Why fund unreasonable science? Afeyan explains

Unreasonable science means funding the hypothesis before the evidence arrives. Noubar Afeyan, co-founder of Moderna and chief executive of Flagship Pioneering, made that case at MIT's HEALS symposium, using individualized mRNA cancer therapy and AI-directed labs as examples of bets that looked implausible when they started and now have clinical data behind them.

What unreasonable science means at MIT HEALS

Unreasonable science is Noubar Afeyan's term for research bets that look implausible before the evidence exists, and he used his MIT HEALS keynote to argue they are where breakthroughs start. Afeyan is chief executive of Flagship Pioneering, the Cambridge venture firm that co-founded Moderna, and he chairs the HEALS External Advisory Council.

His argument rests on a mismatch between how institutions evaluate ideas and when breakthroughs become visible. Groups converge on the common denominator, he said, asking whether a proposal will work, whether the right people are attached, and whether it survives due diligence. Those filters favor the reasonable option, so ideas that cannot yet show data get ignored or underfunded.

Afeyan credited MIT president Sally Kornbluth for driving the institute-wide HEALS initiative and professor Angela Koehler for leading it, along with Iain Cheeseman, Katharina Ribbeck and Caroline Lowenthal. He framed the day's theme as moving from discovery to leadership in translation, policy and ecosystem building.

The claim is testable rather than rhetorical. Afeyan argued that unreasonable ideas should be judged by the burden of producing data, not by the plausibility of the premise. Under that rule, a program that produces no data dies, and one that produces data converts skeptics regardless of how strange the starting hypothesis sounded.

The case for generated biology over discovered biology

Afeyan described a shift from a world that is mostly engineered and discovered to one that is increasingly generated, meaning biological systems and materials designed rather than found. He dates the transition to the convergence of large-scale measurement with machine learning tools that can act on the resulting data.

Biology, in his framing, is information technology. The human genome sequence established that the information exists, and the central dogma describes processing, but the hardware and software of that processing remain largely unmapped. He noted Phillip Sharp, an MIT professor and Nobel laureate in physiology or medicine, as a contributor to understanding how that processing works.

Multimodal measurement matters more than more data of one kind. Afeyan called the shift from seeing to reading to writing to making, and said multi-omics data types inform in different ways than repeated measurements of the same type. That is the technical basis for the generated-biology claim rather than a general enthusiasm about AI.

He also drew a line between point solutions and integrated capability. PCR, discovered by Kary Mullis in 1983, looked like a single transformative tool; decades later, he argued, the field still assembles point capabilities one at a time instead of pursuing integration deliberately.

Individualized mRNA cancer therapy, explained

Individualized mRNA cancer therapy takes a patient's tumor biopsy, identifies neoantigens by sequence analysis, encodes up to 30 to 50 of them in a single mRNA molecule, and delivers it in a lipid nanoparticle to provoke an immune response against that tumor. Afeyan said Moderna and Merck's program has been running since 2017 and now has phase 3 trials.

The manufacturing path is the unusual part. Afeyan described a process that began more than a decade ago and was already testing in humans by 2020, treating patients one at a time with individualized constructs. When COVID-19 arrived, substituting a viral sequence for a tumor sequence and repeating the process at scale became the vaccine production route.

He separated a prophylactic vaccine from a therapeutic one. The program stimulates an immune response in people who already have disease, which he called an immune-modulating therapeutic rather than prevention. He also noted the program's data has been presented publicly rather than only described internally.

Afeyan is careful about the evidence class here: the survival figures come from the companies' phase 2 results with long-term follow-up, not from an independent replication. Readers weighing the claim should treat it as sponsor-reported clinical data reported by the program's own leadership.

How Afeyan frames the mRNA melanoma results

In advanced melanoma, the phase 2 results for the individualized neoantigen therapy showed roughly a 50% improvement in disease-free survival and progression compared with pembrolizumab, Merck's anti-PD-1 antibody sold as Keytruda, according to Afeyan's account of his own company's data. Pembrolizumab itself improves outcomes by about 50% over the comparator in that setting, so the reported effect is an additional gain on top of an active standard of care.

The claim carries an important qualifier. Afeyan said the combined effect comes from adding the individualized therapy to the standard-of-care antibody, which is how the trial was designed, and he described the comparison in relative terms rather than absolute survival months. Relative improvements of that size are meaningful only alongside the absolute numbers and the trial population.

Afeyan also reported that later follow-up, extending toward three and five years, showed the effect persisting, and that the phase 3 readout was expected by the end of the year. Long-term follow-up from a phase 2 cohort is weaker evidence than a randomized phase 3 result, and the program's phase 3 outcome was not available at the time of the talk.

The wider point he drew is about programmability. Because each molecule is itself an information molecule, the same manufacturing approach can be pointed at different tumor types without redesigning the platform. Afeyan said three tumor programs are in phase 3 with Merck, with more in phase 2 and shared-antigen approaches running separately.

Projects Afeyan used as examples of unreasonable bets

Afeyan presented five Flagship-backed projects as evidence that unreasonable premises can survive contact with experiments, and the details differ enough that they are worth separating. Each one started from a hypothesis that could have been wrong, and in several cases the team could not know whether it was wrong until the experiments ran.

Profound: proteins outside the canonical definition

The project started in 2021 from a simple question: are there human proteins the field has never catalogued because they do not fit the canonical definition of a coding gene? The team looked for peptides in mass spectrometry data that map to non-coding regions and for ribosomes attached to matching RNA, then treated those as translated products.

They reported finding tens of thousands of open reading frames of non-canonically expressed proteins, many from non-coding regions. Afeyan connected this to GWAS results that land in non-coding sequence and get dismissed as control regions rather than as expressed protein.

The practical use case is antibody-drug conjugates, which combine a targeting antibody with a cytotoxic payload. Afeyan said the team validated several hundred candidate targets in tumor lines and built about 20 conjugates, with one shown in prostate cancer and described as more specific than PSMA in their hands. That work is preclinical and not in humans.

Abiologics: all-D-amino-acid proteins

This project asked whether a protein built entirely from D-amino acids, the mirror-image form of the amino acids biology normally uses, could work as a drug. Afeyan's prediction was that natural proteases, evolved against L-amino-acid substrates, would not degrade an all-D polypeptide. He described that as speculation at the start, not a finding.

The bottleneck is screening. Because the molecules cannot be produced biologically, libraries must be synthesized chemically, so the team could not screen billions of variants. AlphaFold, the protein structure prediction system, is trained on natural L-amino-acid data and does not predict D-peptide structures, so the team screened brute-force for binders against L-protein targets, then trained models on the first hits.

Afeyan said several generations of that loop, over roughly a year and a half, produced an out-of-the-gate computational hit rate of 10% to 15% for potent binding peptides against multiple targets. Tumor penetration results so far come from animal work.

Quotient and the shift from inherited to acquired variation

Quotient treats somatic mutations, the acquired changes that accumulate in dividing cells, as a source of genetic validation for drug targets. Afeyan said the team estimates roughly 10,000 bases differ between individual cells, which means a person's genome is substantially rewritten across the trillions of cells in the body.

His inference is that where there is variation there is selection, a principle established in cancer. The project deep-sequences millions of cells per patient in disease and control tissue to find acquired mutations enriched in disease, which could mark either a resistant or a protective variant.

Afeyan claimed the approach can match the output of a genome-wide association study using 10 patients and 10 controls instead of tens of thousands of participants. He also said the team found acquired mutation frequencies of 6% to 7% in certain liver failure cases, absent in controls, with p-values he described as extreme. The team has published some of this work, and Afeyan invited the audience to check it rather than accept his summary.

The sample-size comparison is a strong claim and should be read as the project's own framing. GWAS and somatic-mutation analysis answer overlapping but distinct questions: GWAS links inherited variants to disease risk across populations, while somatic analysis looks for acquired mutations within diseased tissue.

AI-directed discovery and the polyintelligence argument

Afeyan described a company founded about three years ago that aims to run the scientific cycle itself: generating hypotheses, designing experiments, executing them, modelling results, and iterating before a human decides what happens next. He said it operates with roughly 250 scientists and engineers and about $550 million assembled to start the effort.

The ambition is an automated scientific factory directed by humans at the level of choosing research areas, while machines run many experimental cycles before human input resumes. Afeyan acknowledged the safety objections directly and argued they are a reason to work on safeguards rather than to stop, adding that competitors would not hold back.

That last point is a strategic assertion, not a measured result. Whether an autonomous loop produces better science per dollar than a conventional lab is an open empirical question, and no independent benchmark of the approach was presented.

He closed with a framework he calls polyintelligence: the triangle formed by human reasoning, machine intelligence, and what he describes as nature's intelligences. Under his definition, any adaptive learning system counts, which brings viral evolution, the human immune system, and synchronized bamboo flowering into the same category. This is a conceptual position about how to allocate research attention, and it is not a falsifiable claim of the kind the rest of the talk dealt in.

The policy argument and its weakest link

Afeyan argued the scientific method itself is under public attack, and that scientists should defend it as vigorously as they defend their own work. He traced part of the erosion to COVID-19 communications, where public health officials stated conclusions with more certainty than the evidence supported, giving critics grounds to question the method itself.

He pointed to a letter he co-authored, titled "Choosing Science," and to work published through the Council on Foreign Relations on biotechnology competitiveness. He described China as a well-funded competitor doubling down on the scientific method while the United States lacks a comparable analysis of where it is losing ground.

He cited the biotech industry's economic weight: an impact of about $3 trillion, 2.3 million employees in 2023, and more than 60% of new drugs originating in small biotech companies. Those figures are the industry's own advocacy numbers and were not sourced on stage to an independent study, so treat them as the case Afeyan is making rather than as audited statistics.

The political science framing is the most contestable part of the talk. Afeyan told the audience that when political science affects their science, they need to engage with it, and that withdrawing from the argument puts funding at risk. That is a claim about strategy, and the talk offered no evidence for the causal link between advocacy and appropriations.

FAQ

• What is unreasonable science according to Noubar Afeyan? Afeyan uses the term for research bets that look implausible before data exists and therefore lose out in committee review, due diligence and funding decisions that favor the reasonable option. His argument at MIT HEALS was that today's breakthroughs already exist in unreasonable form and are underfunded.

• Is the individualized mRNA cancer therapy evidence independent? No. The melanoma disease-free survival improvement of roughly 50% over pembrolizumab comes from the sponsoring companies' phase 2 trial results with long-term follow-up, as described by Afeyan, who runs one of the sponsors. Independent replication and the phase 3 readout were still pending.

• Did Abiologics prove that AlphaFold cannot handle D-peptides? Afeyan said AlphaFold, which is trained on natural L-amino-acid structures, has no basis for predicting D-polypeptide structures. The team worked around this by screening for binders experimentally and training models on the hits, reaching a reported 10% to 15% computational hit rate.

• What is polyintelligence in Afeyan's talk? Polyintelligence is his framing of three interacting intelligences: human reasoning, machine intelligence, and the adaptive systems he attributes to nature, such as viral evolution or the immune system. It is a conceptual argument for how to direct research attention rather than a measured result.

• Why does Afeyan say scientists should engage in policy? He argues that the scientific method is being questioned in public discourse and that scientists who stay out of the debate risk losing the funding their work depends on. He supports the position with a co-authored letter, "Choosing Science," and Council on Foreign Relations work on biotechnology competitiveness.

Source video