On the End of Cancer

On one hand, I love the bullish optimism of ML advocates who cheer "we're going to cure cancer" because, yes, we can in fact realize that technical goal and ML can indeed help get there faster. On the other hand, the goal keeps getting repeated without fleshing out what that means. While the intent is good, the lack of details eventually piles up and risks coming off as insincere. Since I would in fact like to use to ML to help accelerate the end of cancer, I figure it might help to carve out some nuance on what ending cancer may look like and some ways that ML can help us get there faster. Note that what follows is not meant to be authoritative or comprehensive, it's just quick notes on a good future that's starting to loom over the horizon, the one you can start to get little glimpses of at the American Association for Cancer Research or American Society of Clinical Oncology conferences with the NeurIPS-level rush of abstracts and promising results on everything from driver genes to cell therapy production techniques. There is a good future waiting to be built here, although not every bottleneck standing in its way is technological--some are sociological in how we deliver care and evaluate new drugs, but it's nonetheless coming into view and immanetizing into reality so long as we keep up the widely distributed research efforts already underway.

First of all, let's establish a few foundational facts. Humans are not going to stop developing cancer. Not even if we get crazy longevity pills and juiced-up immune systems--it'll just get rarer and weirder at the edges of population distributions. However, we are winning the war on cancer. It is a slow uphill grind but we are moving forward and we're not going to stop now. Through elimination of major environmental hazards (smog, cigarette smoke), earlier detection (routine screening, at-home tests), introduction of more effective therapies earlier in the course of treatment (immunotherapies), spread of various molecular diagnostics (oncogene panels, whole genome sequencing, longitudinal measurements in blood), and rise of Comprehensive Cancer Centers anchoring expanded clinical trials, more patients are being diagnosed early, treated effectively, and living free of relapse than ever before. These trends have led to astonishing results in breast, prostate, lung, colorectal, skin, head-and-neck, and blood cancers while some types, such as pancreatic, brain, and ovarian, remain stubbornly resistant to ongoing advances. Progress is occurring along a jagged frontier: more cell therapies in blood cancers, more targeted molecular therapeutics in lung cancer, more diagnostic power in colorectal cancers. Occasionally walls are hit, like generalizing cell therapies to solid tumors, identifying which melanomas will respond to an immunotherapy vs which will experience hyperprogression, or the long tail of rare cancers ranging from osteosarcomas to clear cell carcinomas of various organs. Nonetheless, we are making progress, that progress is acccelerating, and eventually we will win and regard developing cancer as about as serious as having a weird rash.

So let's run the clock forward and ask: what does winning actually look like? Let's jump forward a few decades to a hypothetical patient, Robert, 73 years old, diagnosed with stage II localized lung cancer. It was detected earlier than it would have been before, a few small scattered nodules in the lower lung that had gone unnoticed except for stubborn fatigue, picked up on a routine peripheral blood test that Robert took at the same time as his annual blood lipids test. 3 key markers showed a rise in cancer-associated immune clearance processes and activation of 2 oncogenes, which led Robert's physician to order a full torso PET scan with a tumor-specific tracer. A radiologist aided by ML-driven image segmentation and comparison of Robert's scans to millions of similar scans spotted the few nodules lurking deep in the lung parenchyma--spots that would have previously been passed over as normal lymph nodes. Upon the radiologist's finding, Robert's primary care physician orders a full peripheral blood screen, which arrives the same day as a package on Robert's door step. The test is easy, a brightly colored mechanical vampire that he latches onto his upper arm and, after a brief sting of a needle, absorbs some blood and processes it, ready to ship out to a lab. The lab receives the blood next day and processes it: full genome sequence for background on normal cells, full spectrum panel watching out for cancer-associated markers, measurement of different leukocyte subsets and their states, detection of cell-free DNA with annotation of which was likely to arise from cancer, and a deep metabolic panel. All of this information is ready and waiting at Robert's first visit with an oncologist the next day, which can now be done virtually but Robert opts for in-person, where, with the aid of a mixture of expert system trained across millions of cancer patients' aggregated and de-identified data, the oncologist walks Robert through a detailed portrait of his lung cancer. 2 different main populations, one KRAS driven and the other TP53-driven, both likely to be EGFR inhibitor insensitive. An active immune response but the monocyte compartment is showing signs of exhaustion. Local lymph node spread but no distant metastasis yet. Pretty routine, according to the oncologist, and great chances for survival--the biggest worry is, as always, tumor evolution towards resistance, but now we can monitor that proactively across several timepoints, checking in every week to read whatever signals we can to measure treatment response vs resistance and aggressively run ahead of the tumor to change drug regimens that pin it down in a vulnerable state: weaker, less adaptive, easier for the immune system to clear out.

The oncologist orders a monitoring appliance and a drug package, both of which arrive at Robert's home the very next day. The monitoring appliance is the size of a coffee maker and the drug package is a thin flexible sleeve with a reservoir compartment that finds and taps a vein in Robert's arm. He just needs to wear it for 2 hours a day: it infuses a custom pack of drugs--refreshed via pharmacy delivery--at specific combinations and rates while collecting blood and tumor cells that get fed back into the monitoring appliance each night when Robert takes off the drug package device and plugs it into the monitoring device to recharge, download biological samples, and sync the treatment plan with the oncologist's office. 3 days have passed. Robert is already on a complex regimen of immunotherapies and low-dose chemotherapy delivered in a staggered schedule to weaken the tumor and prime the immune system. Every day, the devices communicate back to the oncologist for monitoring and, as needed, revise treatment plans with the support of a comprehensive medical ML system. New drug packs for the sleeve are shipped automatically. Robert will need to come in for scans every few months and, if the tumors grow and spread, he may need laparscopic surgery to remove as much of them as possible, albeit now with the introduction of localized, long-acting chemotherapies left behind in the surgical site to suppress regrowth from a cancer cell that may have been missed. The oncologist doesn't think this is likely, however, as both she and the medical expert system she uses have seen complete responses with no relapse in 97% of similar cases recently. He doesn't feel too bad, he isn't losing appetite or hair, and he's optimistic about his prospects. As just part of life at 73, he knows several friends who have been through the same process before it was shrunk and miniaturized for easy home convenience, and all but one of them had survived. That one friend had had it bad: an aggressive glioblastoma that surgeons weren't sure could be removed without causing massive impairments to quality of life (after all, neural tissue regeneration is lagging oncology care by a decade and is especially risky in cancer cases), so they had said their brave goodbyes and made a few good memories before they left. Robert doesn't worry much about this, however, as the localized disease and optimism of every doctor he's seen give him the confidence to plan his next backcountry fishing trips anyway. As a precaution, Robert does call his immediate family and tell them the diagnosis: each of them sends a message to their primary care physician's ML proxy requesting the latest at-home tests and they go through the same initial diagnostic process that Robert did: deep inspection of blood for cancerous or pre-cancerous signals, and with a sigh of relief: nothing found, everyone safe.

Over the course of the next year, Robert's oncologist adjusts his therapy regimen three times to adapt to changes in the underlying cancer: first the EGFR insensitive population is eliminated, but the KRAS mutant survives a bit longer, only to rebound briefly after a 90% reduction with both the inciting KRAS mutation and an EGFRvIII translocation that sustains growth signals in the cancer without any external endpoint. At each point, the oncologist has an arsenal of options for sequences of drugs to administer that block the cancer's signaling adaptations, both slowing its growth and denying it the opportunitity to adapt to treatments over time, in turn giving the immune system an immunotherapy boost to help eliminate all the cancer cells it can. These changes are successful and after 6 months, Robert's cancer is undetectable: not a trace on imaging scans or blood tests. For the next 6 months, the sleeve administers a cocktail of drugs designed to prime the immune system against any potential relapse while monitoring for any re-awakening cancer cells that may have gone dormant in the bone marrow. After that period with no sign of recurrence, Robert is declared all clear with the caveat that more frequent monitoring tests will be done for the next 5 years to catch the cancer even earlier if a few dormant cells manage to start a recurrence. However, if that happens, the same rapid loop of testing, characterization, and treatment is waiting for Robert, continuously improving with his and other patients' data to expand the arsenal of molecular diagnostics that tell us about cancer state alongside the suite of effective drugs that can be used to treat any state that cancer can muster. This is what the end of cancer looks like: early detection, personalized adaptive treatment, and a fast observation-orientation-decision-act loop between the patient and their doctor (and the doctor's systems). It will not work for everyone all the time, there will still be rare aggressive cases, especially on the population distribution margins if longevity treatments pan out, but it will work well and quickly for the vast majority of patients to deliver rapid cures tailored to their cancer and its evolution over time.

Every one of the pieces for that curative loop is currently available in some form (Table 1). Some are well-developed, like the expanding suite of targeted molecular therapeutics (gefitinib, trametinib, sotorasib, basically anything where we have a commonly-mutated gene and a drug to hit it), while others aren't yet ready for everyday use, like longitudinal molecular diagnostics to monitor tumor evolution in the course of routine clinical care. Some of the bottlenecks are technological, like assigning cell-free DNA origins to a tumor vs a normal cell, whereas others are more regulatory, such as getting a companion diagnostic approved for clinical use. But every single one of these issues is technologically tractable and worth aggressive development to turn current academic and startup prototypes into robust scaled technologies that can be deployed far beyond the logistical footprint of Comprehensive Cancer Centers to clinics everywhere. Achieving this vision will not be easy, may not be fast, and it may even turn out that ML helps accelerate the social coordination problems more than it does the technological challenges involved¹. Overall, there are 5 main places I expect to see ML systems making major impacts in this challenge, none of them in target prediction or protein design:

1. World-class diagnostic interpretation and treatment planning at scale. Great radiologists, pathologists, and oncologists remain rate-limiting with specialized cancer centers offering vastly superior care to that available from other clinics. Building systems that let world-class human specialists help stage and treat tons more patients across a lot more locations is a massive lever to raising the floor on cancer care quality across the board.
2. Biomarker development. Knowing which analyte in a panel is worth the investment for a full diagnostics development is tricky and ad hoc. Pooling data across studies and cancer types to help identify even subtle signals with high clinical impact is a massive impact, particularly if we can pre-register prognostic utility and robustness by confounding factors. This would fit hand-in-hand with streamlined regulatory approval paths to develop molecular biomarker tests to make it more economically attractive for small and large players alike.
3. Trial matching and acceleration. Identifying which patients are most likely to benefit from an experimental therapy is bottlenecked by patient visibility in systems capable of conducting experimental trials, especially academic-run trials that aren't acting on already-validated mechanisms. This fits in with (1) and (2) above to accelerate a coordination problem that currently relies on oncologist and patient advocate networks.
4. Translational acceleration. Prediction of which new therapeutic candidates are going to have desirable and clinically feasible pharmacokinetic, pharmacodynamic, and metabolic properties is a huge bottleneck that relies on long animal model-based testing timelines even though these models have low predictive validity. Improving early identification of duds with intractable toxicity liabilities or specifically which models are most representative of clinical pharmacodynamics can be a huge accelerator by substantially trimming preclinical timelines.
5. Virtual cell perturbation modeling. Predicting treatment resistance evolution is partially tractable--the evolutionary paths that resistant cancer cells can take are not unbounded, they are finite, and identifying regulatory constraint configurations common across cancer types can help identify when to use which predictive models of virtual cells, virtual tumor microenvironments, or multi-perturbation responses in identifying which molecular biomarkers or other signs to watch out for in any given patient to get ahead of treatment resistance as early as possible with tailored treatment regimens that adapt faster than the cancer can.

Table 1: Quick Inventory of Technology Readiness for Curing Cancer

Footnotes

1: Research coordination is generally post hoc, leading to duplicative efforts particularly in negative results that never wind up published. I'm generally optimistic about the capacity for ML systems to act as "research escrow" systems with a registered hypothesis and multiple research groups competing to submit results with whatever rewards, financial or social, are tied to TRL advancement being released to first submitter and first replicator, even if it's for negative results. There is a ton of powerful human knowledge about the problems of drug development, diagnostics buildout, and regulatory engagement that narrower research-focused ML may accelerate progress within individual groups, but the potential for increased research gains via better coordinated research efforts may be greater than the impact of any generative ML system in a specific technical aspect that is only one step along the way to making a new therapy, diagnostic, or changing how either are used in sequence or context. Note that this is speculative and not yet a strongly-held opinion, just a thought I'm chewing on.

2: Note that this table excludes current common standard practice for imaging workup and biopsy after initial identification of a palpable mass or other common symptoms.