Research Philosophy Under Construction
Over many years of conversation with my Ph.D. advisor, Donald J. Jacobs, we developed a running thread that merges Bruce Lee’s martial philosophy with how we approach science. We call it, lightly, the Tao of Jeet Kune Xin, the way of the intercepting mind. Jeet Kune Do is the way of the intercepting fist, and Xin (心) shifts the emphasis from strike to cognition, meeting a question early, at the point where it is still forming, before rigid method sets in. The science version is not mysticism. It is a discipline. Absorb what is useful from a model, a dataset, or a theory, and discard what does not survive contact with evidence. Stay formless enough to follow the problem. When binding logic lives at the periphery rather than the active site, intercept there, and when the learner’s hypotheses drift, intercept in projection space before they harden into bias.
In the Tao, understanding the fundamental tools of investigation is part of the training. Hypotheses, postulates, theorems, theses, and the other instruments of formal inquiry are the tools of the sage, not ornaments, but means of navigation. Used well, they chart a path through the unknown without pretending the map is the territory, and used poorly, they become fixed forms that block the intercepting mind. Scientific inquiry is the realm, and these tools are how the investigator moves through it effectively. In a thriving lab they are also the tools of discourse. We use them in conversation, proposing clearly, defining assumptions, arguing consequences, and committing findings to record. The culture that follows is to hypothesize without fear and to take no stance without question. Ideas are offered openly, and they earn their place by surviving scrutiny, not by rank or certainty of tone. Inquiry runs inward toward the unknown and outward toward one another with the same discipline.
“When you go looking for something specific, your chances of finding it are very bad. Because of all the things in the world, you are only looking for one of them. When you go looking for anything at all, your chances of finding it are very good. Because of all the things in the world, you are sure to find some of them.”
Daryl Zero (world’s greatest detective)
That line is a companion to Be like water. Stay formless in aim so discovery can find you.
This page is the technical scaffold. This section is the philosophical one. Both are works in progress. Planned refinements include short axioms in the spirit of Lee’s notes, explicit links to peripheral entropy, Supervised Projective Learning with Orthogonal Completeness (SPLOC), and ALDEN (Automated Learning, Discovery, and Epistemic Navigation), a dedicated note on what “intercepting mind” means in simulation and in automated reasoning, and lab charter language for recruits and collaborators.
| Theme | Draft note |
|---|---|
| Intercept early | Engage recognition at the periphery, of a protein surface, a hypothesis space, or a decision, before the committed path is locked in |
| Be like water | Methods should conform to the problem (disorder, weak homology, high dimension), not the other way around |
| Economy | Minimum sufficient description, entropy where it carries binding information, and projective structure where it carries signal |
| No fixed style | Physics, machine learning, and logic are tools in one kit, not competing identities |
| Tools of the sage | Hypotheses, postulates, theorems, theses, and related instruments of inquiry, learned precisely so they can be wielded, adapted, or set aside as the problem demands |
| Tools of discourse | The same instruments structure lab conversation, proposing, assuming, deriving, documenting, as shared vocabulary for a group navigating the unknown together |
| Hypothesize without fear | Offer ideas before they are polished, since interception works only if questions surface early |
| No stance without question | Hold no position immune to challenge, including method, model, and mentor |
| To refine | Pull quotes, shared lab sayings, and a tighter narrative with Don Jacobs’s input |
| Daryl Zero / water way | Introduce the Daryl Zero quote from the site quotes set - reminder: it captures the same spirit as the water way (formless aim, discovery finds you) |
The final battle of the True Warrior is with the unknown.
Molecular Biophysics Under Construction
The Binding Affinity Problem
What follows is an excerpt from my dissertation on biomolecular binding and thermodynamic recognition. It organizes the problem around three fundamental limitations that strain our notion of what is computable even with modern technology: i) Sampling in High-Dimensional State Spaces; ii) A Need for Context-Specific Descriptors; and iii) Accounting for Global vs. Local Thermodynamics.
Sampling in High-Dimensional State Spaces
The curse of dimensionality is not an abstract concern in biomolecular modeling. It is a concrete barrier that arises the moment one treats binding as an ensemble problem. Even when ignoring explicit solvent, the conformational degrees of freedom (DoF) for flexible proteins and peptides produce a state space whose effective volume grows exponentially with system dimension. With explicit solvent, the number of microstates explodes further, and the equilibrium ensemble becomes a small, structured subset of a vast configuration space. The practical consequence is that no finite simulation campaign can claim exhaustive exploration, and no single protocol can guarantee that the dominant thermodynamic basins have been identified.
This sampling limitation creates two distinct risks. The first is omission, where relevant states may never be visited and low-probability events may matter for functional dynamics or transition pathways. The second is bias, in which the states that are visited may reflect initial conditions, restraints, force-field idiosyncrasies, or protocol choices more strongly than true equilibrium structure. Docking can generate many candidate poses, often under simplified treatments of receptor flexibility and approximate scoring functions. Molecular dynamics simulation (MDS) can incorporate richer physical detail, yet it frequently remains confined to local free-energy basins over accessible simulation timescales, precisely why enhanced-sampling methods have become so important in biomolecular simulation.
An intuitive geometric analogy is to imagine standing in an empty room and tossing a handful of pebbles onto the floor, where they spread uniformly over the surface. In the approximately two-dimensional plane occupied by the pebbles, the coverage may appear extensive. Once the full three-dimensional room is considered, those same points sample only a narrow slice of the available volume. The same logic extends to high-dimensional spaces. Coverage that appears substantial on a low-dimensional manifold may only correspond to a vanishingly small subset of the full state space.
Both docking and simulation operate under partial information, and the central question becomes how to extract reliable conclusions from incomplete ensembles.
A Need for Context-Specific Descriptors
Affinity prediction faces a second challenge beyond sampling. Binding is commonly summarized by the standard thermodynamic relation
where is the Gibbs free energy, is the change in enthalpy, given , and is the absolute temperature-scaled change in entropy. This observable compresses distinct physical contributions into a single scalar. Experimental affinity is reported primarily as the dissociation constant, , or , while decomposition into enthalpic and entropic terms requires additional thermodynamic analysis. Even when and are available, multiple mechanistically distinct combinations can yield similar values. This degeneracy is especially important in biomolecular recognition such as antibody-antigen complexation, where binding often reflects many weak and distributed contributions including hydrophobic contacts, electrostatics, hydrogen bonding, and solvent reorganization. Different mechanisms can converge to similar affinities, masked by enthalpy-entropy compensation.
Equation (1) is deceptively simple. Although provides the correct thermodynamic decomposition, it does not by itself resolve the many distinct physical contributions embedded within the entropic term, including conformational, solvent, and rotational-translational effects. This compression is one reason that binding-affinity calculations often lack generalizability and transferability across unseen systems. As an initial step toward more transferable free-energy determination, this work introduces an entropic component that encodes both the peripheral surface composition of the protein and the structural organization of the first hydration shell, grounded by the known importance of hydration-shell water in biomolecular recognition.
There is something subtle lurking behind Equation (1). The entropy term, , is not a single mechanistic quantity that can be identified with any one isolated source of disorder. In the context of biomolecular binding, it denotes the total thermodynamic entropy of the process, and therefore subsumes multiple distinct contributions arising from molecular motion, internal rearrangement, and environmental reorganization. A more explicit decomposition may be written as
where the terms respectively denote translational, rotational, vibrational, conformational, solvent, and ionic contributions to the total entropy change upon complex formation. Under this view, the persistent lack of transferability and generalizability in computational free energy determination does not arise simply due to numerical insufficiency. It reflects a more basic informational deficit. I lean towards the view that the set of entropic quantities typically retained is not robust enough to represent the full state dependence ingrained in different binding processes of biomolecules.
This limitation constitutes a fundamental barrier in free energy research. Even for the seemingly narrower case of conformational entropy, exact evaluation is formally tied to the full distribution of accessible microstates. In statistical mechanical form,
which requires knowledge of the probabilities over the complete conformational ensemble. For realistic biomolecular systems, this ensemble is neither finitely enumerable nor analytically tractable. Consequently, exact computation of even one component of is already obstructed by the need for information that is inaccessible.
If the full entropic content of binding cannot be directly resolved, then the problem shifts from exhaustive accounting to principled representation. From this view, a global entropy may be interpreted as an entropic proxy designed to capture otherwise unrepresented constraints encoded in the organization of the system. The broader objective is not to enumerate every entropy explicitly. It is to curate a set of entropic proxies that are sufficient for the research question being posed. Progress in binding free-energy prediction depends on whether these proxies can preserve the latent information that conventional decompositions leave behind or never extract at all.
Accounting for Global vs. Local Thermodynamics
A common simplifying assumption in affinity prediction is that binding is determined primarily by local interactions at the interface. This assumption is computationally attractive because interfacial contacts are straightforward to enumerate and convert into approximate energetic terms, and contact-based models are able to recover substantial predictive signal from interface structure alone. The surrounding environment provides thermodynamic context, and that context can reshape the global energy landscape even when local contacts appear dominant. This motivates an extended-interface viewpoint in which nonlocal interactions and peripheral effects can contribute meaningfully to binding strength and specificity.
Several lines of evidence support this broader view. Solvent reorganization can contribute materially to binding thermodynamics through hydration, electrostatics, and water-mediated enthalpy-entropy balance, while binding heat-capacity changes provide an experimental signature of restructuring that extends beyond a purely local contact picture. Allosteric coupling provides a second route by which distal regions can influence binding, since perturbations far from the interface can redistribute conformational populations and alter binding-relevant states. A model restricted to local interface energetics is likely to remain systematically incomplete, especially in systems dominated by weak interactions and solvent-coupled effects.
This work treats binding as a joint constraint with both local and global components. Local interfacial energetics remain necessary, yet they are embedded in a larger thermodynamic setting determined by protein-solvent interactions and peripheral surface composition. The practical question is how to represent that global contribution in a form that is measurable, computable, and predictive. Ensemble-level metrics that quantify organization, variability, or concentration of peripheral states provide a natural route, because they express nonlocal effects as statistical constraints rather than ad hoc corrections to an interface score.
The resulting design philosophy is to optimize interfacial contacts together with the peripheral and solvent-coupled organization that stabilizes binding. Peripheral surface information (PSI) entropy is introduced as one realization of that viewpoint. It combines an enthalpy-aware contact-statistical component motivated by the known predictive value of interfacial contact networks, with a global entropy supported by prior evidence that the non-interacting surface (NIS), its solvation, and its dynamics contribute to favorable binding conditions.
Computational Biology Under Construction
Molecular Docking
If biophysics asks where binding free energy lives, computational biology asks how we operationalize recognition when the search space is vast and the signal is sparse. Molecular docking is the practical face of that question: given structures or models and a hypothesis about complementarity, where do plausible complexes live, and how do we rank them without overfitting the scoring function to the answer we already prefer? My interest is not docking as a black-box oracle but as a pipeline that must coexist with weak homology, ambiguous templates, and twilight-zone regimes where standard similarity metrics stop being trustworthy. Docking workflows here feed the global affinity program by generating and filtering candidates whose thermodynamic plausibility is judged downstream against ensemble-aware criteria.
Feature Engineering
High-dimensional biological data does not yield its structure voluntarily. Feature engineering is the discipline of constructing representations that respect geometry, dynamics, and noise before any learner sees them. My early work on functional protein dynamics recognition used projection pursuit to discover structure in spaces where naive coordinates hide the signal; later work on molecular function recognition extended that logic to supervised settings ( honors thesis; Scientific Reports, 2021; Biomolecules, 2022). Three frameworks carry that projective approach forward, Supervised Projective Learning for Electroencephalography Analysis (SPLEEGA), Supervised Projective Learning with Orthogonal Completeness (SPLOC), and Dynamic Signal Purification. The recurring problem is sampling and computation: the feature space is larger than the budget allows, and the interesting directions are rarely aligned with default axes. The field’s computational challenge is to find representations that carry binding or function information with minimum sufficient complexity, and to know when the representation has captured artifact instead of biology. Future work continues along projection-based and structurally informed feature sets that bridge simulation output, experimental assays, and the affinity models developed in the biophysics pillar.
Cancer Biology
The second thread of my Ph.D. work sits at the intersection of cancer biology and cellular biophysics: how tumor suppressor abundance responds when the cell is pushed out of its comfort zone. Under osmotic stress, I developed what I believe is the first molecular-scale explanation for p53 abundance dynamics, a mechanistic account that links peripheral and conformational response to a quantity the field had previously treated largely at the phenomenological level. That result is not a footnote to the binding-entropy program; it is a distinct discovery that opens a funding line of its own. Peripheral entropy and p53 osmotic response share the same conviction: the answer often lives outside the active site, in the ensemble the standard snapshot discards. Related work on intrinsically disordered segments of p53 and MDM2 extends the same modeling toolkit toward recognition and regulation at the cancer interface. The forward path is to publish and extend this osmotic-stress framework, connect it to therapeutic hypotheses that target peripheral and abundance-level control rather than catalytic pockets alone, and position it as a core aim in upcoming grant proposals.
Medical Counter-Measures
Theory meets urgency in medical countermeasure design. During my internship with Lawrence Livermore’s GUIDE program (Generative Unconstrained Intelligent Drug Engineering), I worked at the intersection of machine learning, simulation, and biodefense-oriented therapeutic design, applying binding affinity and free-energy thinking to problems where the cost of error is measured in human lives. This is where computational pipelines are judged by deployment constraints: time, safety, dual-use risk, and the political economy of preparedness. That pressure is not abstract for me: the same automated design pipelines that accelerate countermeasures can lower the barrier to misuse, which is why governance and screening architecture belong in the same conversation as the science, developed further in my biosecurity brief. My orientation toward biosecurity and biodefense is one forcing function that demands models generalize outside the comfortable benchmark. Future directions include tighter coupling between countermeasure design workflows and the ensemble-based affinity framework, with explicit accounting for what simulations can and cannot certify.
Automated Reasoning Under Construction
Quantum Logic
The third pillar concerns what it means for a machine, or a human-machine pair, to reason reliably under constraint. Supervised Projective Learning with Orthogonal Completeness (SPLOC) is my modern adaptation of quantum logic in the sense of Birkhoff and von Neumann: not physics cosplay, but a formalism for how observations, projections, and complementarity interact when the state space is high-dimensional and incomplete. Supervised Projective Learning for Electroencephalography Analysis (SPLEEGA), my M.S. thesis work, was an early testbed ( M.S. thesis; IEEE SMC 2021). Brain-computer interfaces punish sloppy projection with artifact and nonstationarity. The fundamental question is whether orthogonality and completeness conditions can be enforced in learned representations so that what the system claims to have separated has actually been separated. The limits here are informational: finite data, finite bandwidth, finite energy per inference.
Theory of Mind
Automated learners do not only fit surfaces; they form and select hypotheses, often prematurely. Biased hypothesis formation in projection pursuit is a technical phenomenon with a philosophical shadow: when does a learner commit to an explanation because the representation made that explanation easy? Theory of mind, in this program, is not folk psychology bolted onto a chatbot. It is the study of how an investigator and an automated system coordinate inquiry, each modeling, however implicitly, what the other can know, infer, or overlook. Reasoning with the machine means designing interaction so that hypotheses surface early enough to be intercepted, before they harden into fixed forms. The field’s challenge is trust without credulity: systems that assist without laundering unjustified certainty. That problem does not stay in the lab. As thinking machines move from classifying EEG trials to proposing molecules, policies, and protocols, the question of what a system can know, and what a human operator falsely believes it knows, becomes a biosecurity question in its own right, tied to screening, attribution, and oversight. I develop that thread publicly in the biosecurity brief and in ongoing work on ALDEN’s interaction model. Future work links this subsection to lab discourse norms and to when human oversight fails in the same ways machine bias does.
A Geometry of Logic
It is time for a modern attempt to extend the logical frameworks of Aristotle and the Boolean operations built on them. Years of work in automated reasoning have hit three primary walls. ALDEN (Automated Learning, Discovery, and Epistemic Navigation) trades three paradoxes for one engineering problem, optimization, and that trade is the heart of this line of investigation.
A simple fact motivates ALDEN. When artificial intelligence reaches the singularity, my hope is that entities on the other side can be reasoned with. That is also my hope for biological intelligence within our cosmos, a partnership that would make the modern AI alignment problem obsolete.
The 8-7 companion star system. Eight-point star: perfect symmetry, omnidirectional steadiness, the Old Friend that steadies our reach. Seven-point star: asymmetric humanity, vertically striving between earth and sky, uneven and bias-prone by nature.
ALDEN, now in development, is the model built on that geometry, explainable and inspectable AI that can be reasoned with, not a bag of prompts. The name is Old English for old friend, emphasizing companionship rather than crude relations such as ownership. Boolean logic on a commuting lattice fragment is a starting grammar: enough structure to compose statements, check consistency, and localize error, without pretending the full logical landscape is capturable in one formalism. Alongside that formal work sits a triad I treat as prerequisite for any system that reasons and acts: logic (structure), energy (cost of maintaining and updating state), and time (the horizon over which commitments unfold). Automated reasoning research that ignores thermodynamic and temporal cost is incomplete. The limits of thinking machines are the limits of representation, measurement, and energy budget in tandem. Future directions for ALDEN include executable semantics, inspectable software, and publications that connect this geometry to SPLOC’s projective worldview and to the intercepting-mind philosophy developed with Don Jacobs, one kit of tools, not three unrelated careers.
These three pillars are one program viewed from different elevations, not three careers, not three silos. Thermodynamics asks what binding costs and what entropy carries. Computation asks how we search and represent in spaces too large to exhaust. Automated reasoning asks what must be true of a system, biological or artificial, before its inferences deserve to guide action. The connections are not decorative. Peripheral entropy is also a theory of where information lives when the obvious coordinate fails; projection pursuit is the same move in hypothesis space; ALDEN is the attempt to give that move a logic. And automated reasoning is not a separate ivory tower from biodefense: the more capable the thinking machine, the more urgent the question of who may use it, under what screening, and with what accountability when the model is wrong. Molecular biophysics, computational biology, and automated reasoning are one foundation. Medical countermeasures and biosecurity governance are where that foundation meets consequence. Each pillar is shaped by the field’s hard limits and by the obligation to say plainly where the map ends and the territory continues.