By Brian Finrow and Jim Roberts
We just released a paper on the preclinical development of LMN-201, Lumen’s orally delivered cocktail of therapeutic proteins for preventing C. difficile infection. In the news release we also shared that LMN-201 met its primary endpoint in Cohort 1 of the ongoing Phase 1 pharmacokinetic dosing study. Full data from both will be released after peer review and completion of trial Cohort 2.
We are intensely proud of these results. As detailed below, no one has ever developed a drug like this before, and the results have implications for biopharmaceutical drug development far beyond Lumen Bio.
C. difficile infection: a brief primer
Plenty has been written about how awful C. difficile infection (CDI) is. Estimates for the cash drain on US hospitals range from $5-$8 billion per year. Add in rough economic estimates of the human toll and the total societal cost in the US is probably closer $25–30 billion per year. The global burden is higher still.
Although CDI is classified by the CDC as a top antimicrobial resistance threat, it is not, in fact, resistant to antibiotics. Generic and cheap vancomycin clears up most cases easily. However, vancomycin also damages the microbiome, leading to frequent recurrence of CDI. In fact, broad-spectrum antibiotics, used to treat a diversity of medical conditions, are the leading cause of CDI and — particularly in high-risk patients such as the elderly and immunocompromised — can trigger a downward spiral of infection and reinfection. At worst, this can culminate in colonic resection surgery or death. One independent leading study puts the annual per-patient cost of CDI recurrence at $34,000, mostly in hospital services (ICU time) and surgery. The 12-week fatality rate in Merck’s big CDI recurrence prevention trial was 7%.
In short, CDI prevention represents a massive unmet medical need.
For the reasons above, people no longer prescribe antibiotics prophylactically, so the only FDA-approved preventative available is bezlotoxumab, an IV-infused monoclonal antibody. It works for about a quarter to a third of patients, but access is burdened by several challenges. As always with injectables, needle fear is a significant barrier, and so is the encumbrance and additional cost of booking time in an IV-infusion clinic.
Although not yet FDA approved, another option is fecal microbiota transplant (FMT), which aims to accelerate microbiome re-engraftment by transplanting stool from a healthy donor. One FMT approach showed clear efficacy in a large, well-controlled trial, although direct comparisons with bezlotoxumab are unclear since that particular study enrolled only the small subset of CDI patients with the highest odds of recurrence (those with at least three prior events).
FMT is not risk-free, though: there have been fatalities due to the accidental transfer of gastrointestinal (GI) pathogens. The leading nonprofit stool bank supplying the nascent field initiated a wind-down in 2020 due to the risk of transplanting Covid-19 (which also infects the GI tract). Newer FMT approaches are trying to make the process safer, but it’s unclear whether their supply chains (most involving human donors) can be scaled up to meet the enormous global need. To pay for it all, Wall Street analysts expect these products — if ultimately approved by the FDA — to be priced even higher than bezlotoxumab. Worse, most require even more cumbersome logistics than IV infusion (in most cases, a preparatory bowel lavage or enema administration).
LMN-201: a four-part biologic cocktail for preventing CDI
It’s against this backdrop that Lumen set out to develop LMN-201: an orally delivered cocktail[i] of four protein therapeutics for preventing CDI recurrence.
One of the four is a lytic enzyme derived from a bacteriophage that potently destroys C. difficile cells with a high degree of specificity (i.e., it kills the bad bugs but leaves your commensal bacteria alone). Lysins are not antibodies, so it’s outside the scope of this Medium piece. For the curious, here’s a great survey paper from our collaborator Vince Fischetti on endolysins, and here’s a paper on the particular endolysin incorporated into LMN-201.
The stars here are the three “antibodies”[ii] within LMN-201 that target C. difficile toxin B (TcdB), the key virulence factor in CDI.
What’s different; what’s new?
One appealing feature of this disease target is that there’s over 20 years of research — including a pair of massive Phase 3 clinical trials — that definitively proved that neutralization of Toxin B can prevent CDI recurrence. The toxin must be neutralized within the gastrointestinal tract, so it’s a very straightforward application of Lumen’s platform.
With this in mind, we set out to develop a CDI product that would overcome the drawbacks of all other CDI preventatives on the market or in development. In our view, the ideal product would have the following characteristics:
1. High efficacy: we aim to beat the 33%-66% relative risk reduction previously reported for other CDI preventatives
2. Oral delivery: to eliminate adoption barriers and health services and facilities costs associated with other preventatives (no IV needles; no enema/bowel prep)
3. Safety: relative to injected antibodies, orally delivered antibodies offer important safety advantages by preventing systemic exposure; relative to FMT, there’s no risk of donor pathogen transfer with monoclonal protein therapeutics
4. Scalable manufacturing: the market for a safe and effective CDI preventative is massive and global; costly manufacturing of injection-grade antibodies is not scalable to meet that demand, and FMT-based products involving human donors are even less so
5. Shelf stability: as Covid-19 showed, refrigerated supply chains limit availability
On this list, #2, #3, #4 and #5 are key features of the Lumen platform, and we’ve published on this previously (pre-print; peer-reviewed version now working through the final publication process). The intuitive appeal of oral delivery is also covered nicely in this survey paper on the past 30 years of clinical trials using the approach.
Having already de-risked most elements, our focus turned to objective #1: maximizing clinical efficacy. Our goal is not just to make an incremental advance; we are intent on transforming the management of C. difficile.
Origins of the antibody cocktail idea
We’ve written before on the origins of our oral biologics programs, including how the Gates Foundation first approached us with the general idea in 2017 and connected us with a team at Tufts University. In one of those fun coincidences that happen often in startups, that same group had just published on a set of Toxin B neutralizing VHH antibodies. So in our initial small collaboration agreement, the team there kindly allowed us to investigate those as well (we’ve since licensed them).
The threshold question for us was a common one in drug development: which of the many antibody “hits” should be elevated to the status of “lead.” But why were we looking at combinations? At this point in time (late 2017) all antibody drugs approved had been single-agent products. Even though the FDA’s 1997 guidance on the topic explicitly contemplated high-valence antibody cocktails, 100% of actual approvals were single agents.[iii] So naturally our assumption at the outset was that — like everyone else — we should plan to advance just a single winning antibody into clinical development.
We were, however, thinking about antibody cocktails a lot. The idea was central to our ongoing work with the Gates Foundation developing a preventative antibody drug cocktail for infant diarrheal diseases, where a broad set of bacterial and viral targets must be neutralized. But for CDI, the variant coverage challenge is much less of a concern as compared to, say, a rapidly mutating RNA virus like SARS-CoV-2, so it wasn’t much of a major concern at the time.
From one to many; from good to great
Our ability to develop safe antibody therapeutics and manufacture them at a fraction of the typical cost paved the way for commercially feasible antibody cocktails. So while we were starting to think about CDI, we were also working hard to extend Lumen’s lead in antibody manufacturing scale and cost (more on that here). And it was in that context we became interested in a branch of mathematics called designed experiments (aka statistical design of experiments, or DoE for short).[iv] It’s not been used before in drug development, and it’s an essential step in the results reported in our new paper.
DoE is all about optimizing the output of something where theory predicts that several variables may be key to the outcome but the potential synergies between those variables cannot be modeled with precision. In the mid-20th century, researchers discovered that although such synergistic effects often existed, simply running experiments over and over again while changing one variable each time was an inefficient way to find them. Testing variables one at a time while holding the others constant effectively ignores higher order combinations. And the vastly larger search space representing all possible combinations at all possible treatment levels is far too large to explore comprehensively due to the implausibly large number of experiments entailed.
Expert human intuition is actually pretty good at making guesses in this sort of game, but it’s far from perfect. Complicating things further, counterintuitive results are commonplace. For example, it might be that the optimal synergy requires turning variable 1 up by a small amount, while turning down variable 2 — what experts in that field called “two- or three-factor interaction effects” (analogous to the concept of synergy in our new paper).[v] In fact most real-world situations serve up more than just three plausibly synergistic variables, so in reality the challenge is even larger, requiring the expert to imagine things happening five- or ten-dimensional search space. Impossible.
DoE was developed and elaborated throughout the 20th century to deal with this complex situation, with profound effects embedded in nearly all products on the market today. Yet nobody had ever attempted rationally engineer their way into a more potent antibody drug by applying DoE techniques to the discovery process. At this point, we didn’t have more than a sneaking suspicion that rationally engineered cocktail synergies might be possible, but we took a small leap of faith. Startups are fun places for that sort of thing.
From tentative first steps to final flowering
We began experimenting with different two- and three-way combinations out of the Tufts collection, and we knew immediately that we were on to something.
These in vitro experiments are well established in the CDI field. To know how well your antibody blocks Toxin B, you start by culturing mammalian cells onto a small plastic dish, a common lab procedure. If they’re the right type they’ll flatten out and tightly connect to each other — just like the epithelial cells lining your GI tract. When increasing amounts of Toxin B are titrated onto healthy cells, they start to ball up and die — again just like when your GI tract is attacked by C. difficile. If your antibodies are good, preload them on the dish before the toxin and they’ll protect the cells from rounding up and dying.
Shown below is a diagram of the structure of the C. difficile Toxin B protein, co-crystalized with the binding domains of the three best antibodies from the Tufts collection depicted in light blue (labeled 5D, E3, and 7F).[vi] Right away you can see that the VHHs bind noncompetitively, which makes intuitive sense if they’re to work together toward the shared goal of toxin neutralization.
So far so good, but we quickly encountered mysteries. At middling ranges of toxin and antibody, it was easy to model what was going on, but the relationships broke down in confusing ways at concentration extremes of either one. This is important because Toxin B concentrations can vary widely from patient to patient (and within each patient over the course of the disease). And so will GI luminal antibody concentrations, given the wide swings in intestinal protease concentrations and microbiome composition due to dietary changes and disease state. The ideal product should provide robust protection across all likely concentration ranges of both toxin and antibody. But how to model such an astonishingly complex and dynamic set of conditions? A second set of mathematical tools was required.
With the benefit of hindsight the answer seems obvious, but the critical insight wasn’t to be found in the antibody development field. However, Jim ultimately realized that the perfect mathematical toolkit existed in yet another niche branch of mathematics — one originally developed in the early 1900s to model the behavior of enzyme inhibitors.[vii] In fact Michaelis-Menton kinetics are taught to all scientists in graduate school, but from the point of view of enzymology.
That was the final conceptual breakthrough. At this time, we were fortunate to have a new colleague join the team (Mike Dodds), who stepped in and methodically worked out all the theoretical nuances, which involved a huge number of sequentially designed and executed DoE experimental runs to first build, and then prospectively validate, the best-fit model. The details are unduly technical for a blog post but suffice it to say that the refined model predicts quite reliably the dynamics observed in vitro. Once established it is straightforward to validate with prospectively gathered experimental data. The key takeaway here is that both of these mathematical traditions had to be imported and then integrated to yield the results you see today.
Now we know what to look for, and exactly how to most efficiently search for it. The supplemental materials detail a step-by-step methodology for teasing out similar potency boosts in other antibody drug development targets, assuming of course the relevant biological mechanism lends itself to such synergy. Importantly, the approach does not require ex ante knowledge of the underlying biophysical processes, just a way to rigorously evaluate the preferred outcome — exactly like the production optimization processes that DoE revolutionized in other fields. Modern antibody discovery programs typically start with hundreds or thousands of candidates. Efficiently sorting through these for optimal cocktail combinations will certainly require an analytical approach like we have described.
As useful as this process is for efficiently finding antibodies with synergistic efficacy, perhaps the even more remarkable outcome of our work is how common antibody synergy turns out to be. It is a property of antibody cocktails that can be much more easily achieved than has been anticipated. This is because antibody synergy does not depend on relatively rare and unusual instance of cooperativity or allostery. Rather, in many (maybe most?)[viii] cases synergy can be entirely explained by the principles of mass action acting on antibodies that bind to independent sites on a target. As the enzyme modelers first showed, this is sufficient to cause multiplicative (i.e., synergistic) as opposed to additive effects. Our extension of this concept to antibody cocktails was rigorously validated by the methods described in our publication.
The multi-thousand-fold potency leaps generated in our new paper show that such dramatic performance leaps are every bit as available to drug developers as they have been available to chemical process engineers for 80 years. And in our new paper we published a user manual for antibody drug developers interested in accessing the same results.
This is, as far as we can tell, the first time these mathematical techniques have been used prospectively to rationally design a synergistically acting antibody cocktail for advancement to clinical development. Importantly, this is not just theorizing: alongside the paper’s release we announced completion of Cohort 1 in the Phase 1 pharmacokinetic clinical trial of LMN-201. The paper also neatly illustrates a second flavor of synergy in the same product: the mechanistic synergy offered by the endolysin, which operates through a pathway orthogonal to toxin-neutralizing antibodies.
Like any other tool, this one doesn’t solve all problems. Where high pre-clinical or clinical development costs or safety considerations preclude cocktail formulation entirely, single-antibody approaches will have to do. But the techniques pioneered by Lumen in this paper do hold great potential for new treatments in a wide range of diseases that have proven intractable for traditional drug making approaches.
CDI is one important leading example, and certainly a huge unmet medical need. Inflammatory bowel disease (IBD) is another. Just like CDI, the therapeutic targets are well-known and safe. And just like CDI, at least three prior groups have demonstrated that they can impact IBD outcomes — even systemic markers of inflammation — using orally delivered antibodies. The challenge historically was achieving high potency (dose size is critical in chronic diseases) and, of course, robust safety and scalability (Lumen’s forte). Cardiometabolic disease is a third example. In our research collaboration with Novo Nordisk we are exploring a range of targets. And just like with CDI and IBD, at least one human clinical trial has suggested that oral delivery of an antibody antagonist can trigger meaningful improvements in disease state. The data are thinner, but as with CDI and IBD, the chronic nature and high prevalence place a premium on potency and scalability.
And just in time: biopharma industry’s miserable <10% clinical success rate for new drug candidates compares poorly next to the routine triumphs of other, more profitable industries. It is the primary cause of the well-documented long-run decline in the return on investment for biopharma R&D.
Fresh thinking is clearly needed.
Why hasn’t this been done before?
As noted above, while there are now a handful antibody cocktails on the market, they are simplistic in comparison. Except for Regeneron’s Ebola product, all are two-antibody mixes for Covid-19 (separately administered pairs of anti-cancer antibody drugs look promising too, assuming costs can be brought under control.)
However, none of these were rationally engineered for synergy in the way Lumen pioneered here. Perhaps it’s partly down to COGS. Injection-grade monoclonal antibodies cost $100-$200 per gram by most estimates, and Regeneron’s Ebola drug is dosed at a whopping 10.5 grams for a typical adult (150 mg/kg). It’s hard to build a scalable business with that kind of cost structure. The obscenely high cost of developing a traditional injection-grade, CHO-expressed antibody likely plays a role too — particularly cell-line and process development at ~$10 million per antibody. It’s probably no coincidence that the heavy government subsidies were involved in the development of these examples.
Mostly, though, it seems that the opportunity was simply overlooked. DoE was originally a manufacturing technology, and in the pharmacology industry its application has been limited to improving manufacturing quality systems. It has not been adopted by drug discovery teams. Cross-fertilization across fields is often highlighted in the origin stories of breakthrough innovations, but engineering it purposefully isn’t so easy. So a big part of this is down to luck.
But then again, Lumen Bio is a lucky company. This sort of thing happens around here all the time. Sometimes there really are $100 bills on sidewalks. We simply saw this one first and picked it up.
Brian Finrow is cofounder and CEO of Lumen Bioscience; Jim Roberts is Lumen’s cofounder and Chief Scientific Officer.
[i] Some papers use the term “antibody mixtures,” but we use “cocktail,” the term used by the FDA in its seminal 1997 guidance document on antibody development.
[ii] “Antibodies” is in quotes here because these therapeutic proteins are not, in fact, antibodies. They are engineered, toxin-binding fusion proteins. Antibodies are a particular protein made by a host’s adaptive immune system to fight off pathogens. They come in different flavors (IgG, IgA, IgM, etc.), but each has a particular structure with several obligatory elements necessary for full functionality. In particular, a properly glycosylated Fc domain is necessary for certain secondary immune responses. Lumen’s “antibodies” in LMN-201 lack those elements but include other protein subdomains that are lacking from all actual antibodies. For more general audiences like this one, though, we often refer to them as antibodies to minimize jargon. In LMN-201 they operate very similarly to the polyclonal anti-toxin antibodies from horse serum that are injected into snake-bite victims to neutralize the toxins in venom.
[iii] There were a couple of cocktails in development, most notably Regeneron’s three-antibody Ebola product. But this was a rarity before Covid-19 made two-antibody cocktails commonplace.
[iv] In the manufacturing world, DoE has been fundamental to large-scale production and R&D since the 1950s. It was a cornerstone of the famous “Toyota production system” and its progeny (“lean manufacturing,” “six sigma,” etc.”), and other interesting late 20th century breakthroughs in production efficiency and quality. In fact, it even permeates the biopharmaceutical industry: nowadays the FDA essentially requires it for new drugs as part of their review of the manufacturing process. It just hasn’t really ever been used in the antibody drug pre-clinical or clinical development process, at least as far as we can tell. It doesn’t seem to feature very prominently in the professional training of molecular biologists, for example, despite its near ubiquity in other fields.
[v] For the curious, a more fulsome exploration is included in our paper with Google’s AI/ML team available here as a preprint (including exciting 21st century advances in the field).
[vi] This crystal structure was solved by Rongsheng Jin’s lab at the University of California, Irvine.
[vii] Unbeknownst to us, a pair of Jim’s Fred Hutch colleagues (Tal Einav and Jesse Bloom) were thinking along the same lines and beat us to beat us to the publication punch. Our new paper advances things in at least four ways. We are the first to demonstrate how the model-free insights of the DoE field can be imported into the drug development process to efficiently generate the robust data sets that can inform parameter fitting and selection of best-fit models. Second, we expand the set of mathematical models available to include some others that better capture the behavioral nuances of certain “flavors” of antibody behavior that sometimes emerge in the data: for example, saturation effects that emerge higher and lower concentrations of antibody and/or target are physiologically relevant but are not captured by simpler models. Third, this work demonstrates how the model-agnostic, sequential experimental design principles from the DoE field can be used to inform not just antibody cocktail composition, but also the optimization of relative proportions, a dose-ranging concept (the Einav-Bloom paper was constrained to equimolar concentrations whereas the optimal ratio in LMN-201 turns out to be 1:7:7). Finally, and most importantly, ours is the first such program to demonstrate these principles prospectively in an antibody cocktail development campaign that has advanced to human clinical trials.
[viii] This is an essential point: our paper demonstrates that there is not, in fact, any single model that can predict all the different phenomena that therapeutic antibodies can throw out. Biology is just too complex for such a master key. A good example is the “saturation” effect observed with one of the VHHs in LMN-201 (but not the other two). Antibody cocktails are even more complex due to the potential for allostery (a previously rare phenomenon that the more efficient search techniques pioneered here might plausibly make more common). An example of this provided in our paper is the fact that the optimal relative concentrations are not equimolar, but 1:7:7 — hardly an intuitive composition. A key advance of this work therefore lies in the use of DoE to generate model-agnostic experimental design that provides the multivariable surface response data across a range of cocktail compositions. After model fitting and best-fit analysis, drug developers can then use the model to generate final testable hypotheses for final validation and clinical progression.