Barnett Chiu · AI/ML systems for scientific discovery

About

Current work is in computational biology, along two independent lines. One develops generative and self-supervised methods (VAE families, diffusion, flow matching, JEPA). These are modality-agnostic and aimed at computational biology broadly: gene expression, perturbation response, in-silico experimentation. The other is RNA therapeutics, splicing and decay, which began at Envisagenics.

Alongside both runs agentic AI R&D, in agentic-ai-lab and openclaw-py. Agentic workflows suit scientific discovery in particular, because the ground keeps moving: assays are re-run, annotations revised, and results published between one release and the next. A fixed pipeline has no way to go and look. That is why agentic-spliceai puts an agentic layer over its base and meta layers, and takes its name from it.

Earlier work spans foundation-adaptor frameworks for splice prediction, probabilistic modeling, and model interpretability. The through-line is older than the biology: work on parametric action formalisms for decision-making recast actions as operators on state space rather than discrete symbols, and that idea returns now as the action-operator line in current GRL work and in self-supervised world models. In between came sequence modeling of longitudinal clinical records. The projects below are threads with real lineage rather than a list, and the schematic shows how they connect.

I go by Barnett Chiu and publish as Po-Hsiang Chiu; both resolve to ORCID 0000-0001-8816-9799.

Map

How the Work Connects

Arrows read as flow. Grey edges are code lineage, dashed edges are methods borrowed rather than forked, and the copper edge is a biological coupling, not a dependency. Unarrowed edges are associations. private or not yet public

01

RNA Therapeutics active

Two halves of one question, both begun during my time at Envisagenics and continued since as independent lines. Where does splicing go, and what survives?

agentic-spliceai

github.com/pleiadian53/agentic-spliceai

A model-agnostic framework for adaptive splice-site prediction and novel isoform discovery, in three layers: a base layer of per-nucleotide scores from any foundation model (e.g. OpenSpliceAI), a meta layer that adapts them with multimodal context (conservation, chromatin, RNA-seq junctions, RBP binding), and an agentic layer over both.

The agentic layer is the point. Splicing is context-dependent, and the evidence for it is both incomplete and still arriving: ENCODE eCLIP covers on the order of 150 RBPs against the roughly 1,500 encoded in the human genome, so any fixed feature table is a snapshot of what happened to be assayed. An agent can do what a feature table cannot. It reaches across variant databases, tissue and tumor RNA-seq, and the literature not just to check a predicted site, but to go and find the evidence that would support one.

Builds on meta-spliceai, the foundation-adaptor predecessor. Its research agent, Nexus, came from agentic-ai-lab.

Documentation DOI

nmdiff private

Self-supervised quantification of nonsense-mediated decay efficiency, per transcript. NMD is usually described as mRNA quality control; it is also quantity control. Cells route a fraction of a gene's transcripts into PTC-bearing unproductive isoforms precisely so NMD will destroy them, tuning protein output. That matters for antisense oligonucleotide design: a splice-switching drug only helps if the rescued transcript survives decay.

The obstacle is that no labeled dataset for NMD efficiency exists at scale, and the standard self-supervised toolkit is the wrong shape: no corruption you can apply to a transcript has decay kinetics as its inverse. So rather than freezing the labeling function and learning the encoder, nmdiff makes the labeling rule itself a small, interpretable, biologically meaningful parameter vector and searches over it, selecting the hypothesis whose labels are most predictable from a disjoint view of the data: expression generates the labels, transcript structure predicts them, and no feature crosses between the two. That disjointness is what makes it a cross-view consistency test rather than circular reasoning.

02

Generative & Self-Supervised Modeling active

Novel generative and self-supervised methods, developed modality-agnostically on fast proof-of-concept data and aimed at computational biology broadly (gene expression, perturbation response, in-silico experimentation), not only at the splicing work above.

ssl-lab

github.com/pleiadian53/ssl-lab

JEPA pushed past representation learning in two directions. Generative JEPA adds a flow-matching prior over a frozen latent and a decoder back to data space, making the representation sampleable, i.e. generation given an intervention rather than unconditioned. Action operators promote JEPA's blind positional query to a learned operator the model chooses and conditions on, turning a passive predictor into a controllable world model that can roll futures forward and run counterfactuals.

Deliberately modality-agnostic: methods are de-risked here and returned to genai-lab.

Documentation

genai-lab

github.com/pleiadian53/genai-lab

Generative and predictive models for in-silico biological experimentation: VAEs, diffusion, DiT, latent and flow matching, JEPA, foundation-model adaptation. The flagship line is Perturb-seq perturbation response, with uncertainty quantification and biological validity treated as first-class objectives rather than afterthoughts.

Documentation

GRL

github.com/pleiadian53/GRL · generalized reinforcement learning

The root of the action-operator idea ssl-lab now builds on. Actions as operators on state space rather than discrete symbols: an operator- and measurement-inspired formalism for decision-making, framing policy inference in terms of operators, projections, and energy-based compatibility between states and actions. Originated in my doctoral work and extended since.

Its first application was distributed computing for US ATLAS at Brookhaven National Laboratory: a Condor-based pilot factory for the PanDA workload management system, where physics jobs were scheduled as parametric actions. The formalism was never only a thought experiment.

Documentation arXiv:2208.04822

causal-bio-lab

github.com/pleiadian53/causal-bio-lab

Causal inference, causal discovery, and causal representation learning for drug discovery, target identification, and treatment-effect estimation. A sibling to genai-lab at an earlier stage, and one of the paradigms I am investing in ahead of mainstream use.

03

Agentic Systems

Agentic AI R&D, built on current agent tooling so the work stays at the frontier. My interest is in agentic memory and self-improving systems: agents look like one plausible component of AGI, among several breakthroughs still missing.

agentic-ai-lab

github.com/pleiadian53/agentic-ai-lab

Where Nexus was built (a multi-agent research system that plans, researches, writes, and edits) before being refactored into agentic-spliceai as its literature and evidence layer. The design patterns are the deliverable as much as the system.

openclaw-py private

A typed Python SDK for an agent operating system, and the use cases that earned it. An agent OS is excellent from a chat window, but a chat window is where work disappears: every clever automation lives as a message you will never find again, and nothing is reusable, testable, or version controlled. This turns that surface into code.

The applications are the interesting part: a scheduled research digest; per-topic agent personas and routing; pre-flight checks that tell you what a model migration would actually touch; and semantic memory measured rather than trusted, with retrieval scored against a labeled query set and an extractor–verifier pair that fills memory instead of only searching it.

llm-lab

github.com/pleiadian53/llm-lab

Reusable building blocks for LLM pre-training and post-training (supervised fine-tuning, DPO, RLHF, alignment), aimed at LLM-based solutions in computational biology.

04

Clinical & Health Data

Sequence modeling of patient records, from the Columbia and Mount Sinai years, where treating a medical history as a document first made sense to me.

ehr-sequencing

github.com/pleiadian53/ehr-sequencing

Electronic health records as a language: medical codes as tokens, patient histories as sequences. LM-style encoders (BEHRT, transformers, recurrent baselines) applied to disease progression, survival modeling, and temporal phenotyping. Rooted in my postdoctoral work at the Department of Biomedical Informatics, Columbia University.

Documentation

loinc-predictor

github.com/pleiadian53/loinc-predictor

A three-stage retrieval-and-ranking pipeline for LOINC code prediction and reliability assessment. It addresses the interoperability problem that has to be solved before any of the sequence modeling above is trustworthy. An independently reproduced prototype of clinical-analytics work at Prognos Health.

cf-ensemble

github.com/pleiadian53/cf-ensemble

Ensemble learning through latent-factor collaborative filtering: treat classifiers and instances as the two sides of a recommendation problem, and let matrix factorization reweight unreliable predictions. Built for disease risk modeling and phenotyping during my time at the Icahn School of Medicine at Mount Sinai.

Documentation

Other

Other Projects incubating

Early lines, still finding their shape.

pgm-lab

github.com/pleiadian53/pgm-lab

Probabilistic graphical models: a clean, reusable library of PGM building blocks, and research into novel methods for probabilistic modeling. Aimed at computational biology and physics; it is the one line here that reaches past biology.

protein-ml-lab not public yet

A depth study of the protein-ML stack: protein language models (ESM), structure prediction in the AlphaFold lineage, and generative design (RFdiffusion, ProteinMPNN, Chroma), together with the cross-cutting frontier questions in biology foundation models, from mechanistic interpretability to agentic systems built over them.

Protein structure lives on SE(3), so the equivariant architectures in that lineage are built on a group acting on a state space. That is the same object as the action operators in GRL and ssl-lab, with the Lie algebra supplying the infinitesimal view.

Earlier work lives on the GitHub profile.

Publications

Selected Publications

Agentic-SpliceAI: an open-source framework for adaptive splice-site prediction and novel isoform discovery

Zenodo, 2026 · software
Clinical and temporal characterization of COVID-19 subgroups using patient vector embeddings of electronic health records

Journal of the American Medical Informatics Association, 2023
Generalized reinforcement learning: experience particles, action operator, reinforcement field, memory association, and decision concepts

arXiv:2208.04822, 2022
Machine learning–driven identification of early-life air toxic combinations associated with childhood asthma outcomes

Journal of Clinical Investigation, 2021
Predicting youth diabetes risk using NHANES data and machine learning

Scientific Reports, 2021
EHR-based phenotyping: bulk learning and evaluation

Journal of Biomedical Informatics, 2017
Pilot factory – a Condor-based system for scalable Pilot Job generation in the PanDA WMS framework

Journal of Physics: Conference Series, 2010 · US ATLAS

PhD in Computer Science (AI Track), University of Texas at Arlington. Full list on Google Scholar.

Contact

Get in Touch

Happy to talk about splicing, decay, world models, or anything above. barnettchiu@gmail.com.