mitigation we have is grounding, retrieval-augmented generation, introduced by Lewis and colleagues in 2020, constrains a model to cite real evidence instead of its parametric memory. But retrieval is only as good as the content and semantics it retrieves from. Ground a model in an ontology-backed knowledge graph and its answers become traceable to named concepts and relationships. Ground it in an undifferentiated document dump and you have simply automated plausible-sounding guesswork.

And there is the failure that ends programs in our industry: the absence of explainable lineage. In a GxP, 21 CFR Part 11, ALCOA+ world, an output must be attributable, traceable, and defensible. “The model produced it” is not an answer an auditor accepts. A system grounded in represented knowledge can show its work, these concepts, these relationships, these sources. A system reasoning over un-modeled content cannot, and so it never leaves the sandbox, however good the demo was.

The stack, in order

If there is one thing to carry out of this, it is the sequence:

Data → content layer → semantic layer (ontologies + knowledge graph) → models → agents and inference.

Agentic AI lives at the very top. It is an escalation, something you reach for when a problem genuinely needs autonomous planning and tool use over a foundation that can support it, not a default starting point chosen because it’s the most interesting layer to work on. Reverse the order, as most programs do, and you are not building intelligence. You are building a very expensive, very articulate way to launder ambiguity into decisions that people will eventually act on.

“But we can’t boil the ocean”

This is the honest objection, and it deserves an honest answer. No, you cannot ontologize your entire enterprise before doing anything useful, and you should not try. A semantic layer built in a vacuum, disconnected from a real use case, is its own well-documented way to fail.

The answer is to model the slice your single highest-value use case requires, anchored to that use case, and ship it, then let it compound. Because here is the property that makes the discipline worth it: the semantic layer is one of the few assets in this entire field that becomes cheaper and more valuable with every use case built on top of it. The concepts and relationships you define for target discovery get reused by the safety-signal agent, which get reused by the regulatory-authoring assistant. Each ungrounded pilot, by contrast, pays the ambiguity tax again from zero and leaves nothing behind for the next one. One path bends the cost curve down over time; the other bends it up. The teams that look slower in year one are the ones still standing in year three, which is, not coincidentally, about the horizon over which Gartner finds that AI-mature organizations actually keep their AI in production.

Meaning is the moat

Models are converging and commoditizing, the frontier labs are months apart, and the gap to capable open models keeps shrinking. The agent frameworks are converging too. What is not commoditized, what no vendor can sell you and no model release can hand you, is a faithful, governed, machine-readable representation of your own science and your own decisions. That is the durable advantage. That is the moat.

So the right question was never “which agents should we build?” It is the older, quieter one: is our knowledge modeled well enough that anything, human or machine, can reason over it and be trusted?

AI READY ARCHITECTURE

In summary,

Build the foundation that makes the answer yes.

Do it in that order.

The agents will be waiting, and they will finally work.

Opinions my own.

Connect with me on LinkedIn, https://www.linkedin.com/in/bharadwajpopuri/

or

https://bharadwajpopuri.com/

References

1. MIT Project NANDA. The GenAI Divide: State of AI in Business 2025. (Coverage: Fortune, Aug 18, 2025.) https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

2. RAND Corporation (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed.https://www.rand.org/pubs/research_reports/RRA2680-1.html

3. Gartner (Feb 26, 2025). Lack of AI-Ready Data Puts AI Projects at Risk. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk

4. Gartner (Jun 30, 2025). 45% of Organizations With High AI Maturity Keep AI Projects Operational for at Least Three Years.https://www.gartner.com/en/newsroom/press-releases/2025-06-30-gartner-survey-finds-forty-five-percent-of-organizations-with-high-artificial-intelligence-maturity-keep-artificial-intelligence-projects-operational-for-at-least-three-years

5. Ashburner M, et al. (2000). Gene Ontology: tool for the unification of biology. Nature Genetics 25:25–29. https://www.nature.com/articles/ng0500_25

6. Smith B, et al. (2007). The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nature Biotechnology 25(11):1251–1255. https://www.nature.com/articles/nbt1346

7. Bodenreider O (2004). The Unified Medical Language System (UMLS): integrating biomedical terminology. Nucleic Acids Research 32:D267–D270. https://academic.oup.com/nar/article/32/suppl_1/D267/2505235

8. Wilkinson MD, et al. (2016). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 3:160018. https://www.nature.com/articles/sdata201618

9. Lewis P, et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 33. https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html

10. Himmelstein DS, et al. (2017). Systematic integration of biomedical knowledge prioritizes drugs for repurposing. eLife 6:e26726. https://elifesciences.org/articles/26726

11. Nicholson DN, Greene CS (2020). Constructing knowledge graphs and their biomedical applications. Computational and Structural Biotechnology Journal 18:1414–1428. https://www.sciencedirect.com/science/article/pii/S2001037020302804

What it takes to build an LLM evaluation pipeline you can actually defend to an auditor and what I learned trying to break my own

There is a specific moment that motivated everything I am about to describe. Picture a pharmacovigilance scientist at a large drug company. Her job, in part, is to write adverse-event narratives: structured clinical accounts of what happened when a patient had a bad reaction to a medication. These narratives feed regulatory submissions. They are read by health authorities. They are, in the most literal sense, safety-critical documents.

Now hand her a large language model. The model drafts a narrative in four seconds. It is fluent. It uses the right vocabulary — MedDRA preferred terms, ICH E2B fields, WHO-UMC causality language. It reads like something she would have written. She is, reasonably, both delighted and uneasy.

The unease is correct, and it is not about whether the model is *good*.

It is about everything that surrounds the question of good. Can she trust the evaluation that told her this draft was acceptable? Can she prove, months later, that the evaluation result was not changed? When the evaluation ran, did it quietly copy a patient’s name and date of birth into a logging system or a metrics dashboard? And if the model that *scored* the draft was offline that afternoon, did her pipeline tell her or did it hand back a confident, plausible, entirely fictional number?

I spent a concentrated stretch of work building a system to answer those questions, and then I spent a good deal more time trying to prove the answers were wrong. This is the story of both halves.

Why pharmaceutical work is genuinely different

It is tempting to treat “evaluating LLM output” as a solved or solvable problem with general-purpose tools. For the core quality question, that is nearly true, and I will come back to the excellent tooling that exists. But pharmaceutical R&D operates inside a regulatory wrapper that changes what “evaluation” has to mean.

Three frameworks matter here, and they are worth naming because they are not optional opinions; they are the rules the industry is audited against.

1. The first is 21 CFR Part 11, the U.S. Food and Drug Administration’s regulation governing electronic records and electronic signatures. In spirit, it requires that records be trustworthy, attributable, and protected from undetected alteration. If an evaluation result is a record that influences a regulated decision, then “we think it’s the same as when we computed it” is not an acceptable posture. You need to be able to demonstrate integrity.

2. The second is GAMP 5. the Good Automated Manufacturing Practice guidance for validated computerized systems. Its through-line is that systems doing consequential work must be validated, reproducible, and behave predictably, not occasionally, but as a documented property. A scoring system that returns different numbers on identical inputs, or that silently changes behavior when a dependency is missing, fails this on its face.

3. The third is ALCOA+, a set of data-integrity principles: data should be Attributable, Legible, Contemporaneous, Original, and Accurate, plus Complete, Consistent, Enduring, and Available.

   Read those words slowly with an AI evaluation pipeline in mind.

• Attributable, who or what produced this score, and under what configuration

• Original and Accurate, is this the real computed result, or a placeholder

• Enduring and Available, can you retrieve and verify it later?

And underneath all of it sits patient privacy. The inputs to a pharmaceutical evaluation; the prompts, the source documents, the reference answers, routinely contain protected health information. The outputs can too. An evaluation pipeline that logs raw inputs and outputs to a dashboard, a database, or an observability platform is, in this context, a data-handling problem wearing the costume of a developer convenience.

So the bar in Pharma is not “is the output good.” It is “is the output good, *and* can we prove the evaluation was intact, reproduce it exactly, and demonstrate that producing it did not expose patient data.” That compound requirement is the gap.

Where the generic tools stop

I want to be clear and fair, because this is a point where it would be easy to be unfair. The open-source LLM evaluation ecosystem is genuinely good. DeepEval gives you well-constructed metrics for relevancy, faithfulness, hallucination, bias, and toxicity, and a clean way to write custom criteria. RAGAs is purpose-built for retrieval-augmented systems, scoring faithfulness and context precision against retrieved evidence. Evidently brings deterministic text descriptors, LLM-as-judge scoring, and real statistical drift detection, Kolmogorov–Smirnov and chi-squared tests on production distributions. I did not replace any of these. I use all of them.

But every one of them answers the quality question and stops there, which is exactly correct for what they are. None of them signs its results so you can prove they were not altered. None of them treats de-identification as a precondition of scoring. None of them, out of the box, guarantees that a missing backend produces a visible failure rather than an invisible fabrication. They are evaluators, not systems of record. The regulatory wrapper is not their job but in pharma, it is *someone’s* job, and that someone has historically had to build it themselves, ad hoc, around tools that were never designed for it.

That is the space PharmaEval occupies: not a better metric, but the trustworthy envelope the metrics need in order to count as evidence.

The architecture: one choke point, four guarantees

The central design decision is almost embarrassingly simple, and I think its simplicity is the point. There is exactly one path through which an evaluation can happen, and that path enforces the guarantees in a fixed order. I think of it as a choke point: a single function that every evaluation must pass through, so the guarantees cannot be accidentally bypassed by a future code path or a well-meaning shortcut.

The order is: de-identify, then score, then sign, then persist.

First, the input, the model output, the reference text, and the context are all run through the PHI scrubber. Then and only then the *redacted* text is scored. The scores, the de-identification summary, and cryptographic hashes of the redacted text are assembled into an audit record, which is HMAC-signed. Finally, a de-identified projection is written to the queryable store, and the signed record is appended to an immutable journal. Reviewer decisions follow the identical shape: the comment is de-identified, a signed decision record is appended, and the projection’s status updates.

Because this is the only way through, the properties hold by construction rather than by discipline. You cannot score raw text, because the scorer is downstream of the scrubber. You cannot persist an unsigned record, because the store refuses one. You cannot quietly mutate the journal, because there is no update or delete path and every record is signed. The guarantees are structural, not procedural which is what GAMP 5 actually wants.

Let me take each pillar in turn, including where each one falls short, because the shortfalls are as important as the strengths.

Pillar one: de-identification, and the honest shape of its limits

The de-identifier is deliberately model-free. It detects structured personal identifiers using regular expressions backed by checksum validators, Social Security numbers validated against their real area and group rules, credit cards against the Luhn algorithm, National Provider Identifiers against their check digit, IBANs against mod-97. It recognizes emails, phone numbers in several formats, dates, medical record numbers, ZIP codes, IP addresses, and account numbers. When it finds something, it can mask it (`[SSN]`), replace it with a format-preserving fake (a fake SSN that keeps the `###-##-####` shape, generated deterministically so the same input always yields the same redaction), or hash it to a stable token. Re-identification is possible only when explicitly enabled, and otherwise no reversal data is retained at all.

I chose model-free detection on purpose. It is reproducible, it requires no GPU, it runs in an air-gapped validation environment, and its behavior is auditable line by line, all of which matter under GAMP 5. But that choice has a hard consequence, and here is where I have to be honest in a way the field too often is not.

Regular expressions cannot recognize names.

There is no pattern for “this string is a person’s name.” When I tested the scrubber against `Patient John Smith, treated by Dr. Sarah Johnson, recovered well`, it detected zero identifiers and passed both names through to storage verbatim. The single most common HIPAA identifier the name is exactly the thing my deterministic detector is structurally incapable of catching. The same goes for street addresses and for non-standard formats of things it otherwise handles: a Social Security number written as nine bare digits with no delimiters, an ISO timestamp with a time component, certain written-out dates.

I did two things about this, and the second is the one I care about.

The fix: I added a name-denylist hook. A deploying team can supply an explicit list of known sensitive terms — patient names, clinician names, study aliases — that are always redacted, case-insensitively, at every occurrence, ahead of the pattern matches. Real names live in an environment variable, never committed to the repository. With the denylist supplied, the same sentence redacts cleanly to `Patient [NAME], treated by Dr. [NAME], recovered well`, end to end, through the live API.

The honesty: I wrote an explicit limitations section into the documentation and the user interface stating plainly that this is a *structured-identifier scrubber, not a certified de-identifier*; that names require the denylist or a machine-learning layer; that the result is a *reduction* of PHI exposure, not a guarantee of its elimination; that it is not equivalent to HIPAA Safe Harbor or Expert Determination; and that a detection count of zero means “no in-scope identifiers were found,” not “no PHI is present.” The original documentation had claimed, in plain language, that “raw PHI never reaches the database.” That claim was false for names, and I removed it.

I will say more about why that removal matters later, because I think it is the real point of the whole project. For now: the de-identification pillar is genuinely useful and genuinely bounded, and the system tells you exactly where the boundary is.

Pillar two: cryptographic provenance

This is the pillar I am most confident in, because I attacked it hardest and it held.

Every evaluation and every review decision becomes an `AuditRecord`. The record carries the event type, the actor, a timestamp, the non-PHI payload (scores, thresholds, de-identification summary), and SHA-256 hashes of the redacted input and output. On construction it computes a reproducility hash, a hash over a canonical JSON serialization of its contents, stable regardless of key ordering — so that any later mutation of any field is detectable. Then it is signed with HMAC-SHA256 using a key that lives only in the environment, never in the code.

Verification re-derives the signature and the reproducibility hash and compares the signature in constant time. The properties this produces are not aspirational; I tested them adversarially. I mutated the payload and recomputed the reproducibility hash to make the record internally consistent, verification still failed, because the attacker does not hold the signing key. I swapped a valid signature from one record onto a different record, failed. I changed the hash that binds the record to its redacted text, failed. I forged a signature with the wrong key, failed. I passed an empty or null key , the system raises rather than silently accepting, which is the correct fail-closed behavior. Crucially, the reviewer’s comment, which is stored inside the record’s metadata, is covered by the signature, so you cannot quietly rewrite a reviewer’s words after the fact.

The journal that holds these records is append-only by construction: the store exposes no update and no delete method, refuses to write an unsigned record, and assigns each record a monotonic sequence number. A verification endpoint re-checks every record in an evaluation’s trail and returns an explicit integrity error, a distinct HTTP status, if anything fails. When I tampered with a stored record directly in the database and then called that endpoint, it correctly reported the trail as unverified.

This is what provenance for AI evaluation should look like. Not “we logged it,” but “here is a signed, reproducible, tamper-evident record, and here is the mechanism by which you or an auditor can verify it independently.”

Pillar three: the refusal to fabricate

This pillar addresses a failure mode that is easy to miss precisely because it is silent, and I have come to think it is one of the more important ideas in the whole system.

Evaluation code frequently has a fallback path. The intended scoring backend, an LLM judge, a metric library, is unavailable or misconfigured. Rather than failing, the code returns a default: a hardcoded number, a simulated score, a placeholder that keeps the pipeline running. The dashboard fills with green. The batch job completes. And the scores are fiction. In a casual setting that is a minor bug. In a regulated setting it is a data-integrity catastrophe, because you now have records that look like evaluations, carry the authority of evaluations, and were never actually computed. They violate the “Original” and “Accurate” of ALCOA+ silently and completely.

PharmaEval takes the opposite stance. Its deterministic metrics genuinely compute from the text, I will return to their crudeness in a moment, because they are crude. Its LLM-judge mode, when the backend is unavailable, *raises an explicit error* rather than returning a number. A missing score looks missing. There is no path by which the system produces a plausible result it did not actually derive.

Now the honesty, again. The deterministic metrics, which exist so the system can run reproducibly with no external credentials, are lexical heuristics. When I stress-tested them, I found that the string “drug drug drug drug” earns a perfect faithfulness score against a clinical reference, because the metric measures the fraction of output words that appear in the reference and every word matches. One of the regulatory-language metrics returns roughly the same value for almost any input, carrying little real signal. The metrics cannot detect negation, so a narrative asserting that a patient did *not* take a drug scores similarly to one asserting they did.

This matters, and it does not. It matters because nobody should mistake the deterministic scores for a substitute for an LLM judge or a human reviewer; they are a deterministic smoke-test, appropriate for a reproducible continuous-integration gate, and the documentation now says so. It does not undermine safety, because and I verified this directly, garbage does not auto-approve. The composite score dilutes the gameable individual metrics, and the auto-approval bar is conservative enough that borderline and nonsensical inputs route to human review rather than passing automatically. The system fails safe. But the metrics are honestly labeled as what they are: crude, deterministic, and not the final word.

Pillar four: rigor that fits the task, and a human in the loop

A pharmacovigilance case and a regulatory-affairs response are different kinds of documents with different failure modes, and they should not be held to a single undifferentiated bar. PharmaEval ties thresholds to named use cases. A pharmacovigilance preprocessing case applies a stricter hallucination threshold and a higher auto-approval bar, weighted toward adverse-event narrative quality and causality assessment. A health-authority response case weights faithfulness most heavily and demands the highest auto-approval bar of all. A document-authoring case foregrounds ICH E3 structural compliance. These are not cosmetic labels; they resolve into the actual thresholds and routing logic the evaluation applies.

And nothing consequential happens fully automatically. Anything that does not clearly clear its bar lands in a human review queue, where a subject-matter expert approves or rejects it, and that decision, too, is de-identified and signed into the audit trail. The human is not a rubber stamp bolted on at the end; they are a first-class participant whose judgments carry the same provenance guarantees as the machine’s.

The adversarial review, and why I ran it on myself

After the system worked, I did something I think more builders should do: I tried to break it, on purpose, from a clean environment, treating my own prior claims as suspect rather than settled.

That review is where the name leak surfaced. It is where the gameable metrics surfaced. It is where I found a misleading comment in the database layer claiming that an in-process lock made concurrent multi-worker writes safe, when in fact that lock only coordinates threads within a single process and cannot coordinate across the separate worker processes a production deployment runs. (I fixed that too, enabling write-ahead logging and an explicit busy timeout so concurrent writers wait out contention instead of erroring, and I corrected the comment to describe reality.)

I want to dwell on the posture, not just the findings. It would have been easy, and common, to run the test suite, see it pass, and declare victory. A passing test suite is the thing you should trust *least*, because tests encode what you already thought to check. The adversarial review checked the things I had a stake in not finding. The signing held. The auth held, every bypass attempt I tried returned an authentication failure. The production configuration correctly refused to start without its secrets. The concurrency held under two hundred simultaneous writes. But the de-identification and the metrics did not fully hold, and the only responsible thing to do was to say so, in the product, where a deploying team would see it before making a decision.

What I think the contribution actually is

Let me be careful and unglamorous about this, because the temptation to inflate is exactly the thing I am arguing against.

I did not invent a new evaluation metric. I did not produce a research result. What I built is a practical, working, tested pattern. a reference implementation that demonstrates something I believe the field needs to internalize:

LLM evaluation in regulated settings can and should carry the same integrity guarantees we already demand of every other system of record.

Concretely, the contribution is a small set of primitives, composed into one enforced pipeline:

Provenance, by signing every evaluation and review decision into a reproducible, tamper-evident, append-only record that anyone can verify independently. PHI-safety, by making de-identification a structural precondition of scoring and persistence rather than a logging afterthought. Anti-fabrication, by guaranteeing that a missing or broken scoring backend produces a visible failure instead of an invisible fiction. Use-case-fit, by binding rigor to the regulatory task. And, the one I have come to value most, honest disclosure, by shipping with an adversarial review and an explicit limitations section that tells a deploying team exactly what the system cannot do.

None of these primitives is pharma-specific. They generalize to any domain where AI output carries consequence and accountability, legal drafting, financial analysis, clinical decision support, public-sector adjudication. Pharma simply has the regulatory clarity to make the requirements legible, which makes it a good proving ground. The pattern is portable; the regulations just supply the vocabulary.

If there is a single idea I would want to leave behind, it is the last primitive. We are in a moment of extraordinary overclaiming about what AI systems can do, and the overclaiming is most dangerous precisely where the stakes are highest, because that is where someone downstream relies on the claim. A system that hides its blind spots cannot be reasoned about by the people who have to take responsibility for it. An honest limitations section is not an admission of failure. In a regulated context, it is a precondition of use, and I think, increasingly, it should be a precondition everywhere.

Limitations, and what comes next

To keep my own counsel, here is where this work genuinely ends. The de-identifier needs a machine-learning or named-entity layer to catch names and free-text identifiers reliably; the denylist is a bridge, not a destination. The deterministic metrics are a smoke-test, not a substitute for an LLM judge or expert review, and a serious deployment would lean on the LLM-judge path and human reviewers for anything that matters. The file-based default datastore is appropriate for modest scale; sustained multi-writer load wants a managed relational database. And none of this is a substitute for formal computer-system validation, which is an organizational process, not a property of code.

What comes next, if I carry it further, is the obvious set: an optional NER de-identification stage behind the same interface, richer metrics that detect negation and contradiction rather than counting word overlap, key rotation for the signing layer, and a managed-database backend. But the foundation, the choke point, the signing, the refusal to fabricate, the honesty, is the part I wanted to get right first, because everything else is an improvement on a trustworthy base rather than a patch on an untrustworthy one.

We are going to keep handing consequential drafting to language models. The narrative that pharmacovigilance scientist is reviewing will increasingly have been drafted by a machine, and so will the next one, and the one after that. The question was never whether we would evaluate that work. It is whether we can *defend* the evaluation, prove it intact, reproduce it exactly, and show it did not expose the patient at its center. I set out to show that we can, and to be precise and honest about how far that goes. That precision, I think, is the whole job.

PharmaEval is a reference implementation and engineering pattern. I’d welcome conversation with anyone working on AI governance, data integrity, or evaluation in regulated environments. The framework is open source .

I built it to be readable, a reference for how I think these pipelines should be structured, not a product. If you’re working on LLM deployment in a regulated context and any of this resonates, I’d genuinely like to compare notes.

PHARMA LLM EVAL DASHBOARD

Opinions my own.

If you found this useful, I write about Agentic AI architecture, evaluation, and governance in regulated enterprise.

Connect with me here:

https://bharadwajpopuri.com/

Read my previous articles on my Substack:

Most LLM evaluation demos I see score one thing: output quality. They run a hallucination check, a faithfulness score, maybe a RAG retrieval metric, and call it done. That’s the easy 80%. The hard 20% is everything that determines whether you can actually deploy the thing in a regulated environment and that 20% is where projects stall.

I spent the last few weeks building an open-source evaluation framework to work through that hard part end to end. Not the scoring; the scoring is a solved problem with DeepEval, RAGAs, and Evidently. The part nobody demos: what happens to a borderline result, whether patient data leaked to your judge model, what the evaluation actually cost, and how any of it survives an audit.

Here is what I learned building it, and why I think the industry is optimizing the wrong layer.

The score is the input to a decision, not the decision

A composite quality score of 0.78 is not an answer. It’s a question: who decides whether this output ships?

In a regulated setting, the honest answer for most outputs is “a human expert.” So the framework routes on score; high-confidence results auto-pass, low-confidence results auto-fail, and everything in the ambiguous middle goes to a review queue where a subject-matter expert approves or rejects it with a mandatory comment. That comment, the reviewer identity, and a UTC timestamp all land in an audit trail.

Human-in-the-loop is usually treated as a demo feature, a nice screenshot for the steering committee. I’d argue it’s the opposite. HITL is the governance mechanism. The criteria that decide when a human gets pulled in are the actual compliance control. The score just feeds it. If you build the scoring and bolt on review later, you’ve built the easy part and skipped the part regulators care about.

If you send patient data to your judge, you’ve already failed

Here’s a failure mode I almost never see addressed: the evaluation pipeline itself can be the data breach.

Real regulatory data: adverse event narratives, clinical notes, drug labels, frequently contains patient identifiers. Your evaluation harness takes that text and sends it to an LLM judge to be scored. If that judge is an external API, you just transmitted protected health information to a third party, inside the tool that was supposed to be enforcing quality.

So the framework de-identifies inputs before they reach any judge, using a local on-device model, no outbound call. It also runs a second check in the other direction: scanning the LLM’s output for leaked identifiers, scored as a first-class evaluation metric with a zero-tolerance gate. A model that hallucinates a realistic-looking patient name into a generated narrative will pass every quality metric and still be a data-integrity violation. You only catch that if you’re explicitly looking for it.

PII handling can’t be a preprocessing afterthought. In a regulated pipeline it’s a metric, a gate, and an audit record.

Inference cost is the smallest number on the invoice

The framework tracks the cost of every evaluation run and writes it into the audit manifest. This started as a convenience feature and turned into the most clarifying part of the build.

The token cost of a judge call is real but small. What dominates total cost is everything around it: the engineering to build the pipeline, the operations to keep it running, the iteration cycles, and in regulated industries specifically the validation overhead. When you account for all of it, the true cost per evaluated output is orders of magnitude above the inference line item. Naming that number changes how people make decisions. “It’s pennies per call” is true and irrelevant. The pennies aren’t the cost.

I made cost a first-class signal in the audit trail for the same reason I made PII one: if it isn’t measured and recorded per run, nobody can govern it.

Use the simplest architecture that works

A through-line in all of this: I deliberately kept the deployment lean. The reference web app boots in a deterministic demo mode with no heavy dependencies and no API keys, so it deploys to a free tier in minutes. The full evaluation stack with live judge models is one configuration change away, not a prerequisite.

That’s a principle, not a shortcut. The instinct in this space is to reach for the most sophisticated architecture available autonomous agents, complex orchestration, the works, when a far simpler pattern delivers the same value with a fraction of the operational and validation burden. In a regulated enterprise, the binding constraint is almost never infrastructure. It’s governance throughput: how fast a use case clears your approval bodies. A simpler architecture clears review faster. That speed is worth more than the sophistication you gave up.

The point

Evaluation frameworks that only score quality are solving the part that’s already solved. The unsolved part is operational: routing decisions to humans, keeping sensitive data out of the judge, measuring true cost, and producing an audit trail that holds up. Those aren’t features you add at the end. They’re the architecture.

The framework is open source . I built it to be readable, a reference for how I think these pipelines should be structured, not a product. If you’re working on LLM deployment in a regulated context and any of this resonates, I’d genuinely like to compare notes.

PHARMA LLM EVAL DASHBOARD

Opinions my own.

If you found this useful, I write about Agentic AI architecture, evaluation, and governance in regulated enterprise. The framework, the evaluation notebook, and the human-in-the-loop reference app are all available.

The right one is older and less exciting: what is the simplest thing that reliably, safely, and auditable solves this problem?

AgenticAI is an escalation pattern, not a default starting point. It earns its place when a problem genuinely needs autonomous planning, tool use, role separation, or coordination that simpler designs can’t deliver. Most problems don’t. And the ones that do still have to survive a cost question that rarely gets asked until the bill arrives.

So I built a small tool that enforces the order of operations I think every team should follow before committing to an agent. It walks a problem through three gates, in sequence, and won’t let you skip ahead.

You can try it here:

AgenticAI Decision Wizard

Here’s the thinking behind each gate.

Gate 1: Should we automate this at all?

Before AI enters the conversation, decide whether the problem is worth solving with software. Volume, hands-on effort, error stakes, the value of the outcome, and critically, how stable the underlying process is.

The trap is automating a process nobody has actually defined. You can’t reliably automate something you can’t describe; you just bake the variability into code. An undefined process gets stabilized first, not automated. And a low-volume, low-stakes task often shouldn’t be automated at all, however tempting the technology.

Most ideas that fail here fail quietly; which is exactly why an explicit gate matters.

Gate 2: Should it be Agentic?

This is where discipline pays off. The instinct is to reach for the most advanced architecture available. The better instinct is to climb a ladder of complexity only as far as the problem forces you to: deterministic logic, a single model call, retrieval-augmented generation, a fixed workflow with model steps, a single agent with tools, and only then multiple agents.

You cross into agentic territory when the task needs to plan its own steps at runtime, take actions across systems, or handle genuinely open-ended inputs. If the path is known and linear, a workflow is more reliable, more auditable, and far cheaper. If the core need is finding and grounding information, that’s retrieval, not an agent.

When a problem does clear that bar, a second question decides single agent versus many: do you actually need separate roles, separate permissions, independent review, parallel work, or multiple domains reconciled? If you don’t, one agent with tools is the answer and adding more agents simply multiplies latency, failure modes, and attack surface for no benefit.

Multiple agents are justified by real separation, not by ambition.

Gate 3: What does it cost to own?

This is the gate where the room goes quiet.

Inference is cheap. Everything else is not. The per-token cost of a model call is a small line in a multi-year bill that also includes build and integration, evaluation and testing, orchestration and hosting, monitoring, human-in-the-loop review, maintenance, security, and version upgrades. In a regulated environment, validation and audit requirements add a multiplier on top of all of it.

Put those together and the true cost per run is routinely an order of magnitude, sometimes far more, above the raw inference cost, especially at pilot scale, where the build hasn’t been amortized across enough volume to disappear. None of that shows up in a pricing-page demo.

The tool models it explicitly and compares the result against the manual baseline you’d be replacing, so you can see whether the economics actually work before you commit.

The point isn’t to say no to agents. It’s to say yes for the right reasons, in the right order.

Disciplined architecture decisions aren’t anti-innovation, they are what makes innovation survivable at scale. The teams that win with AgenticAI won’t be the ones that adopted it earliest. They’ll be the ones who knew which problems deserved it, and what it would really cost to keep alive.

The right question was never “can we build an agent?” It’s “what is the simplest architecture that reliably solves the problem, and what will it actually cost to own?”

Decide in that order. The tool that walks through it is here: https://personalstudent9.github.io/agentic-wizard/

Opinions my own.

• CMS + ONC just issued a new Request-for-Information (RFI) that lays out dozens of questions on how to modernise every layer of Medicare’s digital plumbing — claims, clinical data, patient-facing apps, AI decision-support, and enforcement of existing interoperability rules. STATCMS

• The effort is being steered by new tech leadership drawn from Palantir and Main Street Health. Palantir’s former analytics leads are already in senior HHS/CMS IT roles, and they have a well-publicised playbook for rapidly integrating siloed data and adding AI on top of it. OrangeSlices AI

What the agenda means in practical terms

Near-term implications for someone working in pharma

1. Accelerated real-world-evidence cycles
   Medicare claims + EHR data will hit analytical platforms days, not months, after a service. Safety-signal detection, label-expansion studies, and HEOR projects will need to run on shorter timelines.

2. Higher bar for demonstrating “value”
   When Palantir-style dashboards can compare drug outcomes across millions of beneficiaries, CMS can demand outcomes-based contracts (OBCs) with far finer performance metrics. Be ready to negotiate on real-world endpoints, not proxy measures.

3. Digital companion strategy becomes table stakes
   If CMS curates an “app formulary,” drugs that arrive with FHIR-enabled adherence, PRO-capture, or titration support tools will stand out. Align your digital-health BD roadmap with FHIR R4 APIs and CMS privacy/Security profiles.

4. Compliance & privacy exposure
   The same analytics that find fraud will spotlight questionable hub-services, referral patterns, or aggressive co-pay assistance. Audit internal data-flows now; ensure they meet HIPAA + new FHIR bulk-export safeguards.

5. Opportunities to co-create with CMS
   The RFI explicitly invites industry ideas. Submitting comments (deadline likely mid-July) can shape rules on AI explainability, patient-generated data, and coverage-with-evidence-development. Coordinate responses across RWE, regulatory-policy, and market-access teams.

Bottom line

CMS is about to operate much more like a data platform than a paper claims payer.
For pharma, that raises the bar on evidence generation and transparency but also creates a much richer data environment to prove value, target the right patients, and partner on preventive or outcomes-based models. Teams that invest early in interoperable data standards and agile analytics will turn this regulatory push into competitive advantage.

Sources:

https://www.federalregister.gov/documents/2025/05/16/2025-08701/request-for-information-health-technology-ecosystem

https://www.statnews.com/2025/05/30/medicare-ambitious-tech-agenda-former-palantir-main-street-executives/

https://www.federalregister.gov/documents/2025/05/16/2025-08701/request-for-information-health-technology-ecosystem

https://orangeslices.ai/hhs-it-office-recruits-new-chief-from-palantir/

Share

Subscribe now

For that, consult a licensed expert—or maybe your clairvoyant aunt—but definitely not me.

Introduction

The biotech and pharmaceutical sectors have always been subject to ebb and flow. Over the past several years, we’ve witnessed a dramatic shift in market sentiment—from high-risk discovery platforms in vogue one day to a demand for late-stage, “derisked” product candidates the next. While some investors chase the newest buzzword, more seasoned observers know these fluctuations are hardly new. If anything, they’re part of a rhythm the industry has repeated for decades.[1]

In this article, we’ll explore the drivers behind the ongoing swings in biotech valuations, discuss how startup formation and consolidation contribute to a “supply-and-demand” dynamic, and highlight why understanding these cycles is crucial for anyone investing in or operating a biotech company.

1. A Sector Shaped by Cycles

From the “genomics bubble” of the early 2000s to the more recent mega-round frenzy for scientific “platforms,” biotech has always been cyclical.[2][3] In risk-off periods, capital tends to chase clinical-stage (or near-clinical) product candidates, addressing clearly validated targets. When markets open up, bigger ideas—novel mechanisms of action or new discovery platforms—suddenly garner enthusiastic checks.[4]

• Past Boom-and-Bust Periods

  • Genomics bubble (late 1990s–early 2000s): Fueled by the hype around reading the human genome, huge sums poured into early-stage platforms. When the bubble burst, focus shifted toward “spec pharma” (e.g., repurposed or reformulated older molecules) as the supposedly “safer” play.[2]

  • Great Financial Crisis (2009–2012): Here again, funding dried up, and few wanted to commit to uncertain early discovery. Meanwhile, large pharmas selectively in-licensed or acquired known quantities.[2][3] Once macro pressures eased, attention shifted toward riskier but high-reward innovation.

  • Recent Bull Run (2013–2021): After years of low interest rates, active IPO markets, and expansions in biotech venture financing, science-driven platforms became hot—sometimes too hot. A surge of startups launched on the promise of revolutionary gene editing, mRNA, next-gen chemistries, and more.[4][5]

Every wave ends, and from early 2021 onward, we’ve seen public and private markets become more guarded.[3] The pendulum has swung back toward “assets” with visible clinical endpoints.

Yet, Does dismissing early-stage, innovative R&D jeopardize the industry’s long-term vitality?

2. Valuation Through the Lens of Supply and Demand

Biotech is powered by ideas, people, and capital. Each of these has a direct bearing on how new companies get valued:

1. Supply of Startups

   • When the market floods with freshly funded platforms, many compete for the same limited set of managers, clinical investigators, and patient populations for trials. This influx can drive valuations up in the short run, but it also leads to a “startup glut,” especially if many of these young companies focus on similar or incremental approaches.[6]

   • Conversely, fewer new ventures launch in tougher times, which can temper competition for talent and clinical resources.[7] A smaller pipeline of new companies (and well-managed existing ones) may later find themselves in high demand if innovation rebounds.

2. Investor Appetite (Demand)

   • “Risk-on” describes periods when public and private markets willingly fund ambitious science.[8] Biotechs without immediate clinical validation can still score large rounds or high IPO valuations.

   • “Risk-off” cycles push investors to safer territory—molecules with Phase 2 or Phase 3 data, validated biology, and straightforward commercial paths. While these strategies can yield near-term value, the pace of truly transformative breakthroughs can dwindle if the early-science pipeline starves.

3. Exit Environment

   • IPOs: During bull markets, a fluid IPO window significantly lifts valuations, as biotech companies can access large capital pools—sometimes earlier than their pipelines might merit.[3][9] Once that window narrows, private rounds become harder to close, and valuations come under pressure.

   • M&A: Large pharma remains hungry for new products, particularly in areas with validated mechanisms. When capital is scarce, pharma can often acquire late-stage assets at a discount, influencing how private investors underwrite future deals.

3. The Case for Balanced Portfolios

Long-term biotech investors often strive for a mix of asset-centric and platform-oriented companies.[2][4] Each model has distinct strengths:

• Asset-Centric Approach

  • Pros: Clear near-term milestones, often simpler narratives for investors, can position for faster, risk-reduced exits.

  • Cons: More prone to “binary events.” Failure of a single lead program can endanger the whole company.

• Platform Model

  • Pros: Potential to discover multiple drug candidates, spawn broad pipelines, and attract partnership deals that offset R&D costs.

  • Cons: Early platforms require time to generate leads, are highly technical, and may be perceived as riskier by the market in times of high interest rates or risk aversion.

Over the decades, we’ve repeatedly seen “what’s in vogue” flips.[4] At one moment, a deep, innovative platform is the golden ticket to blockbuster returns. Conversely, that platform can’t secure a bridging round without pivoting to a late-stage in-licensed product.

4. Scarcity as a Strength: Fewer Startups, More Focus

Interestingly, several data sets from Pitchbook and others suggest a decline in newly formed biotech startups, hitting multi-year lows in the past two years.[7] Paradoxically, this might be good for valuations in the medium term:

• Less Competition for Resources

  • With fewer concurrent preclinical or Phase 1 programs, teams face less crowding at investigator sites. Recruiting key talent also becomes more feasible without inflated salary demands.

• Quality Over Quantity

  • A leaner startup ecosystem could mean deeper vetting of scientific reproducibility and clinical feasibility. Investors will fund the best ideas rather than every big idea.

• Future Upside

  • By 2027–2030, today’s newly formed (presumably more carefully curated) startups could find themselves in a healthier exit environment. A smaller roster of serious contenders with robust pipelines generally commands higher valuations if broader market sentiment improves.[6][7]

5. R&D Efficiency and Capital Discipline

Regardless of where you fall on the asset vs. platform debate, today’s environment demands operational excellence:

• Tranche Your Spending

  • Smaller, carefully timed raises can limit dilution and ensure a company hits clear milestones before scaling.[2][3] Rushing to grab $200+ million without near-term data may lead to poor “governance by abundance.”

• Outsource vs. In-House

  • Fixed infrastructure is expensive and becomes a liability in lean times. Strategically using CROs (contract research organizations) and external manufacturing, when possible, keeps the burn rate manageable.

• Partner Strategically

  • Collaborations with Big Pharma or larger biotech peers can help offset R&D expenses and validate your approach.[4] In risk-off periods, pharma’s interest might spike if you have a compelling, partially derisked program.

6. Looking Ahead

While no one can predict the market’s next sharp turn, some lessons recur:

1. Innovative Science Eventually Finds a Home: Even when capital is hesitant, breakthroughs that convincingly address unmet medical needs attract partners.[2]

2. Focus on Data, Not Hype: Overblown claims about “transformative” technology can do more harm than good if they fail basic validation. Reproducible experiments and early clinical successes speak volumes.[1][10]

3. Balance Is Key: A diversified approach—both at the portfolio level for investors and the pipeline level for companies—helps weather cyclical downturns and capture upside when enthusiasm returns.[2][4]

The BioPharma sector’s future may look choppy, but history suggests that disciplined programs and thoughtful valuations can pave a sustainable path. For those willing to look beyond immediate sentiment swings, there remains a wealth of opportunity in developing the next generation of therapies that truly matter to patients.[1][4]

Want to Stay Ahead of the Biotech Curve?

Subscribe to our Substack for more insights into biotech valuation, R&D trends, and financing strategies. We’ll keep you updated on the market’s latest swings—and how to position yourself for both the short term and the long haul.

Click the button below to subscribe and never miss an update.

References

1. A brash biotech VC speaks out on headwinds in the industry and what the future may hold. Excerpts adapted from an interview with Omega Funds founder Otello Stampacchia, 2025.

2. Biotech Risk Cycles: Assets and Platforms. Posted October 28, 2024, in Biotech financing, Biotech investment themes, Capital efficiency, Capital markets, Exits, IPOs, M&As.

3. Biotech Funding: Times Are Tough, Maybe For The Better. Posted April 14, 2023, in Biotech financing.

4. Startups, Exits, And Ecosystem Flux: Bullish For Biotech. Posted September 8, 2014, in Biotech financing, Biotech startup advice, Exits IPOs M&As, General Venture Capital, VC-backed Biotech Returns.

5. Biotech Venture Creation: The Benefits Of Scarcity. Posted April 8, 2025, in Biotech financing, Capital markets, Exits IPOs M&As.

6. The Biotech Startup Contraction Continues… And That’s A Good Thing. Posted April 26, 2024, in Biotech financing, Biotech investment themes, Capital markets, and Fundraising.

7. Pitchbook Data. Various Q1 2023–Q1 2025 Funding and Startup Formation Reports. Retrieved from https://pitchbook.com.

8. NVCA (National Venture Capital Association) & PwC MoneyTree. Historical data on venture funding, 2010–2023. Retrieved from https://nvca.org/research and https://www.pwc.com/us/en/industries/technology/moneytree.html.

9. TD Cowen Analysis. Market dynamics in biotech crossover and IPO participation, 2022–2024.

10. Foundings Matter: Thiel’s Law Applied To Biotech. Posted June 11, 2013, in Biotech financing, Biotech startup advice.

For that, consult a licensed expert—or maybe your clairvoyant aunt—but definitely not me.

This week in life sciences, we witnessed significant developments across the industry. From groundbreaking FDA approvals to substantial biotech funding rounds and remarkable R&D breakthroughs, the sector continues to advance at a rapid pace.

Recent FDA Approvals

The FDA has recently approved several novel drugs, offering new treatment options for patients. Notably, "Darroway" received approval for treating specific types of breast cancer, providing hope for patients with this condition. Additionally, "Grapafex" was approved for use in combination with fludarabine as a preparative regimen for stem cell transplantation in acute myeloid leukemia. These approvals represent significant progress in addressing unmet medical needs.

Biotech Funding Highlights

In the world of biotech funding, "Latigo Biotherapeutics" secured a substantial $150 million in a Series B funding round. The company is focused on developing non-opioid pain treatments, a critical area of research given the ongoing opioid crisis. This funding will enable Latigo to advance its pipeline and bring innovative pain management solutions to the market.

R&D Breakthroughs

The R&D landscape continues to evolve with groundbreaking discoveries. One notable breakthrough is the clinical validation of CRISPR technology, which holds promise for treating previously undruggable diseases like cancer and neurodegenerative disorders. Additionally, AI-driven advancements like DeepMind's AlphaFold have revolutionized protein structure prediction, opening new avenues for drug discovery and development.

Latigo Biotherapeutics $150M Series B Funding (2025)

Latigo Biotherapeutics indeed closed a $150 million Series B financing round in March 2025. The company’s press release, dated March 17, 2025, announced the $150M Series B, led by Blue Owl Capital (Latigo Biotherapeutics Closes $150 Million in Series B Financing to Advance Non-Opioid Pain Therapeutics - Latigo Biotherapeutics). Proceeds from the round are focused on advancing Latigo’s pipeline of non-opioid pain therapeutics, particularly its highly selective Nav1.8 inhibitor programs in clinical development for pain, as well as expanding its broader drug pipeline (Latigo Biotherapeutics Closes $150 Million in Series B Financing to Advance Non-Opioid Pain Therapeutics - Latigo Biotherapeutics). In other words, the funding is earmarked to develop best-in-class non-opioid pain treatments that target pain at its source (Nav1.8 sodium channels) instead of using opioids (Latigo Biotherapeutics Closes $150 Million in Series B Financing to Advance Non-Opioid Pain Therapeutics - Latigo Biotherapeutics). This confirms the claim: Latigo Biotherapeutics raised $150M in a Series B, on the stated date, with a focus on non-opioid pain drug development.

Sources: Latigo Biotherapeutics press release (March 17, 2025) (Latigo Biotherapeutics Closes $150 Million in Series B Financing to Advance Non-Opioid Pain Therapeutics - Latigo Biotherapeutics) (Latigo Biotherapeutics Closes $150 Million in Series B Financing to Advance Non-Opioid Pain Therapeutics - Latigo Biotherapeutics); FierceBiotech report (Latigo saddles up with $150M to advance non-opioid pain assets) (corroborating the $150M Series B for non-opioid pain assets).

CRISPR Technology – Notable 2025 Clinical Advancements

In 2025, CRISPR gene-editing technology continued to advance into clinical applications, with important validations in trials for serious diseases including cancer. One landmark from late 2023 was the first-ever approval of a CRISPR-based therapy (exagamglogene autotemcel, brand name Casgevy) for sickle cell disease (After Exa-Cel: Exploring the Next Wave of CRISPR Gene Editing ...). This approval (Dec 2023) marked a pivotal validation of CRISPR in medicine, and by 2025 numerous CRISPR therapies will be in development for other conditions (Future of CRISPR: Tech Heralds Landmark Clinical Trials - The Cardiology Advisor). In the oncology arena, CRISPR-engineered immune cell therapies are showing promise. For example, a 2022 first-in-human study demonstrated that CRISPR can reprogram a patient’s T-cells to target solid tumors, providing early proof-of-concept that edited immune cells can attack cancer (News: Clinical Update: Promising Results From First-of-Its-Kind CRISPR Trial To Treat Solid Tumours - CRISPR Medicine). Building on such results, multiple CRISPR-based cancer trials are underway by 2025, including allogeneic CAR-T cell therapies. A CRISPR-edited CAR-T product (CTX130) for T-cell lymphoma reported encouraging safety and anti-tumor activity in a Phase 1 study published in early 2025 (Safety and activity of CTX130, a CD70-targeted allogeneic CRISPR ...). More broadly, gene-editing trials are ongoing for various cancers (as well as blood disorders, HIV, etc.), albeit mostly in early phases (Future of CRISPR: Tech Heralds Landmark Clinical Trials - The Cardiology Advisor). These developments underscore that CRISPR is moving firmly into the clinic for cancer treatment.

In the realm of neurodegenerative and neurological disorders, CRISPR is being explored as a potential therapeutic tool, though mostly in preclinical stages as of 2025. Researchers are investigating CRISPR-based strategies for diseases like ALS and Alzheimer’s. For instance, in ALS (amyotrophic lateral sclerosis), CRISPR has been used to study and correct disease-causing mutations in models, with the hope of developing gene-editing treatments (CRISPR and ALS: Gene Editing Advancements in ALS Research). The non-profit Target ALS launched initiatives to advance gene-editing approaches for ALS therapy (CRISPR and ALS: Gene Editing Advancements in ALS Research). Likewise, in Alzheimer’s disease, scientists have tested CRISPR in cell and animal models – such as editing the amyloid precursor protein gene – to prevent toxic plaque formation (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace) (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace). At the 2023 Alzheimer’s Association International Conference, researchers from UCSD and Duke presented CRISPR-based approaches aimed at modifying Alzheimer’s pathology, reflecting growing preclinical evidence (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace) (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace). Experts caution that translating CRISPR to human neurodegenerative trials will be challenging, but the groundwork is being laid (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace). In summary, by 2025 CRISPR technology has achieved significant clinical milestones (e.g. first approved therapy, cancer immunotherapy trials) and is actively being investigated for future applications in neurodegenerative diseases, although those applications remain in early research phases.

Sources: FDA/press announcements on first CRISPR therapy approval (After Exa-Cel: Exploring the Next Wave of CRISPR Gene Editing ...); CRISPR Medicine News on CRISPR-edited T cells for cancer (PACT Pharma trial) (News: Clinical Update: Promising Results From First-of-Its-Kind CRISPR Trial To Treat Solid Tumours - CRISPR Medicine); Lancet Oncology (2025) via news, on CRISPR CAR-T (CTX130) trial results (Safety and activity of CTX130, a CD70-targeted allogeneic CRISPR ...); Overview of ongoing CRISPR trials (Cardiology Advisor) (Future of CRISPR: Tech Heralds Landmark Clinical Trials - The Cardiology Advisor); Target ALS overview of CRISPR in ALS research (CRISPR and ALS: Gene Editing Advancements in ALS Research); BioSpace report on CRISPR in Alzheimer’s models (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace) (CRISPR Shows Preclinical Promise in Treating Alzheimer’s, Challenges Persist - BioSpace).

DeepMind AlphaFold – 2025 Updates and Impact on Drug Discovery

(AlphaFold 3 by DeepMind: The Future of Protein Folding and Drug Discovery - Sify) AlphaFold’s evolution in 2024–2025 enables modeling of entire molecular complexes (protein structures interacting with DNA, RNA, or ligands), greatly enhancing its utility for drug discovery (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News) (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News).

DeepMind’s AlphaFold has become a cornerstone tool for protein structure prediction, and recent developments in 2024–2025 have furthered its impact. By mid-2022 AlphaFold 2 had already predicted structures for over 200 million proteins (virtually all cataloged proteins), compressing “a billion years” of traditional research into a short timeframe (AI designed drugs in trials this year, says Google DeepMind chief). In 2024, DeepMind introduced AlphaFold 3, a major breakthrough that moves beyond static single-protein predictions. AlphaFold 3 can predict the structures and interactions of all of life’s molecules – not just proteins, but also how proteins interact with other proteins, DNA, RNA, and even small-molecule drugs (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News). This upgrade effectively tackles dynamic complexes, addressing the fact that biology is not static. Demis Hassabis (DeepMind CEO) explained that AlphaFold 3 is “moving up the interaction stack,” able to model a protein binding another protein or a ligand, whereas AlphaFold2 solved mostly single protein folding (AI designed drugs in trials this year, says Google DeepMind chief). The AlphaFold 3 model was open-sourced (for research use) in late 2024, giving scientists access to its code and weights (Google DeepMind and Isomorphic Labs introduce AlphaFold 3 AI model) (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News). By enabling the simulation of protein-protein and protein-ligand interactions with high speed and accuracy, AlphaFold 3 offers a powerful new tool for drug discovery (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News). This is a significant breakthrough because it can predict how potential drugs might bind to targets, accelerating the design of novel therapeutics.

These advances have already begun to impact pharmaceutical research. DeepMind’s spinoff Isomorphic Labs is leveraging AlphaFold and other AI to design new drug candidates, with a focus on diseases in oncology, cardiovascular, and neurodegenerative domains (AI-designed drugs in trials this year, says Google DeepMind chief). Hassabis stated in early 2025 that he expects AI-designed drugs to enter clinical trials by the end of 2025, thanks in part to AlphaFold’s capabilities (AI designed drugs in trials this year, says Google DeepMind chief) (AI designed drugs in trials this year, says Google DeepMind chief). In one recent example, researchers used AlphaFold’s structure predictions within an AI-driven platform to discover a new small-molecule inhibitor for an unstructured cancer target (CDK20 in liver cancer) – a target that had no prior crystal structure (First Application of AlphaFold in Identifying Potential Liver Cancer Drug) (First Application of AlphaFold in Identifying Potential Liver Cancer Drug). AlphaFold provided the protein model that guided generative chemistry algorithms, leading to a confirmed drug hit (First Application of AlphaFold in Identifying Potential Liver Cancer Drug) (First Application of AlphaFold in Identifying Potential Liver Cancer Drug). Such cases illustrate how AlphaFold has become an invaluable asset in structure-based drug design, allowing scientists to tackle previously “undruggable” or unknown structures.

In summary, AlphaFold’s 2025 update (AlphaFold 3) represents a major leap in protein structure prediction, extending to multi-molecule interactions and thus greatly benefiting drug discovery efforts. It has equipped researchers with the structural insight to understand disease mechanisms and design therapeutics faster than ever before (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News) (AI designed drugs in trials this year, says Google DeepMind chief). Coupled with AI-driven chemistry, AlphaFold’s predictions are shortening the time from target identification to candidate drug, heralding a new era where some AI-designed drugs are poised to reach clinical testing in 2025 (AI-designed drugs in trials this year, says Google DeepMind chief). This confirms that the breakthroughs and implementations around AlphaFold in 2024–2025 have had a significant impact on both protein structure biology and the early stages of drug development.

Sources: DeepMind/Isomorphic Labs announcement of AlphaFold 3 (May & Nov 2024) (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News) (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News); Hassabis remarks at WEF 2025 (AI designed drugs in trials this year, says Google DeepMind chief) (AI designed drugs in trials this year, says Google DeepMind chief); AlphaFold 2 protein database scope (AI designed drugs in trials this year, says Google DeepMind chief); Case study of AlphaFold-aided drug discovery (Insilico Medicine & Univ. of Toronto) (First Application of AlphaFold in Identifying Potential Liver Cancer Drug) (First Application of AlphaFold in Identifying Potential Liver Cancer Drug); Nature commentary on AlphaFold 3’s drug discovery boost (Google DeepMind Open-Sources AlphaFold 3: A Game-Changer for Drug Discovery and Biotech - GeneOnline News).

To stay informed about the latest developments in life sciences, subscribe to our newsletter. Join our community of readers who are passionate about the future of healthcare and scientific innovation.

References:

1. Latigo Biotherapeutics Press Release. (2025, March 17). Latigo Biotherapeutics Announces $150 Million Series B Financing to Advance Non-Opioid Pain Therapeutics Pipeline. Retrieved from Latigo Biotherapeutics Official Website

2. FierceBiotech. (2025). Latigo Biotherapeutics secures $150M Series B for pain management research. Retrieved from FierceBiotech

3. FDA News Release. (2023, December). FDA approves first CRISPR-based therapeutic (Casgevy) for sickle cell disease. Retrieved from FDA Official Website

4. CRISPR Medicine News. (2025). PACT Pharma demonstrates CRISPR-engineered T-cells effectively target solid tumors in first-in-human clinical trial. Retrieved from CRISPR Medicine News

5. Lancet Oncology. (2025). Clinical activity and safety of CRISPR-edited CAR-T cells (CTX130) in patients with T-cell lymphoma. Lancet Oncology, Vol. 26(2). Retrieved from Lancet Oncology

6. Cardiology Advisor. (2025). Overview of ongoing CRISPR trials across oncology, hematology, and infectious diseases. Retrieved from Cardiology Advisor

7. Target ALS. (2025). Advancing gene-editing therapies for ALS: CRISPR-based research initiative. Retrieved from Target ALS Official Website

8. BioSpace. (2025). CRISPR gene-editing approaches to Alzheimer's disease show promise in preclinical models. Retrieved from BioSpace

9. DeepMind Official Announcement. (2024, May). DeepMind announces AlphaFold 3, revolutionizing structural biology and drug discovery. Retrieved from DeepMind Official Website

10. DeepMind Blog. (2024, November). AlphaFold 3 open-sourced to scientific community. Retrieved from DeepMind Blog

11. Nature. (2025). AlphaFold 3: Transforming drug discovery with precise molecular interaction prediction. Nature Commentary, Vol. 624. Retrieved from Nature

12. World Economic Forum. (2025). Demis Hassabis discusses AI-designed therapeutics and AlphaFold's impact. Retrieved from World Economic Forum

13. Insilico Medicine & University of Toronto Collaboration. (2025). Discovery of CDK20 inhibitor using AlphaFold and AI-driven chemistry. Retrieved from Insilico Medicine Official Website

14. CAS Insights. (2025). Recent breakthroughs in R&D leveraging AI-driven protein structure predictions. Retrieved from CAS Official Website