Make Dents in C̵a̵n̵c̵e̵r̵ the Universe

Do you want to make a dent in the world’s most lethal cancers? We propose four ways to do this below. Do you want to go further? Do you want to tackle more of the world’s biggest problems — like climate change and habitat destruction? We consider this also.

In April 2020, Simon and I set out to make a dent in the world’s most lethal cancers. We were not alone. We were helped by Jack, James, Joanne, Sebastian, and so many other kind experts. Emilie, Evgeny, Faisal, James B, Evis, Mireia, Colin, Richard, Jess, Andrew B, Andy G, Andrew S, Bram, Simon H, Vegard, Barry and Pierre, thank you!

We sadly failed in our mission. Outcomes for these cancers remain unchanged. However, if there are ways to make a difference, then we believe that we have identified some of these:

• Roivant for impact-investment markets;
• personalised vaccine development as a service;
• Scale for sensitive data;
• AI guides for scientific intuition.

We introduce these four ways below. We hope that by sharing our journey others may succeed where we failed. Of course, if other teams would like our help, then we are here for you.

In our journey, we also learnt that a funding gap exists for the world’s biggest problems. It has been said that the world’s biggest problems present the world’s biggest business opportunities. However, the brightest minds of our generation are still not working on these problems. We believe that this is because of the way funding is optimised. We share three proposals by Nielsen and Qiu to address this.

Why Fight These Cancers?

We set out on this journey because this is one of the world’s biggest problems. There is also no solution in sight.

Cancer is the world’s biggest killer. Cancer kills ten million people per year. It is also the fourth biggest killer of children. Of the broader cancer types, the most lethal are pancreatic, lung, oesophageal, brain, stomach and liver cancers. In countries with the most sophisticated healthcare systems, these cancers have fatality rates of 83–95%. Collectively, these lethal cancers are also responsible for 55% of all cancer deaths.

Big pharma is unlikely to solve this problem. As explained by senior biopharma executives, “Pharma follows the science, and the science isn’t there yet. However, this isn’t just about pharma. Even academic researchers have to follow the science. To do otherwise would be to commit career suicide.” Biopharma’s relative disinterest in the most lethal cancers can be seen in this table. It shows the number of commercial trials per fatality across cancer types.

Newer health-tech companies like Tempus and Flatiron are also unlikely to solve the problem. This is because these companies are service providers that optimise existing pharma operations. No new path is being followed. Tempus optimises the existing drug-approval process. It does this by optimising the trial-matching step of the process. Flatiron optimises clinical adoption of approved drugs by providing an analytics platform for real-world outcomes research.

Three Categories of Approach

We explored three broad approaches to make a dent in the most lethal cancers:

• turn treatments into trials;
• turn healthcare data silos into AI healthcare ecosystems;
• turn scientific papers into new scientific hypotheses.

Within these categories, we found four specific strategies. We introduce these below.

Turn Treatments into Trials

The first approach that we explored was to give every patient with a lethal cancer the chance to trial a new treatment option. Fewer than one in twenty patients with these cancers get to join a trial today. The remaining nineteen receive ineffective chemotherapy or radiotherapy. They are left hopeless, with no chance of life. We set out to change this.

Why Now?

Recent technological advances suggest that now is the time to do this. Specifically, we believe that recent advances in personalised cancer vaccines are the key driver.

Championed by pre-eminent researchers like the Broad Institute’s Professor Ott (Harvard and MIT), personalised cancer vaccines are designed to induce a T cell response in patients that is specific to each patient’s tumour. They are a third type of immunotherapy, alongside immune-checkpoint inhibitors and CAR T cell therapies. Recent data suggests that many more treatment options will soon need to be trialled across many more patient cohorts.

Four types of data indicate that we have now reached a technological tipping point. Firstly, recent trial data shows that vaccine developers can now achieve secondary-outcome goals in a safe and effective manner. Specifically, developers can induce tumour-specific T cell responses, where these T cells penetrate the tumour and kill cancerous cells. Importantly, this is true even for patients whose cancers have the fewest target mutations. These cancers include glioblastoma. This is arguably the very worst of the worst cancers. Secondly, recent trial data is already providing anecdotal evidence of primary-outcome (survival) success. In particular, this is when personalised cancer vaccines are combined with immune-checkpoint inhibitors. Trials are now being expanded to focus on survival. Thirdly, concierge oncology clinics are also achieving anecdotal but profound survival improvements using personalised cancer vaccines to treat ultra-wealthy patients with lethal cancers. This is something that our team saw firsthand. Fourthly, and this is something that happened while we were exploring the approach, researchers have achieved unprecedented primary-outcome success repurposing the approach of personalised cancer vaccines to fight Covid. This real-world success has fundamentally changed how pharma now views this technology.

Problems to Solve

Three problems need to be solved before we can turn treatments into trials for all patients with the most lethal cancers:

• access to the required drugs;
• money to pay for the trials;
• trial adoption by healthcare providers and patients.

Of these problems, the first one is key. Pharma companies refuse to make their experimental drugs available to third-party trials. This is because poor design or management of these trials could lead to misleading outcomes or notable adverse events. This would, in turn, jeopardise the drug’s standing with regulators or payers.

The second problem hinges on the first. Pharma companies will only fund trials if their own drugs are being tested; and if anyone else is to fund the trials, then this also requires access to suitable drugs.

When we started on this journey, we thought that the third problem would be significant. We believed that it would be difficult to get healthcare providers to adopt trials as treatments. However, the path to solving this now seems clear. We heard at the outset that it takes too long for patients to enter existing trials for the most lethal cancers. For example, we heard that the majority of eligible patients die before they join a leading trial for pancreatic cancer. Since then, the teams behind the Tessa Jowell BRAIN MATRIX trial, the pandemic’s RECOVERY trial, and Medable have shown that this is a solvable problem. BRAIN MATRIX and RECOVERY have shown that healthcare leaders, clinicians and patients are now willing to integrate trials directly into the clinical path. They are also willing to do this at a national scale. Medable has shown that tools can be built that bring trials directly to the patient and the clinician, rather than requiring that the patient go to a distant trial centre.

Two Strategies

Our investigations led us to identify the following two strategies to solve the drug-access problem:

• Roivant for impact-investment markets;
• personalised vaccine development as a service.

Roivant Sciences is Europe’s second-fastest startup to a $5bn valuation. It is a crowdfunding platform that gives pharma’s “unloved” drugs another shot at a clinical trial. We term drugs unloved if their owners no longer deem them strategically important, or if they have exhibited some exceptional trial responses but not enough good responses. Roivant has shown that pharma companies are willing to hand these drugs over in return for a share of revenues if Roivant can get these drugs to market. Roivant packages these drugs up into subsidiary companies (“Vants”) that it lists on public stock exchanges to raise money for the required trials.

While Roivant continues to make strides, unlocking potentially life-saving value that would otherwise have been left on pharma shelves, the company will not be making a dent in the most lethal cancers. This is because Roivant, like traditional biopharma companies, necessarily follows the science. It must do this because it raises funds from the stock market.

This brings us to our first strategy. We believe that a team can make a dent in the most lethal cancers by bringing Roivant’s crowdfunding approach to the impact-investment markets rather than traditional stock markets. Impact investors are investors who want to do well by doing good, but it’s the doing good that dominates. They are happy to lose for an exceptionally good cause. Impact-investment markets are already worth $715bn, and they are fast-growing.

Three key activities would be involved in bringing Roivant’s approach to impact-investment markets. These are: identifying suitable drugs; doing deals with the biopharma companies; and packaging drugs up appropriately for investors. Given this focus, we believe that the right team would comprise industry insiders or financial analysts that specialise in drug value.

The second strategy was born out of Covid . It builds on a step change in economic fundamentals resulting from the pandemic. Before Covid struck, only a small number of teams wanted to trial personalised cancer vaccines. This left each team managing all parts of vaccine development themselves:

1. target discovery;
2. vaccine formulation — using some vaccine platform (RNA, peptide, dendritic cell, DNA or viral vectors);
3. administration — spanning dosing and sequencing (in particular, with immune-checkpoint inhibitors);
4. assessment of immune response.

Covid changed this. The explosion in demand for these types of vaccines means that there is now room for a third-party provider of personalised vaccine development as a service. We expect such a service provider would initially focus on vaccine formulation. However, the most important contribution would be a self-learning tool for target discovery and administration that uses every assessment of immune response after administration to refine programmatic target discovery and administration. We anticipate data network effects driving substantial long-term defensibility for such a tool.

The cost of such a service provider is not yet clear. However, small teams have been developing vaccines themselves. This includes at least one concierge oncology practice. Given this, we are confident that a third-party provider could make personalised cancer vaccines available in a cost-effective way to nationwide investigator-led trials like the BRAIN MATRIX that are crying out for therapeutics. If we couple this with the fact that the first immune-checkpoint inhibitors will soon be coming off patent, thereby permitting cost-effective trials of vaccines in combination with immune-checkpoint inhibitors, we are optimistic that this strategy could make a dent in the most lethal cancers.

Turn Healthcare Data Silos Into Healthcare AI Ecosystems

The second approach that we explored was born of our experiences at Google. “The best minds of our generation are still working in support of ads. That sucks!” (Apologies to Jeff Hammerbacher for the tweaked quote.) We wanted to change this. We wanted to give life-giving purpose to our brightest AI minds. Specifically, we wanted to make it possible for our brightest AI minds to extract life-giving insights from sensitive healthcare data that they cannot see.

Why Now?

We heard that big pharma is calling for Scale to be applied to clinical healthcare data. A recent step change in enabling technology suggested that now was the time to do this.

Scale is a service that gives superpowers to human labellers of data. It works across text and imaging data, but it is across imaging data where Scale has had the greatest impact to date. Specifically, Scale’s image labelling is powering today’s autonomous driving revolution. Scale gives human labellers superpowers through a type of machine learning called active learning. This involves machines first pre-labelling and pre-segmenting images from car cameras. The pre-labelled and pre-segmented images are then given to humans for validation or correction. The machines then use this feedback to learn and improve for the next iteration. In time, the improving accuracy of the machines means that humans only need to consider the most difficult and most interesting cases.

As transformative as Scale has already been for autonomous vehicles, Scale cannot be used today across sensitive data. In particular, Scale cannot be used to help radiologists label medical images. This is because sensitive data cannot be shared with the Scale service. This is especially frustrating to big pharma. After a pharmaceutical company achieves regulatory approval for a drug, the company still needs to grow adoption by payers and clinicians. The best way to do this is to show real-world successes of the drug, and the gold-standard way to do this is with RECIST1.1 labelling of imaging data following treatment. Unfortunately there is a critical shortage of radiologists and so even Flatiron Health, the very best clinical data repository, does not have the RECIST1.1 data needed to accelerate big pharma’s go-to-market strategies. For other data repositories, there is no chance of RECIST1.1 annotations.

Oasis is a new technology provider that promises a fundamental change. By design, Oasis enables services like Scale to learn from data that is never shared with the Scale team. Specifically, it provides a sandboxed hosting environment for Scale that has the following properties. Firstly, the environment is provably secure. This means that we can prove that no data is ever shared with the Scale team (or anyone else). It also means that we can prove that Scale’s code does not touch any other code; and we can prove that the code is not accessible by anyone. Secondly, the environment is attestable. This means that we can prove that Scale’s code executes as intended across the data. Thirdly, it is auditable. This means that all activity is transparently logged for billing and access control.

Strategy

Building on the reasoning above, our primary goal was now to build a healthcare-grade middleware layer for labelling services like Scale. This middleware layer would provide fine-grained control over where data is processed; and it would provide fine-grained access controls. Learning from the missteps of Google, all activity would also be made transparently, immutably auditable.

Although we saw this as a giant step forward, we also believed that much more was possible. Specifically, we believed that the following was possible:

• bring AirCloak to sensitive healthcare data;
• bring data scientists and AI practitioners to healthcare data;
• bring AI services to healthcare data.

Let’s consider.

As things stand today, the biggest clinical data repositories expose raw data to researchers. This leaves them exposed to Anderson’s Rule. According to this rule, the scale of the repositories means that they can no longer improve functionality while also maintaining security. By pursuing new functionality, commentary from another leading security researcher is also starting to bite: “Federation is the new de-identification.” In just the same way that de-identification was never a silver bullet for protecting privacy, so “federated” analysis is also not a silver bullet. This is particularly true for the loose way that some startups are using the word.

Our plan for solving this problem was to block researchers from having direct access to the raw data. Instead, we would give researchers indirect access by wrapping AirCloak into the new secure middleware layers. This indirect access would give researchers analytics from which they would not, in practice, be able to infer any raw data.

Having brought safe analytics to sensitive healthcare data, the next goal was to transcend the Five Safes and provide provably safe use of the data by anyone. This particularly lofty goal requires that we give data scientists and AI practitioners the following:

• a way to inspect the shape of the data, while not having access to the raw data itself;
• primitives for data processing and AI algorithms that are expressive enough to generate the required insights, while at the same time not allowing anyone to take the insights and infer specifics about the source data.

OpenSafely’s efforts are perhaps most notable in trying to satisfy the first of these requirements. OpenSafely generates synthetic data that it exposes to data scientists. It also provides high-quality data dictionaries. We agree with OpenSafely’s approach, although we believe that it is possible to produce more representative synthetic data than is currently exposed. We believe that this can be done by baking differential privacy into the generation of this data. We see this as a tractable problem because synthetic data would only need to be generated periodically. Humans can also audit these data sets before they are exposed.

The second set of requirements is what makes this goal particularly lofty. It involves taking key primitives, like MapReduce and stochastic gradient descent, and reformulating these to provide a mathematically provable guarantee that inference is not possible. Although the OpenDP team is making good progress, this remains a very challenging problem.

If the Five Safes can be transcended, the next goal would be to build a healthcare-focused version of Fred Ehsam’s proposed AI Marketplace. We see this as a straight-forward extension of the work above. The smart-contract infrastructure would be extended to enable data owners and algorithm developers to build mutually owned AI models as services. Such a marketplace would allow any data provider to participate. It would allow any AI practitioner to participate. If it is designed in a way that encourages the development of compositional, modular AI services, it would also turn today’s healthcare data silos into a vibrant, impactful AI healthcare ecosystem.

Why Not Yet?

During the course of our investigations, we learnt that our strategy was a couple of years premature.

Firstly, contrary to our early expectations, the enabling technology for secure computation is not yet healthcare-grade. Oasis currently relies on an AMD chip that does not allow customers or partners to verify that Oasis cannot access data. This means, in particular, that Oasis cannot currently satisfy the Data Governance Act. AMD has an updated chip that will remedy this, but it has not yet been deployed by commercial cloud providers.

A temporary workaround is possible using Intel’s chip. The Secret Network, for example, makes this possible. However, the cost of this workaround would be high as this chip would require all code to be rewritten in Rust. The benefit would also be short-lived. Indeed, the benefit may not even be significant. This is because GPUs may be required to process the images, and this is not yet possible.

Secondly, the best partner repositories do not appear to be ready. The strategy that we put together was endorsed by a large, long-standing biopharma customer of one of the largest repositories. The Chief Commercial Officer of this repository also gave the plan his backing. Disappointingly, the repository’s operational leads are not yet ready to prioritise this work.

Turn Scientific Papers into New Scientific Hypotheses

The third approach that we explored was whether an AI-powered guide could improve the intuition of human scientists, and thereby accelerate scientific discovery. Our thought was that a Scale-like approach could surface unknown patterns in the scientific literature, and we suspected that these patterns could be used to guide the intuition of scientists researching cancer vaccines.

Why Now?

Unknown to us, a team at DeepMind had a similar thought to us. And although their focus was on guiding mathematicians rather than scientists, their early results from December 2021 are encouraging. This is at least one reason for thinking that now is the time to do this.

The more important reason, however, is the stage of development of personalised cancer vaccines. As detailed by Douglas Hanahan and Andreas Bender, researchers in molecular oncology already accept that they need to take the whole biological system into account. This view is amplified among scientists working on cancer vaccines. Firstly, these researchers are advancing from a focus on secondary outcomes (inducing an immune response that kills some cancer cells in a tumour) to primary outcomes (killing all cancer cells in a tumour). Secondly, early trial data has left researchers with many system-wide questions. Thirdly, early trial data suggests that personalised cancer vaccines will be most effective when used in combination with other treatments, and this compounds the system-wide complexity.

The following are just some of the system-wide questions thrown up by recent trial data. Researchers are finding that CD4+ T cells are being unexpectedly induced during trials. Have researchers who are focused on different diseases or therapeutic types seen this happen? If so, what are the roles of these CD4+ T cells? Researchers are finding that immune cell populations are being induced with different phenotypes and different distributions. What can be learnt from other settings? Researchers are finding that Treg cells are being induced. These Treg cells appear to have an immunosuppressive influence in the tumour and in the microenvironment. In which other circumstances does this happen? Researchers are finding that the cancer cells that are not specifically targeted by vaccine peptides sometimes lead to tumour outgrowth. Does this always happen? Does the relationship matter between the chosen peptides and the hallmarks of cancer? Or does the relationship matter between the chosen peptides and the expression of genes in signalling pathways? Can we use these pairwise relationships to combine peptides synergistically, rather than simply combining peptides that are currently considered independent of each other?

Problems to Solve

We believe that three key problems need to be solved if an AI-powered guide is to have a material impact:

• get human domain experts to annotate;
• create ongoing value for customers;
• capture a sufficient share of the value created.

Scale uses money to incentivise their community of labellers. This approach is unlikely to work here. Instead, we need an approach that aligns with the interests of highly qualified domain experts. We believe that, with a little help from a recent Google paper, CLEF holds the key. CLEF provides academic researchers with large, annotated datasets. To gain access to these datasets, researchers must first annotate additional datasets. With a view to getting the first academics interested, the recent Google paper showed that a relatively simple technical approach can seed a set of useful cross-paper scientific links without any human input.

We tend to view the second problem through the lens of the classic DIKW pyramid, as shown below. If an AI-powered guide is to contribute meaningfully, then the ongoing human-machine interactions need to move up the value chain from describing to explaining to predicting, and this needs to happen continuously. DeepMind managed to achieve this in a restricted mathematical domain. Can the same be achieved here? Perhaps, but we know this to be a hard problem.

The third problem reduces to a problem of attribution. Customers will only pay appropriately, if they can attribute newly created value to the enabling service. This is difficult here because developing a successful new drug is like finding a needle in a haystack. Most hypotheses do not matter, but some are worth billions of dollars. This led a senior pharma executive to explain to us that he does not need more hypotheses. What he needs is a better conversion rate from hypotheses to regulatory approval.

Why Not Yet?

Frustratingly, we stopped pursuing this approach because we could not find a path to capturing value. We recognised the many challenges in trying to create value for customers — both at the outset and on an ongoing basis. However, these challenges would not have put us off, had there been a clear path to value capture. Unfortunately, without such a path, we were left unable to see how real-world impact might be scaled.

Fellowships & Prizes to Make Dents in the Universe

It was in November 2021 that time ran out on our cancer work. Funding was sadly exhausted. In this section we explore what this means for all the world’s big problems. We also share three proposals to improve things.

When it comes to the world’s biggest, hairiest problems, Peter Thiel laments: “[They] promised victory over cancer in six years’ time. They promised us flying cars. All we got was 140 characters.” (Apologies, Peter, for combining your quotes.) Peter argues that we have become masters of relentless, incremental progress, but this has come at the expense of fundamental step-change progress. He argues that incrementalism is reducing our appetite for risk, and this is causing fundamental stagnation.

We agree. We also believe that this is the rational outcome of a funding/incentive gap. The biggest, hairiest problems are the ones in which nobody knows how to make a dent. Nobody knows what to build. Nobody knows which incentives to align, and how. Nobody knows what to scale. Frustratingly, all of this means that these problems are a poor fit for traditional funders. Nonprofit and public funders optimise for specific product deliverables (build X to this specific brief), but this does not help if we do not know what to build. Venture capitalists optimise for rapid, defensible, near-boundless scaling, but this does not help if we do not know what to scale.

In this post, Nielsen and Qiu suggest that scientific funding has a similar issue with big, hairy problems. They reflect on how scientific funders are increasingly focused on specific metrics of progress, like improving publication output, or improving publication output of a particular type (e.g. in the top 1% of citations). They go on to show that optimising in this way reduces the number of outlier discoveries. This matters — as per Peter’s lament — because “performance in science is plausibly dominated by outlier discoveries, not typical discoveries.”

Nielsen and Qiu propose three ways to address the funding gap. We share these below:

• Manhattan Projects for X;
• Thiel Fellowships, but supersized;
• XPRIZE, but scaled.

Taking inspiration from the Manhattan Project, their first proposal is to build funding and branded prestige around each big, hairy problem. This has clearly worked to devastating effect in pursuit of the first nuclear bomb. However, the approach is suitable not just for technically challenging problems. Parley has shown that this approach can align the interests needed to make a dent in ocean plastics.

Taking inspiration from the Thiel Fellowships, the second proposal is to build funding and branded prestige around fellowships aimed at the exceptional people most capable of making a dent in the universe. These fellowships should target the world’s leading lights — not just exceptional talents who are only just starting their careers. Fellows would be given space for deep, distraction-free thought. They would also be encouraged to participate in structured collaboration, using approaches that have proven successful elsewhere. The level of fellowship funding should reflect the calibre of person being targeted. This should not mean that Google is necessarily outdone in terms of financial incentive, but it should not be irrational for someone to turn down an offer by Google.

Taking inspiration from XPRIZE, the third proposal is to use insurance to aim higher with each prize. The idea is that sponsors take insurance out against very difficult challenges being met. “The more unlikely the outcome, the higher the prize-to-premium ratio, and the more prize money you can award for a given premium.” It is not clear from Nielsen and Qiu’s post whether they expect all insurance payouts to go to creators. However, if Nielsen and Qiu allow for some payout money to go back to the sponsors, then we believe that this could significantly scale the impact of this proposal. Specifically, it would change the type of sponsor from charitable benefactor to impact investor, and there are many more of these. If they allow these impact investors to collaboratively pay premiums, through a type of crowd sponsoring, this would scale the impact still further.

Thank You!

Thank you again to all the kind souls who helped us. To Jack and James, especially: you helped us to dream, and, at every difficult turn, you were always there for us. Thank you! We hope we can pay this forward.

To all the crazy dreamers reading this, we hope that you found a nugget of something useful. We hope that we have helped you in some small way to aim higher, and with more support. May you make many dents in the universe. We’re cheering for you!