A deep dive into sensing and analytics for biopharma production.

For about six months, under the superb biotech-y name Abrio, I investigated business opportunities for biosensing technology in the production of biopharmaceuticals.  Part-time contract work at the time gave me significant flexibility to travel and interview stakeholders in this industry. I targeted a specific date to either pursue seed funding or to accept an unrelated offer for a full-time position.  Judging that my odds of rapid traction were low, I chose the latter. But recently, I have revisited the lessons that I drew from this exercise, which I now see as a net positive career step.

Bioproduction in industry

Biopharmaceuticals are mainly produced by living cells.  Drug development progresses from lead discovery to selection of a few candidates for preclinical validation.  By the time a drug candidate is ready to file for investigational new drug (IND) status to begin trials, it must be exactly specified, pass many assessments of efficacy and toxicity in animal models, and be manufactured by a standard, locked-down process.  Manufacturing development follows parallel phases of maturity: first small-scale bioproduction to support lab experiments, then scale-up to support IND filing and trials, then full-scale manufacturing where cost and throughput are main concerns. This spans several years and usually each step changes hands to a different technical team.

By the time a process is used in an IND filing, the protocol is rigorously documented and frozen according to good manufacturing practices (GMP).  Quality is formally monitored and even small changes require revalidation. Changing anything about a process once it is in GMP is a major deal, requiring quality tests, documentation updates, and in many cases some degree of re-approval from regulatory agencies.  

My technology proposal

As part of my PhD research, I adapted pulsed lasers and single-photon detectors to monitor the random diffusion of biomacromolecules.  During that time, my labmates and I used this to study properties of proteins attached to cell membranes. One such effect we could study, practically in real time, was the slowing down of two proteins’ diffusion rates when crosslinked with antibodies.  Antibodies are routinely used to measure a huge variety of biological signals of interest with high sensitivity, such as by enzyme-linked immunosorbent assays (ELISA). But ELISA and related methods usually require incubation and rinsing steps that can take hours.  I suspected that I could adapt my approach to turn around answers in a minute or less, and sketched out applications in biosecurity, point-of-care diagnostics, and bioproduction.

I started this project with the goal of identifying an agent that was readily measurable by ELISA, but where an answer was needed in a minute or less, not hours. Antibodies, which are expensive, would be consumed in the process, and the photonic instrumentation needed to make it all work would cost tens of thousands of dollars.  So the application had to be big enough where that speed improvement was highly valuable. In principle, biopharma production met this requirement, as a single weeklong production run can yield (or ruin) millions of dollars of drug product.

Knowing that a working prototype would be a significant investment, I decided to first study the industry and solicit feedback from potential customers.

Biosensing use cases

Bioproduction is generally separated into upstream and downstream phases.  In upstream, carefully selected cells are cultured in media in a reactor, producing a mixture of many things that includes the drug or a critical part thereof.  Downstream consists of separating and purifying the product, which in drug production often includes many steps to eliminate contaminants that can harm a patient.  

There are many variations on this workflow, and it pervades many industries, but many biopharma lines converge on a relatively common process: the drug contains monoclonal antibody (mAb), produced by mammalian (commonly CHO) cells upstream, and separated by  chromotography downstream.  The levers that bioproduction engineers turn are typically small variables within this process, not wholescale and risky changes to the general approach. By the time a drug candidate reaches bioprocess development, the precise cell line is largely established and scientists are making subtle changes to media formulation, reactor conditions, and separation protocols.

Drug production obviously creates enormous demands on purity and quality: small misteps can kill or injure humans, also derailing trials, product growth, and hurting corporate brands.  This is even more dramatic in biologics, because live cells are extremely sensitive to their environments in a way that makes them somewhat unpredictable. In addition to open-loop bioproduction controls implicit in GMP, the industry closes production loops with a number of bioanalytical methods, collectively known as process analytical technology (PAT).  In biologic production, these include real-time sensing of upstream reactor conditions, such as temperature, pH, glucose, carbon dioxide and oxygen. Reactors are typically outfitted with dedicated sensors for these variables.  

Analysis also includes cascaded measurements of drug product concentration, or titer; and measurements of non-drug biological agents, including host cell protein (HCP), non-protein byproducts of upstream production like endotoxin, and crosslinked and fragmented species of the drug itself.  These are typically performed in a separate laboratory by biochemical assay, analytical chromotography, and mass spectrometry. In large, mature companies, turnaround time for routine measurements was typically 1-2 days, and most people that I interviewed did not find this to be a pain point.

Testing value propositions

My original technical vision, a chip for near-continuous fluorescence immunoassays, would adapt commercially available research antibodies with affinity for specific analytes.  This was best suited to measuring biological products such as proteins, peptides, lipopolysaccharides, and complexes of these. My hope was to find specific analytes for which lab immunoassays existed but where turnaround time was limiting.  The value proposition was that answers could be delivered in minutes, thus closing upstream feedback loops to change or abort a process before doing damage. But my interviews led me to conclude that this was not a compelling problem: measurements that were needed quickly were already possible with existing chemical sensors, and tricky biological measurements could wait for reports from central analytical labs.  

Developing better assays, for example for drug product crosslinking, would be welcome, but this did not align with my core competency on the sensor engineering side.  And even then, liquid chromatography with and without mass spectrometry (LC-MS) was maturing rapidly: instrument integrators increasingly market in-line LC-MS solutions tailored specifically to bioproduction, and turnaround time was very impressive.  This would obviate many analytical laboratory orders where production lines could justify installing LC-MS.

Another enlightening discovery was that not all users were enthusiastic about measuring more variables.  All sensors produce some degree of noise, and this is particularly prevalent in immunoassays where assay interference (cross-talk with other analytes) is common.  Particularly for the GMP phase, aggregating more signals could lead to worse overall results: their context and significance were new, and new information could be confusing but not obviously actionable.  Manufacturing operators could be liable for either ignoring or overreacting to these new signals.  This particular lesson is relevant to any potential project in sensing and analytics.

Furthermore, recent product introductions further alleviated the pain points surrounding biological measurements.  Because a wide swath of biologic production was mAbs, and these share a constant Fc protein domain, a single affinity assay could measure titers of a wide variety of drug production lines.  BioForte had recently entered the market with a surface plasmon resonance-based sensing instrument with protein A affinity against mAb-Fc. It was targeted toward bioprocess development users that did not want to rely on analytical laboratories for titer quantification, a highly routine task.  But, at the time of this project, reactions to its accuracy and usefulness were mixed among the people I interviewed.

But given the need for rapid, stable measurements over the course of days, there was a surprising need to measure even mundane chemistry, instead of biological products.  At least one company was working on a disposable pH sensor that immersed in single-use bioreactor bags.  Another was Finesse, later acquired by ThermoFisher, whose product measured oxygenation.  

One variable that stood out as a pain point was, surprisingly, glucose.  Of course, assaying glucose is straightforward in the lab. And measuring it at the reactor could employ electrochemical sensing with the glucose oxidase enzyme, the exact same technology that diabetics use to measure their blood glucose at home.  But this would be difficult because of constraints imposed by adjacent technologies.

Recognizing the technological ecosystem

Engineers often picture their inventions as standalone solutions, overlooking the many ways they connect with adjacent technologies.  But any potential product needs to fit into an ecosystem, and this introduces both constraints and opportunities.

One constraint in this application was sterility.  With few exceptions, reusable bioreactors are rigorously steam sterilized, removing trace bacteria and fungi that can quickly ruin high-stakes production runs.  Any sensor that interfaces with a bioreactor needs to undergo sterilization and still perform. This rules out the concept of inserting affinity bioreagents, such as sensing antibodies, inside of reactors themselves.  

Where lab-based ELISAs were used to measure bioreactor contents in the middle of a production run, an operator would need to access the contents by an aseptic sampling port.  Indeed, a few beta products addressed this, automating sampling at predetermined intervals and retaining samples in well plates or pouches. In fact, I considered pivoting my focus to dramatically reducing the sampling volume in these autosampling systems, from milliliters to microliters.  But this would really only be valuable for the early phases of bioproduction development, where development runs often use many small bioreactors, and drawing many milliliters for sampling would noticeably change fermentation conditions.

Another ecosystem-driven constraint was the bioreactor itself.  Traditionally, these have been glass or stainless steel tanks that hold up to many cycles of production and sterilization.  But precisely because sterilization is a time-consuming and faulty process, the industry was rapidly moving to single-use plastic bioreactors.  These could be gamma sterilized in the factory and were inexpensive enough to be disposed of after each production run.  This complicated the sensor business: threaded sensor ports were no longer standard, and operators would not be happy sterilizing sensors when the reactor itself was disposable.  Presterilized sensors that were also single-use would have a competitive advantage.

So while I found a valuable opportunity in glucose sensing, sterilization and bioreactor integration stood out as engineering challenges.  These quickly digressed from my comfort zone – photonic sensing and fluid handling – as reengineering of detection enzymes would be a likely need.

Looking back

When I began this project, my inner engineer deeply wanted to start by building something – anything.  But taking a page from Lean Startup, I decided to first study market opportunities and talk to potential customers.  I was able to do so at minimal expense, and learned a lot in a short amount of time.  And I strongly feel that I avoided building and pitching an expensive venture that would not have been a huge win.

Along the way, many people took time out of their busy schedules to help me articulate this problem.  I am particularly grateful for some senior technology managers in biopharma – they probably entertain hundreds of such requests a year from mature and well funded companies, but still hosted me with patience and respect.  As we cross paths again, I look forward to sharing this article with them.

The potential applications of real-time ELISA sensing remain exciting: improving nutrition with real-time feedback of entire metabolic pathways, or monitoring firefighters in high-stress situations via endorphin levels.  With growth in cellular and gene therapies, there is also a growing need to engineer biological feedback loops within the short timeframe of therapy administration, such as for cytokine dynamics associated with CAR-T cancer treatment.  So needless to say, I remain on the lookout for major opportunities to develop such a technique and make it useful.