Selling AI Products + Services to Big Pharma

It's really hard selling AI to big pharma, so let's talk about how to do it

Aug 08, 2023

The top 700 pharmaceutical companies are some of the largest and most influential organizations in the world with a collective market cap of over $5 trillion. They have learned the science of innovating at scale leading to hundreds of products that are used to save and transform lives around the world. As early, nascent builders of AI, we have the opportunity to equip these organizations with insanely great tools that have the potential to massively accelerate their impact. In reality, getting our tools into these large organizations is an immense challenge that is a blackbox to many. After building AI tools for years, selling them to every top 20 pharma company, and experiencing many tough failures + big wins, I want to share my perspective on selling AI to big pharma.

The Bar

The bar for many big pharma to make large software / service / partnerships purchases is:

What specific drugs is this going to help us get approved?

Many of the big pharma are very explicit about this bar and it sits at the crux of any justification for a large purchase. This blog focuses on understanding what this bar truly means and tactical steps for convincing big pharma that you meet this bar.

Creating and Capturing Significant Value

There are three fundamental truths you need to understand if you want to use AI to create huge value in drug discovery:

The further you are from the drug program decisions, the less value you create
The further the decisions you influence are from the commercial + clinical stages of drug discovery, the less value you create
The decisions that create the most value are choosing the right target/phenotype, reducing toxicity, and predicting drug response + choosing the right target population

You can view the drug discovery process as a series of sequential decisions that are based on data. In this sequence, there are three major stages that matter: preclinical, clinical, and commercial. Each of these stages have substages which are important but the decrease in risk and increase in value between each of these higher level stages is immense.

https://www.baybridgebio.com/drug_valuation.html

This simple graph shows why you create more value if you influence decisions on the right side of the graph compared to the left side. It may also explain why Veeva, a SaaS company who creates software for the commercial and clinical stages, is valued at $30B while Schrodinger, a SaaS company who creates software for the preclinical stage, is valued at $3B.

While the stage of the decision is important, the type of decision is also important. 90% of drugs fail in clinical trials and these failures are either due to lack of efficacy or safety. The efficacy problems almost always boil down to choosing the wrong target / phenotype or choosing the wrong patient population. Drugs rarely ever fail because they don’t bind to the target or the developers did not know the protein structure or the developers could not generate a drug-like chemical with certain properties or the expression yield was too low or there was not a good summary of the literature. All these problems and decisions simply do not create significant value because influencing them does not change the failure rates of moving from one stage to another. This leads to $50M - $150M dollar outcomes, not $10B - $100B outcomes. Solving these problems is like getting better at casting the perfect supporting cast in a movie. Thats great but if you do not have the right lead cast then none of it matters. So if you are building an AI product / service, you want to think deeply on the types of decisions you are influencing and whether they will actually have an effect on helping big pharma develop approved drugs or whether you’re helping pharma cast the perfect extras.

Screw Your Digital Revolution

If your pitch deck or website has any messaging like “Usher in the Digital Revolution” or “Unleash the Disruptive Power of AI” then you are not going to make it. Phrases like these do not mean anything. If you go to BioIT World or any other SaaS for biotech conference, 90% of the messaging will be some flavor of these themes:

AI to develop your drugs faster and cheaper
Discover better, safer drugs with AI
Discover how AI will transform your R&D
Speed up scientific discovery with AI
Use AI to breakdown data silos and generate better insights
Integrate AI and scientific discovery at scale

As a young startup with a new AI software, you will get lost in the crowd if you copy any of these themes. They are generic, fluffy, soft, meaningless language. To a pharma executive, the digital revolution does not mean a $50k ARR contract with your unknown startup. It means a 8-figure cloud migration deal with Google, a multi-year partnership with Microsoft’s AI team, or a cross company effort to integrate ChatGPT. You need to accept that your tiny little startup just does not matter in the vague world of the pharma digital revolution.

For your startup’s messaging to be noticed by big pharma, it needs to clearly and directly communicate how your company influences the one thing pharma cares about: developing approved drugs. Here are some examples of successful messaging:

Systematically and rapidly optimize AAV capsids, overcoming the limitations of naturally occurring virus capsids
By applying revolutionary insights in 3D protein-ligand co-folding, we seek to target previously undruggable disease targets
Make better discoveries with your high throughput drug screens by combining human intuition with unbiased AI
Know the toxicity of your drug before running animal models

These messages are successful because they clearly focus on a specific area of the drug discovery process and directly communicate how they impact that area with simple, powerful, direct language. This type of messaging gets scientists working in these specific areas immediately interested.

Building Trust in AI

The use of AI in their day to day work lives is a completely novel concept to the majority of people in big pharma. There is a huge amount of interest and excitement but a lack of authentic trust when it comes to relying on AI for the high stakes work of drug discovery. This puts the responsibility of establishing trust in the hands of the tool builders. Here are some of the key ways I have worked with my customers to build trust in AI…

Training data

One of the primary ways scientists build trust in AI is by deeply understanding what it was trained on. They want to understand the training data at the foundational level. For clinical data, they want to know if it was prospective/retrospective, the collection protocol, the assay protocol, quality control/review standards, how the sites were chosen, and much more. For preclinical data, they really care about batch effects, size of the dataset, experimental parameters (e.g. model / cell type / exposure time / dosage), the number of chemicals / conditions tested, the type of chemicals / conditions tested, device settings for generating the data, algorithms used to process the data, and more. These are really important details for scientists because its how they are trained to assess the quality of the scientific results they deal with everyday. There is a reason why the Methods section usually comes first in scientific papers and why it commands a prominent spot in the paper. Any results without the Methods are meaningless.

I typically find it to be critical to build a robust data sheet, raw table, and summary of the training data to share with scientists. They spend a ton of time really digging in here, asking questions, and building a thorough understanding of the training dataset. Hiding the training data from the customers is a sure way to start off on the wrong foot.

Interpretability

Machine learning researchers and drug discovery scientists have a different understanding of what it means to interpret an AI model. For ML researchers, they usually see it as trying to understand what the guts of the model weights are doing through saliency maps, feature importance, and a bunch of other fancy statistical methods.

For scientists, they see interpretability in terms of the ground-truth biology/chemistry/physics/science that they have very strong knowledge of. This known, ground-truth is typically called “controls” and scientists have strong expectations of how controls should perform in various experimental conditions. Scientists use the relationships to controls to understand new variables. Given this paradigm, they interpret the AI model by studying how it performs on controlled conditions where they know what should happen. They then use that to build out of map of relationships between a handful of controls and the AI model. So if you have three drugs where one drug is a known biological control and the other two drugs are unknown, a scientist will study the output of the AI model on the known control in order to interpret the model’s output on the unknown drugs. If your AI algorithm predicts the activation of immune cells then scientists will want to see how the algorithm performs on control conditions like TLRs which increase activation and control conditions like IL-10 which are known to decrease activation. This type of relationship mapping and interpretation is the primary method for which a scientist understands and builds trust in a new AI model. In many ways, this is similar to the ML concept of a test set, but with many fewer data points and more knowledge-based rather than statistics-based.

So if you’re building these tools, one of the first things you want to do is find out what controls scientists need to see in order to build trust in your model. These controls will be important in helping you communicate the value of your model and helping scientist know how to use it.

Sometimes scientists push to try to understand the guts of the AI algorithm and exactly what it is learning. To ML/AI practitioners this is obviously an impossible task and our achilles heel as we do not know what these neural nets are learning. Here are some key talking points which I use to help the scientists understand this phenomenon and become comfortable with it:

We use a massive amount of diverse data…
To teach the AI to take an unbiased approach…
To using all the information in the data…
To learn the best features for accomplishing the task.
Contrast this with a biased human approach…
Where a human reduces the data to the singular feature they want…
Which may or may not be the best for the task.

This is how I summarize what AI is doing and set scientists up to understand its value. I do not tell the scientist exactly what the AI algorithm learns (I and nobody else knows this information), I just setup a new world in which they do not need to understand exactly what the AI is learning. This is only a setup. The punch line comes from helping the scientist understand why the AI approach is better or complementary to what they are doing now. After the setup of the new world, you need to make the scientist want to live in the new world.

Do not try to do saliency maps or some other fancy algorithm to attempt an interpretation of neural networks. I have wasted so much time here and its never worked. You need to approach interpretation in a simpler fashion from the perspective of the scientist.

Benchmarks vs use cases

Scientists do not care about how you do on the latest academic benchmarks in your field. Drug hunters have a hundred concerns and your 3% improvement on the SOTA for the benchmark of the month ain’t one of them. They care about how your algorithm performs on real use cases they are dealing with day to day. So instead of spending time talking about your 3% improvement over Alphafold on CASP, talk about how you used your model to help a scientist discover the structure of a very difficult class of proteins that Alphafold repeatedly fails at. Spend the majority of your time talking about real use cases grounded in what scientists are doing day to day rather than abstract benchmarks on curated datasets.

AI vs traditional approach

The one benchmark that scientists do care about is how your AI model is better than what they are doing day in and day out. These benchmarks are not well defined and usually customized so the onus is on you to get creative with how to best empathize with the scientists’ work. A text book example of this happens in computer vision analysis where the majority of scientists are still using classical computer vision techniques and just starting to explore deep learning. So a great benchmark in this case is to take a problem that scientists are doing everyday then contrast what they are doing now against what you enable them to do. Lets take an example like modeling diseased and healthy samples with histology images. To generate an effective benchmark, you can run this analysis using the classical computer vision approach that most scientists are using today then run it again using your AI approach. You can then present these results as a practical benchmark which helps them understand the delta between what they are doing now and what they can be doing if they use your product / service.

Architectures

Scientists do not care about neutral network architectures. If you mention transformers, convolutional networks, graphical neutrals, attention or any other highly technical term to try to sound advanced and trustworthy, you are not going to make it. Focus on establishing trust of the model based on what goes into it and what comes out of it, not how cool, novel, and advanced its inners are.

Ease of use

Scientists deal with a lot of very complicated technology, protocols, decisions, etc. This means that they have been burned by things that are so hard to do that they produce no meaningful results and just waste time. So when assessing any novel technology, they are constantly asking how feasible it will be to integrate with their workflow and how practical it is to start producing value. You need to paint an extremely clear picture for how your product / service is used for it to be trusted by scientists. If its very complicated and difficult then you will trigger an immune response from scientists. If scientists can clearly envisage using your offering then they start to lean in.

Many AI tool developers underestimate just how difficult it is to effectively integrate into a scientist’s workflow as they can be very specialized and complex. There are typically three key stages: 1) getting the input for the model to predict on, 2) getting the predictions to the scientists, and 3) helping the scientists use the predictions to drive towards actions. When reviewing your product, scientists will be assessing how easy it is to do all three of these things. If any one is difficult or impossible then it runs the risk of losing the trust of scientists.

Reducing failure rate

There is this extremely tricky situation where the gold standard evaluation for AI products/services is using it to get a drug approved but this gold standard takes 7-10 years + billions of dollars. There is currently no AI that can claim they can get multiple drugs approved with a failure rate better than the industry standard. I hope this changes over the next couple of years but its going to be hard! Until then, we need some proof points that AI can actually improve how industry develops approved drugs. Here are some ways to get intermediate proof points without developing your own drug program…

In vitro data: Setup a common drug discovery problem relevant to your AI product / service. The problem should have a ground-truth follow up step which tells you whether it worked or not. A good example is choosing “hits” in a broad drug screen then validating that those “hits” do what you want in a focused followup experiment. Make the decision with and without your solution then show that using your solution gets better in vitro followup validation data. This is hard and requires AI tool builders to think like and actually do drug discovery experiments! It helps to get scientists or consultants on your team whose entire job is running these realworld drug discovery validations. Its important to realize that if your AI is only validated and used in the world of bits then it will not be trusted by scientists. Its impact needs to be felt in the physical world with the bare minimum being in vitro proof points.
In vivo data: Refer, to the value chart above. In vitro data is the first step but its on the very, very left side of the value creation slope. What starts to show more value is getting great in vivo results for your AI product/service. These experiments typically cost between $100k - $500k and take several months to complete. This is why its so important that your 1 year AI development plan has these real-world experiments built in. They take a long time and their iteration cycles are very, very slow compared to software. Once you have this data, it serves as a great milestone for validating your claims and helping scientists believe your AI may actually be useful. You should present this data in a scientific paper but also in beautiful, elegant blogposts + slides. The combination of the two modes of communication is really important for hammering home the messaging. The in vivo experiments should be relevant to the claims your product / service is making. So if you are claiming you can choose better oncology targets, then show animal cancer models being run with and without the use of your AI. If you are claiming that you can improve toxicology predictions, predict the output of novel chemicals then setup an animal study to see how different the results are from your predictions.
Design partner: The in vitro + in vivo data should be enough to help you get your first design partner! Ideally, this partner is working on active drug programs that they are advancing to clinical trials. An absolutely fantastic point of validation is that they use your solution to help them move from the preclinical to clinical phase. The ultimate validation is if they then move from the clinical to approval phase with your solution used somewhere in the process. This is another reason why its important to consider the value chart above and ideally try to work with a design partner who is 1-2 years out from a clinical trial.

Proving AI actually improves the failure rate of drug discovery is hardest part of validating AI and everybody is waiting to see how it turns out! Its where the rubber meets the road between the hype and reality. Its also the biggest paradigm shift for AI tool developers who do not come from the world of drug discovery and its where many of them fail. It is a painful shift at first but one that you will come to deeply appreciate and respect over time.

You should not develop your own drug programs but you should run physical experiments that are relevant to drug programs. Being forced to validate your AI in the physical world helps you develop deep empathy with your customers who live in this world day to day. Running these experiments, making these decisions, and seeing 6mos + $500k get wasted on null results because your mice dead from infection is the only way to know what scientists go through as they are developing drugs.

Top of Funnel

Lets say you have built a great AI product / service that can create significant value by helping to develop approved drugs. Now its time to sell! The first aspect of selling to or partnering with pharma is building out a robust top of the funnel pipeline such that you are talking to many different companies every day. The best way to increase top of funnel is to be plugged into a well connected pharma network yourself or through your investors, advisors, or board members. So much can happen in this industry simply because of who you know or who someone close to you knows. However, this is a painfully obvious truth you probably already know (or will soon find out) and does not really help our situation. What should we do if we are technologists or just starting out and don’t have that deep seeded network? Here is a high level laundry list of things I’ve used in the past ordered by effectiveness…

Publish great science with respected scientists
Distribution partnerships - contract researchers (CRO), consultants, device manufacturers, contract manufacturers (CMOs), advocacy groups
Conferences
Hiring the industry insider with 5-15 years in the niche space
Hosting or responding to the viral science of the moment
Social media streams of great science and blogposts
Fostering a niche community of highly specialized experts (meetups, dinners, journal clubs, etc)
Supporting important academic opinion leaders for free
Meticulously mapping everybody you, your employees, your investors, your family, your friends, your employees family, and your investors family know in any context
Cold emails / calls

The Champion

One of the most important aspects of selling to big pharma is finding a great champion you can build a strong, mutually beneficial relationship with. Your champion should be insanely excited about what you are doing and act as a guide for navigating the amorphous enormity of big pharma. The most important aspect of a great champion in big pharma is that they have tremendous impact on advancing their company’s drug programs forward. Remember, all big pharma cares about is developing approved drugs. So having a champion who is core to this will put you in the best position to close a massive deal. There are two ways for company employees to have impact on drug programs: 1) make the decisions or 2) support the decisions. Different roles within big pharma are either making decisions or supporting decisions for drug programs. Lets talk through some of these key roles…

Drug program lead

The people with the largest budgets and most authority at big pharma are those who are leading drug programs. Drug programs are typically grouped based on specific products or on therapeutic areas like neuroscience, metabolism, oncology, vaccines, etc. There is usually a single leader per a drug program or therapeutic area then an executive committee who oversees the entire portfolio of all the programs. A drug program leader is usually working with an executive committee to make major go/no go decisions whose price tags can range from $10Ms to $100Ms. Once the large decisions are made by the committee, the drug program leader and their team usually have significant authority on how the budget is spent. This makes drug program leads extremely powerful champions to have as they can directly buy your offering. Senior scientists, researchers, clinicians, and other members of drug program teams can also be great champions if they are effective at navigating their team. Drug program teams typically want to use the most practical, simple, well known methods to advance their drug program forward. Drug discovery is hard and so they try not to complicate it further with untested readouts. They typically only invest time in novel technology if it is extremely clear how it can be used to advance their drug problem or if they are stuck on a important problem whose only solution is using a novel technology like AI.

Drug program leads typically cannot buy recurring subscriptions and need approvals for spend over a couple million. So a quick way to penetrate the organization and start creating immediate value is to work with your champion to scope a time restricted “fee-for-service” appearing deal to get the ball rolling. These deals are really just subscriptions to the product for a fixed amount of time but they are not recurring and typically have some clear service deliverable relevant to the deal program. Doing these types of deals allows your champion to quickly approve the budget, gets you immediately creating value, and helps your champion build a data package to take to IT/their executive committee for approval of the larger recurring spend.

Platform lead

Providing critical support to the drug programs are a large number of platform teams with each team having its own speciality. Teams are typically grouped by areas like sequencing, screening, high content imaging, xray crystallography, proteomics, clinical data, antibodies, and other specialities which are highly advanced and used by a lot of different drug programs. Many drug program teams rely on these platform teams to advance. A drug program team will typical decide they need to do X (e.g. run a high throughput screen). In order to accomplish X, they go to the platform team that specializes in X. This results in the platform teams receiving large amounts of inbound requests to do X for many different drug programs. I have seen platform teams servicing 15 - 200 different drug program teams at one time. The amount of budget a drug program team has is usually correlated with the amount of drug programs they work with and the impact they have on each program. Platform teams like sequencing, screening, and high content imaging can be core to many drug programs so their budgets can be huge. They typically see many different use cases so are very familiar with the core problems in their speciality and the drug programs as a whole. Platform teams are a lot more willing to experiment with novel technologies because they usually have the extra budget+resoucres for exploratory work and willingness to fail on trying something new. If a new technology can make them 100X better then thats a huge pay out! If it does not work then things continue as normal. This makes platform teams and the leaders of these teams really great champions.

Computational research

Computational research teams are usually a lot harder to sell AI products / services to at the moment. There are a few reasons for this…

Their budgets are typically smaller and mostly focused headcount + basic needs
Their value is measured by their impact on drug programs
They do not have authority over the decisions drug programs make
Their organizations are still evolving within big pharma and learning how they best impact drug programs
They want to provide the computational value to the drug programs themselves, they do not want your product / service to do it for them

Every single time I have sold to a computational research team, the intermediate response is either "this is amazing, I need to convince X, Y, Z program lead to see the value here” or “we are building an internal team to do this so we don’t need you” or “this is amazing but our team does not have the budget”. The state of big pharma right now is that computational researchers are not directly in charge of developing approved drugs and are a service function with little autonomy compared to other platform teams. How computational researcher teams fit into the org chart and drug discovery process is still evolving. We are already seeing the very early stages of change with prominent computational biologists being put in charge of early discovery organizations at a couple of the big pharmas. Still, their deputies leading the various individual drug programs are all not computational researchers. Slowly, the distinction between computational researcher and scientific researcher will blur as they get better tools and more knowledge.

IT

IT departments are in a similar situation as computational researchers when it comes to being a champion for novel AI products. IT departments are typically shopping for very well defined technology problems like data storage, security, compute, etc. They do not have the context on the core problems facing drug programs. It is very hard to convince IT teams that you are solving a problem which is core to the drug discovery process because they just do not deeply understand these problems. For these types of problems, they usually get inbound requests from the drug program / platform teams then make decisions based on what internal teams are pushing for and their budget + top down priorities. They rarely go to these teams and say "you need to use this technology to solve X problem”, the teams usually come to them and say “I need to use this technology to solve X problem”. This makes them less than ideal internal champions.

C-level executives

Very senior executives and individuals managing high level strategy are usually hit or miss when it comes to championing AI products / services. Usually for AI products to create maximal value in drug discovery, they need to be used by the people on the ground with the most context for making decisions on drug programs. Some executives are very tuned into this lower level context and they can immediately see where novel technology can create value on their team. For most, it can be too low level and they need to see it work bottoms up to really understand the value. Even if an executive sees the value, they will need to convince their team to use it in their day to day. I have found that it is much more effective to use a bottoms up approach to establishing executive level buy-in rather than a top down approach. You want the scientist on the ground saying “I need this to do my job to advance our drug program” vs an executive asking their scientist “Do you think this will help you do your job to advance our drug program?”.

The Big Meeting

Once you’ve established a good relationship with your champion and have met with them a couple times, they will likely setup The Big Meeting. Scientists love bringing a bunch of people together to discuss novel results and technology. Your champion will likely do the same and organize a big meeting with 15 - 30 people from their company. These meetings are typically multidisciplinary with scientists, engineers, computational researchers, clinicians, executives, IT, and more. This is a huge opportunity to really shine and make it to the next major step in the sales process.

For the presentation, you almost always want to present a story in the form of a realistic case study for your AI product / service. The story can be real or fake but it needs to be grounded in reality. Ideally, the story is given life through your product or some graphically manifestation of your service. Slides are good to support but the main narrative should be told through a medium which makes the story lively. If you use a product to tell the story then the people in the audience can clearly see how they can also use the product.

Live demos really make an insanely huge difference and are the secret to truly great pitches. Every great sales pitch I have been apart of has been carried by a great live demo. The most visceral customer reactions I have seen have all happened during the demo as a result of letting the product speak for itself. There have been astonished quotes, teams slacking “WE MUST USE THIS” behind the scenes, people jumping up/down, and dumbfounded amazement all from a demo - nothing else. I have found that there is no graph, quote, feature list, or image which invokes the same emotion as a live demo. A boxplot, confusion matrix, or accuracy metric simply does not connect with humans as deeply as seeing the AI work in real product with results that go beyond just numbers.

For these meetings, the key high level things you want are…

More

Live demos and walkthroughs or real stuff
Examples of your product / service being used to solve relevant problems
Areas where your model thrives
Values and ideals you adhere to
Dramatic shares of incredible results influencing a drug program

Less

Metrics like accuracy, r square error, AUC, F1, etc
State of the art comparisons
Feature lists

What you want to push towards your audience is awe-inspiring vibes and clear communication of your values supported by stories and live demos. What you want to pull from your audience is specific areas of skepticism they need to have validated with data. You need to create a dynamic where there is a constant cycle of:

“Wow, thats amazing” → “but do you have data validating X…” → show slide/publication with data to validate X → “Wow, thats amazing” → “but do you have data validating Y…” → show slide/publication with data to validate Y → etc…

You want all the hard data and facts ready to go in your back pocket so you can instantaneously whip them out to nail their question. This is harder than it looks and takes a lot of time and iteration to get right. It is especially hard given the multidisciplinary nature of these big meetings so you need your bases covered from multiple different areas of expertise. The art of these Big Meetings is nailing the back in fourth between vibes and hard scientific data with the audience.

The perfect big meeting for me is:

25-30 minutes of presenting stories to transmit awe-inspiring vibes + values
10-20 questions asking for validating data / facts to support the vibes + values
70%+ of questions addressed instantly and concisely through strong answers and relevant data
15-30 minutes brainstorming ways to work together and next steps

If you prepare well enough to hit these four success bullets then you are well on your way to closing a big pilot! Obsessiveness and learning through iteration are the only ways I have found to craft the perfect meeting over time. Get your reps in!

The Pilot

Awesome job in the big meeting! There was some great conversation and you left feeling like you really impressed their team with your product or services. Thats great! Unfortunately, scientists are inherently skeptical and will need a lot more than a meeting to trust your offering. They need data. Whatever claims you made about your product or services will get scientists excited but that excitement will lead them to start thinking of ways to validate your claims with strong data. This is where a pilot project usually comes in. Pilot projects should address the most important pieces of skepticism that scientists face. These usually fall into one of these themes: ease of integration, accuracy, business impact, and ability to address their specific use cases. A good pilot covers multiple themes of core skepticism at once.

You should always have a pilot template / proposal ready that addresses the key themes of skepticism for your offering. You never want the customer to debrief without your pilot proposal in hand. Having customers go in blind typically leads to more much painful next steps as the customer may not have a clear vision for next steps or may propose a pilot that is way off base. To protect against this, you want to send over a pilot template / proposal within 24 hours of the big meeting. The proposal should have detailed steps, a clear outcome, a clear objective, and a specific call to action for your champion. Here is an example proposal:

We would like to propose a pilot project for you all to assess how easy it is to send us samples, how fast we can process the samples, and the accuracy of our AI predictions. Here are the proposed steps for the pilot:
Start reviewing and finalizing the NDA and MSA with our legal teams
You send us 20 blinded samples with 10 diseased samples and 10 healthy samples
We process and run our AI model on each sample
We send you the predictions for each sample then you can assess whether it matches the label
Champion, if you all are interested in pursuing this pilot, can you please let us know where we can send the legal documents for the review and sample shipping instructions?

After you close the pilot, you need to turn down the sales pitching vibes and turn up the science. That means you let the data do most of the talking. If you try to sugarcoat the data with any hype or salesmenship then scientists will be utterly repulsed and lose trust. If you try to hide any mediocre or bad data, they will find it and lose trust. If your results don’t match your claims in the pitch then they will start losing trust. The only response to bad pilot data is to be completely up front with it, communicate you deeply understand the situation, clearly state why it happened, and lay out an earnest plan to fix it over the coming months. If you are sincere, straight forward, and painfully honest then most scientists will give you a second chance. As much as you may be tempted to pretty up a pig, do not do it. Instead state how you know you only have a pig right now but you are going to kill yourself over the couple months to get some horses.

These pilots are brutally hard but we must face these tests to validate what we build and close the best customers.

The Close

After about 2 months if you’re lucky, 3-6 months normally, and 6-12 if the gods hate you, you will be in the closing process on a great deal! In general, this process is straight forward but wanted to add a couple of points I think are very important…

Ensure that you know exactly what needs to happen to get the deal closed (who needs to approve/sign, what the legal process is, what the bandwidth on legal is, what potential internal blockers can arise, budget cycles, etc)
Jump on a phone call or meet in person with your champion every week to get the full picture of whats going on in real-time
Continue to be patient and balance obsessing over the deal with peace in knowing that everything will turn out alright, but careful not to let too much worry/obsession or too much peace/comfort derail the deal

You are a closer and will get it done. Thank you so much for taking the time to read my blogpost and I sincerely it helps you make a huge impact at these world changing organizations.

Reviewer’s Feedback

Please comment or send me your feedback as I’d love to have this blogpost evolve from just my perspective. I wanted to start the conversation with everything thats in my head but hoping to expand my knowledge with other people’s perspectives. Here is some of the most thoughtful feedback I’ve encountered when writing this blogpost.

Dylan Reid

Sales Cycle - This probably applies to big pharma more than biotech and clinical development more than early discovery but I’ve found the sales cycle for life science software to be super weird and very different from other enterprise SaaS in a few ways:

Annual Budgets - Annual budgets seem to be set earlier in pharma (9-12 months ahead) than other sectors but there's often excess funds that if become available towards the end of the year that go away if unused — creating a long sales cycle for big contracts and a narrow window for pilot programs, and upsells; a little like selling to government.
Program Timelines - As you get into clinical development and are selling to a specific program you become hostage to study start dates and readouts which are often delayed (and sometimes accelerated). That creates a big path dependency you have to navigate (eg. be ready to move fast, but expect delays).
IT / Security - Have found that IT, especially at big pharma is where a lot of contracts go to die, even after all the other stakeholders have signed off, and there doesn't seem to be a single agreed upon standard for security/ privacy among the big pharma companies so getting IT involved early can be helpful and understanding sensitivity should be part of prospecting
Delivery / Deployment - Pharma is so used to buying services (from CROs, Vendors) it can be tough for them to buy software even when the product is clearly self-service SaaS. Hopefully this will change, but for now I’ve found companies that can sell into that framework (eg. package software as professional services, invest in training, implementation and change management) are more successful. It’s a fine line to walk and easy to become a service/ NRE vs. product company, but Veeva (which has a big service arm, but a true software product) has shown it can be done.

Value Creation / Capture - Think this is super important and agree with your framework here, though some of the examples feel a little skewed towards small-molecules and traditional ML drug discovery (HTS, H2L). I have more experience with biologics side and generative/ rational drug design and think there are a few differences - which may or may not be in scope of your article.

Highest Value Problems: While tox and target selection always matter there are some higher order bottlenecks in biologics that are where a lot of the value is being created today: for established modalities like antibodies developability (eg. stability, immunogenicity, specificity, aggregation) matter a lot. For newer modalities like gene therapies - delivery is probably the biggest bottleneck which could look like tox, immunogenicity, specificity (biodistribution) or editing efficiency depending on which delivery vehicle (AAV, LNP, EV, etc.) or approach (novel editor, truncated payload) you are dealing with. These tends to be modality specific but as a general principal, figuring out what the highest value problem in the area you are working is important.
Pricing / Business Model: This maybe out of scope for the piece, but something I think about a lot is whether software based business models (eg. usage, seat based pricing) make sense for life science software companies, especially at the discovery stage where the number of users can be small compared to the amount of directly attributable value technology can create. Shared value or risk based pricing (eg. Upfront, clinical milestones, royalties) are common for platform biotech partnerships, but as more of that work becomes computational - I think there's an opportunity for software companies to explore similar business models.

Building Trust in AI - Think this is super important and doesn’t get talked about enough. Agree with what you have said, but I wonder whether shift to generative AI and foundation models will change how we build trust. A few thoughts / observations:

Predictability: Generative models are unpredictable "stochastic genies" - even with guardrails - and evals only capture a set of possible outputs at a fixed moment in time. In addition to guardrails and grounding, we're going to need continuous monitoring and evaluations for model degradation and emergent properties; along with a way of communicating that to users.
Confidence: Generative models tend to be over confiden and more impressive on first blush than continued usage, meaning it make be easier to build trust but harder to maintain it. Companies will have to communicate the limitations of models to users and scope out appropriate use.
Explainability: The ability to interact directly with models in natural language creates an opportunity to build trust and give users greater insight into how models work and when they should be trusted (eg. how confident are you in this answer, explain your chain of thought for how you arrived at this answer, etc.)

Alex Beatson

What problems create value - I agree that target ID and ADMETox issues are two of the main reasons programs fail and thus two of the highest-value areas to work in. However, there are many drug targets that we can't yet adequately drug, and many startups which develop new technologies in order to drug previously undruggable disease targets. More than a few of these efforts to develop high-value first-in-class drugs can result in multi-billion-dollar outcomes, like the recent Nimbus-Takeda deal. So people might push back on the idea that there isn’t value in early-stage discovery. I think there's a more nuanced reason that AI/SAAS companies can’t capture much value within early-stage discovery: the problem of drugging undruggable targets and developing new therapeutics isn't actually one problem which AI can solve (such as finding a protein structure, or finding binders, or any other computational task), it's actually a whole sequence of problems (including target validation, assay design and validation, prioritizing molecular series to balance potency with patent and ADMET risk, etc), and each AI product will more or less only help with one sub-problem, generating limited value. For this and other reasons (algorithms or models are not a good moat without proprietary data, particularly sold as SAAS), it seems to be an immutable law of physics that AI companies in early-stage discovery who try to be SAAS companies all eventually pivot to being therapeutics companies running in-house drug programs.
Communicating confidence - It’s important to consider how you communicate a model’s confidence / error bars / level of trust. If you predict ADMETox or potency, the scientist needs to know how much they can trust that prediction. A prediction that has some nonzero error is fine if you know what the error bars are. For example, if the model predicts a molecule is bad the scientist can make a decision whether they should synthesize a molecule anyway (because it's promising in some other ways and the bad prediction is weak confidence), or whether they should avoid spending the $ and time to synthesize the molecule all together (because the prediction is sufficiently high confidence).
Pricing Pilots - One question is how pilots should be priced. From my experiences in drug discovery, I’m in favor of being mercenary about setting high prices, because I've seen both how important upfront payments from discovery partnerships with pharma can be to AI discovery companies' finances and how unlikely renewal of even partnerships that go well is, but things might be different in other areas of AI for pharma.

Leonard Wossnig

Concerning Start-up Messaging to Big Pharma: The examples one is using in a presentation should illustrate how your company can impact big pharma's main goal: developing approved drugs. Even catchy statements like "drugging the undruggable" need proof of success. Many companies, such as Charm, focus on clear targets and support their pipelines with proprietary and differentiated software. Pure software plays are rare in successful deals, especially large ones.
Specificity in Building Business Relationships: I agree that more specific details, such as pipeline steps or indications, can strengthen business relationships by enabling connections with an internal champion who can identify and promote a concrete application.
Creating and Capturing Value: A critical missing element in the idea of "Creating and Capturing Significant Value" is the escalating investment needed at each stage. Even doubling success rates (e.g., from 4% to 8%) may still be a gamble, and investors looking for substantial upsides may be dissuaded by the inability of software to capture value like tangible assets unless it revolutionizes trials' PoS.
Early Stage Business Models: While later stage impacts are undeniably more impactful (e.g., patient selection impact on PoS), the combination of software with unique experimental setups, techniques, and data generation in the early stages can still create a defensible difference and advantage. Commercial success may follow the trend set by companies like Entos, LabGenius, and Charm (see prev.), who have integrated software solutions with lab development and drug pipeline construction.
Building Trust in AI: Transparency in training data is paramount. Lack of these details often leads to the rejection of pitches, claims, and papers, even published ones. Challenges remain in areas like protein folding where data leakage is still an issue since similarity is not clearly defined. The minimal improvements on benchmarks, such as a 3% improvement on AF, are often insignificant to drug discovery but machine learning practitioners and startups overstate the importance to get the papers in journals or get funding.
ML/AI vs Traditional Approaches: The community needs to ensure fair head-to-head comparisons between AI and time-based methods, avoiding overfitting or improper optimization. Proper metrics and relevant considerations are often ignored but are essential for a balanced view. This is particularly challenging if public benchmarks are used, since these are pretty useless. Pat Walters wrote just another great blog post explaining why this is the case.
Ease of Use and Target Audience: The importance of ease of use is contingent on your audience. For replacing systems like the Schroedinger suite, it's vital. However, for expert users, a library might suffice. I think understanding your audience is key.
Importance of Connections in the 'Top of Funnel': You could emphasize that connections are paramount in the industry which is due to trust. Pharma Directors, VPs, EVPs, etc., continually approached by companies, often rely on trusted connections to filter viable opportunities among the many they come across. So it's in some ways a natural behaviour.
Selling to C-Level/Portfolio Executives: While engaging executives can be helpful, most executives need to rely on internal opinions (of people who work on the ground and know the tech) to make decisions. Thus, having an internal champion who is interested in your product is often essential for a successful pitch even if you have exec buy in.

Thank you to Nan Li, Jay Rughani, Jacob Effron, Tony Kulesa, Pablo Lubroth, Siyu Shi, Shervin Ghaemmaghami, and Carl White for taking time to review this piece!