NomosLogic Inc

10/15/2026

AI drug discovery is becoming more powerful. The cost of being wrong is rising with it.
Permanent modalities demand more than plausible predictions. They demand mechanisms that can be inspected, kill-conditions defined before the result exists, calibrated predictions that know when to abstain, and provenance that survives scrutiny.
This is the architecture behind falsifiable AI drug discovery at NomosLogic | Heuresis: probabilistic systems generate hypotheses; deterministic gates decide what counts as evidence.
Because in therapeutics, a model can be right—and still be inadmissible.
NomosLogic. Deterministic. Exact. Yours.

10/15/2026

There is a fundamental difference between evidence that is good enough to publish and evidence that is good enough to build a drug program on. The former requires novelty. The latter requires certainty. We consistently conflate the two at our peril.

Academic publishing incentivizes finding a statistically significant p-value on a novel finding. Committing to a ten-year, nine-figure development program requires a different standard: a governable, reproducible, and causally sound body of evidence. The goal is no longer just to be right once, but to be unable to be wrong.

The ~90% clinical failure rate in our industry is, in part, a testament to this category error. We take publication-grade findings and treat them as decision-grade blueprints. How do we build the institutional discipline to recognize the difference and the infrastructure to bridge the gap?

10/09/2026

The new modalities are permanent. A gene edit is not an inhibitor you can wash out. A degraded protein is not a transiently occupied binding pocket. These are architectural changes to the system.

We celebrate the power of these interventions, but we rarely discuss the consequence: permanence dramatically raises the cost of being wrong. When an assay fails with a small molecule, you lose the investment in that screen. When a program based on a permanent modality fails because the target was wrong, you have engineered that failure into the biology itself.

The ~90% clinical failure rate was tolerable when our tools were transient. It is an existential threat when our tools are permanent. This suggests our methods for target selection and evidence governance must evolve faster than our methods for chemical intervention. Have our decision-making frameworks kept pace with the power of our molecular tools?

09/25/2026

Roughly nine out of ten candidates that enter clinical trials will fail. This is a widely cited and humbling statistic. But what if some of those failures were the right drug tested in the wrong population? We have graveyards of late-stage assets and mountains of genomic data from the trials. We should be asking if deterministic re-analysis of that data could reveal a genetic subpopulation where the drug actually worked. Could we find the signal in the noise of a failed Phase 3? It is a hard question. It requires a computational framework that can process that historical data with absolute consistency. But if we could rescue even a handful of those assets for the right patients, the effort would be more than justified.

09/20/2026

Determinism in a computational system means one thing: the same input will produce the exact same output, every time. It is a simple concept with profound implications for drug discovery. This is not an argument against probabilistic models. Their ability to explore vast possibility spaces is essential for hypothesis generation. But the moment a decision is made to advance a candidate, to spend the next million dollars, to move toward the clinic, the logic must be deterministic. The creative, probabilistic engine can propose ten thousand doors. The analytical, deterministic engine is what tells you which nine thousand, nine hundred and ninety nine are locked. We need both, but we must be clear about which tool is right for which job.

09/12/2026

An inherited loss-of-function variant in a DNA mismatch repair gene presents a clear lesson in somatic evolution. The germline variant is not, by itself, oncogenic. Instead, it disables the primary cellular machinery for correcting replication errors. It removes the brakes on the somatic mutation rate.

This creates a state of genomic instability. The cell's normal operations begin to generate a steady stream of new mutations. Most of these are harmless passengers, but eventually, by pure chance, a mutation will arise in a proto-oncogene or a tumor suppressor. The germline variant provides the fertile ground for the subsequent somatic driver events.

This two-step mechanism, an inherited predisposition followed by a stochastic somatic event, explains the penetrance patterns we observe. The cancer risk is probabilistic, not deterministic, because the second hit is a random event. It also explains why the resulting tumors are often hypermutated, carrying thousands of somatic variants not present in the individual's healthy tissue.

09/09/2026

We are shifting the paradigm from fixing failure to optimizing outcomes before failure happens. Our infrastructure establishes the predictive foundation required to engineer therapeutic success at the molecular level.

https://youtu.be/1aorp1_2SBY?si=YogImJV1gtD3ZEGN

We are shifting the paradigm from fixing failure to optimizing outc...

09/09/2026

We have become exceptionally good at identifying statistical associations in human genetics. Genome-wide association studies can now pinpoint hundreds of loci linked to a complex trait. Our ability to interpret these findings mechanistically, however, lags far behind.
What is the biological meaning of a polygenic risk score? When risk is distributed across a thousand common variants, each with a tiny effect size, the classical model of a single broken component in a pathway no longer applies. It suggests a system that is subtly but broadly biased toward a disease state.
How do these small effects combine? Is the architecture simply additive, or are we missing key epistatic interactions? Does the risk manifest as a uniform five percent reduction in the efficiency of a pathway, or does it create a vulnerability that only becomes apparent under specific environmental stress? Answering this requires moving from statistical descriptions to predictive, dynamic models of the underlying networks.
ScientificMethod

09/09/2026

We have become exceptionally good at identifying statistical associations in human genetics. Genome-wide association studies can now pinpoint hundreds of loci linked to a complex trait. Our ability to interpret these findings mechanistically, however, lags far behind.
What is the biological meaning of a polygenic risk score? When risk is distributed across a thousand common variants, each with a tiny effect size, the classical model of a single broken component in a pathway no longer applies. It suggests a system that is subtly but broadly biased toward a disease state.
How do these small effects combine? Is the architecture simply additive, or are we missing key epistatic interactions? Does the risk manifest as a uniform five percent reduction in the efficiency of a pathway, or does it create a vulnerability that only becomes apparent under specific environmental stress? Answering this requires moving from statistical descriptions to predictive, dynamic models of the underlying networks.

09/07/2026

I was trained to believe that a sufficiently high-resolution genome would provide the ultimate explanation for disease. For a long time, my work reflected this belief, focusing almost exclusively on DNA as the source code for biology. I no longer think this is sufficient.

A genome describes a set of possibilities, not a guaranteed outcome. It is a blueprint for a complex machine, but it cannot tell you which parts are currently running, under what load, or how they are interacting with their environment. Without observing the downstream effects in the transcriptome, the proteome, and the resulting phenotype, the genome is a map of roads not taken.

My own work has shifted to demand this consilience. A genomic finding is now the beginning of an inquiry, not the end. A hypothesis is only considered strong when the genomic evidence is corroborated by orthogonal data from other biological modalities.