GWASPoker: Multidisciplinary Genomics & Poker Research

• GWASPoker is a context-dependent term with distinct uses in genomics pre-download triage, No-Limit Hold’em agent design, and continuous recursive game theory models.
• In statistical genetics, it streamlines GWAS summary statistic screening using partial downloads, header detection, and canonical mapping to enhance polygenic risk score pipelines.
• In poker research, it blends self-play optimizations with real-time opponent adaptation for profit maximization, and models continuous guts poker through fixed-point recursion and coalition analysis.

Searching arXiv for papers mentioning “GWASPoker” and closely related context. GWASPoker is an overloaded research label with three distinct technical meanings in the arXiv literature. In statistical genetics, it denotes a phenotype-driven, GWAS-Catalog-specific pre-download triage tool for polygenic risk score calculation that scans candidate GWAS summary statistic files for PRS column availability through partial downloads and header detection, without requiring full-file transfer (Muneeb et al., 6 Mar 2026). In poker research, the same label is used for a profit-maximizing No-Limit Hold’em agent that combines a GTO-like baseline with real-time exploitation (Yi et al., 28 Sep 2025). In generalized recursive game theory, it also denotes continuous guts poker modeled as a single-state generalized recursive game with increasing stakes, enabling fixed-point analysis, coalition optimization, and numerical computation of long-run expected returns (Buck et al., 2022).

1. Terminological scope and disambiguation

The term is used in three non-equivalent senses.

Usage of “GWASPoker”	Technical definition	Source
Statistical genetics tool	A phenotype-driven, GWAS-Catalog-specific pre-download triage tool for PRS column availability	(Muneeb et al., 6 Mar 2026)
No-Limit Hold’em agent	A profit-maximizing poker agent with a GTO foundation plus opponent exploitation	(Yi et al., 28 Sep 2025)
Continuous guts poker model	A continuous, threshold-based model of Guts in a generalized recursive game framework	(Buck et al., 2022)

In the genomics usage, the name is attached to a practical system for finding, screening, and standardizing GWAS summary statistic files before committing to large downloads. The problem setting is explicitly the heterogeneity of GWAS Catalog summary statistic files, whose structures vary across file formats, column labels, and metadata layouts, while file sizes range from 15 MB to 2 GB (Muneeb et al., 6 Mar 2026).

In the poker-agent usage, the name refers to a model that first navigates toward many simulated poker hands against itself and keeps adjusting its decisions until no action can reliably beat it, then adapts by observing opponent behavior and adjusting its strategy to capture extra value accordingly (Yi et al., 28 Sep 2025).

In the generalized recursive game usage, GWASPoker refers to continuous guts poker with i.i.d. uniform private signals, simultaneous “hold” or “drop” decisions, recursive stakes multiplication, and a one-shot reduction based on immediate returns $\alpha$ and stakes multipliers $\beta$ (Buck et al., 2022).

This suggests that “GWASPoker” is best treated as a context-dependent term rather than a unique identifier.

2. GWASPoker as a GWAS summary statistic triage system

In statistical genetics, GWASPoker addresses a specific bottleneck in building PRS pipelines: identifying suitable GWAS summary statistic files across thousands of GWAS Catalog entries before performing full downloads (Muneeb et al., 6 Mar 2026). GWAS summary statistics are the core input to PRS because they encode per-variant effects and association evidence aggregated from large cohorts without exposing individual-level genotypes, and PRS frameworks such as PLINK, PRSice, and ldsc munge_sumstats.py depend on consistent columns and allele conventions (Muneeb et al., 6 Mar 2026).

The system is organized into five modules. Module 1 performs phenotype-driven triage by parsing the GWAS Catalog metadata dump, fuzzy matching user-provided phenotype names in reportedTrait and efoTraits, optionally filtering by discoverySampleAncestry, and optionally using a context-based QA model, ahotrod/electra_large_discriminator_squad2_512, to extract sample sizes or case/control counts from initialSampleDescription (Muneeb et al., 6 Mar 2026). Module 2 scans candidate download pages and first-level subdirectories, creates Information.csv, chooses the largest file when the specific GWAS file is not obvious, performs a partial download with a 10-second timeout, detects delimiters and headers, handles compressed and domain-specific formats, and maps detected headers into a canonical PRS schema (Muneeb et al., 6 Mar 2026). Module 3 performs full download only after pre-download triage. Module 4 extracts and normalizes the downloaded file to a canonical schema using offline synonym dictionaries, with an optional LLM-assisted code-generation step. Module 5 produces a final rules-based transformation template and a presence/absence report for PRS-relevant columns (Muneeb et al., 6 Mar 2026).

The implementation is in Python 3, was developed and tested on Linux HPC with Python $\ge 3.8$, uses pandas for tabular parsing and shell utilities such as gunzip, zcat, and cat for compression handling, and is released under the MIT license (Muneeb et al., 6 Mar 2026). The LLM-based code-generation step is entirely optional, and a rules-based column-mapping template is provided for fully offline use (Muneeb et al., 6 Mar 2026).

The central design choice is pre-download triage. Instead of downloading large summary statistic files and discovering incompatibilities later, the tool reads only the first chunk or chunks needed to determine delimiter structure and header content. This lets researchers select files that satisfy PRS column requirements quickly and reproducibly and only fully download files that will work downstream (Muneeb et al., 6 Mar 2026).

3. Parsing logic, canonical mapping, and PRS integration

GWASPoker’s parsing logic combines time-limited HTTP streaming, delimiter detection, header extraction, compression handling, and rules-based schema normalization (Muneeb et al., 6 Mar 2026). For compressed files with extensions such as .gz, .GZ, .zip, and .tar, the tool sequentially applies gunzip, cat, a “double-gunzip” trick, and zcat to extract just enough content for header detection. For delimiter detection, it tests common delimiters and chooses the one that yields more than one column when read by pandas, then removes extraneous quotes and whitespace to avoid misaligned column detection (Muneeb et al., 6 Mar 2026).

The canonical PRS attributes tracked by the system include chromosome, SNP ID, base pair position, effect allele, non-effect allele, sample size, p-value, imputation info, minor allele frequency, beta, standard error, odds ratio, Z-score, and direction (Muneeb et al., 6 Mar 2026). The rules-based mapper uses synonym dictionaries learned from scanning 54,026 files and, for each canonical attribute, searches the detected headers for any variant of that attribute and returns the matched original header, otherwise NA (Muneeb et al., 6 Mar 2026).

The study reports the following numbers of unique header variants per attribute: Chromosome (12), SNP (34), Base Pair (40), Effect Allele (32), Alternative/Non-effect Allele (34), N (46), P Value (219), Info (22), MAF (109), Beta (21), SE (48), OR (16), Z-score (17), and Direction (12) (Muneeb et al., 6 Mar 2026). Across 54,026 parsed files, the system identified 724 unique header signatures, defined as distinct combinations of column names (Muneeb et al., 6 Mar 2026).

The tool is explicitly positioned as complementary to post-download harmonization systems. Existing tools like gwasrapidd, pandasGWAS, and gwas-download simplify Catalog querying and downloading but do not offer pre-download header triage. MungeSumstats and GWASLab standardize and validate already-downloaded files; GWAS-SSF and gwas-sumstats-tools advance community standards and validation; and GWAS-VCF offers an indexed representation for efficient storage and querying (Muneeb et al., 6 Mar 2026). GWASPoker complements these by screening for PRS column readiness before full transfer and cataloging header signatures across the Catalog (Muneeb et al., 6 Mar 2026).

The downstream PRS interface is conventional. The normalized outputs are intended for tools such as MungeSumstats, PRSice, PLINK, ldsc, or GWAS-SSF workflows, after effect-size conversion, allele alignment, genome-build harmonization, and QC on INFO, MAF, and sample-size fields (Muneeb et al., 6 Mar 2026). The typical PRS formula is

$$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$

where $\beta_j$ is the per-SNP effect size and $x_{ij}$ is the genotype dosage for individual $i$ at SNP $j$, and odds ratios are converted by $\beta = \ln(\mathrm{OR})$ (Muneeb et al., 6 Mar 2026).

4. Empirical coverage, validation, and edge cases in the genomics tool

The reported coverage metrics are extensive. The study analyzed 60,499 GWAS Catalog records; 60,281, or 99.6%, contained accessible download links; and 54,026 of 60,281 files, or 89.6%, were successfully partially downloaded and parsed across 20 file formats, yielding 724 unique header signatures (Muneeb et al., 6 Mar 2026). Across 13 phenotypes, 84 of 85 manually curated GWAS files, or 98.8%, were automatically retrieved and processed (Muneeb et al., 6 Mar 2026). Header validation against fully downloaded files showed exact agreement in 23 of 28 cases, or 82.1% (Muneeb et al., 6 Mar 2026).

The 13 phenotype areas evaluated were asthma; blood_pressure_medication; body_mass_index_bmi; cholesterol_lowering_medication; depression; gastro-oesophageal_reflux_gord_gastric_reflux; hayfever_allergic_rhinitis; high_cholesterol; hypertension; hypothyroidism_myxoedema; irritable_bowel_syndrome; migraine; and osteoarthritis (Muneeb et al., 6 Mar 2026). Module 1 returns all Catalog entries matching the phenotype, after which users manually filter out gene-burden studies, non-exact matches, body-part specific subtypes, and duplicates, typically selecting at least one GWAS per population with maximum discovery samples (Muneeb et al., 6 Mar 2026).

Validation was defined precisely: header validation compared the set and order of column headers detected by the partial-download method against those found after fully downloading the same files, and exact agreement means that the headers matched one-to-one, including order and naming, with no missing or extra columns detected by either method (Muneeb et al., 6 Mar 2026). The five discrepancies were attributed to comment lines preceding the header, multi-row metadata blocks placed above the actual header, and non-standard character encoding (Muneeb et al., 6 Mar 2026).

The main failure modes are also explicit. Servers that do not respond within the 10-second timeout or impose rate limits may thwart partial downloads; non-standard compression wrappers or nested archives may require manual extraction; malformed headers, lengthy pre-header metadata blocks, multiple comment lines, or unusual encodings can defeat header detection; .yaml and .RData files were excluded in the study; and some domain-specific table formats such as .1_summary_table and .modified may require custom parsers (Muneeb et al., 6 Mar 2026). Multi-trait and multi-cohort files can also produce ambiguous mappings, in which case the system returns NA and recommends manual review (Muneeb et al., 6 Mar 2026).

A plausible implication is that the tool is strongest as a reproducible screening layer rather than a replacement for downstream QC or manual adjudication in complex edge cases.

5. GWASPoker as a profit-maximizing No-Limit Hold’em agent

In the 2025 poker paper, GWASPoker denotes a profit-maximizing No-Limit Hold’em model that goes beyond pure Game-Theory-Optimal play by combining a strong GTO-like baseline with real-time exploitation (Yi et al., 28 Sep 2025). The paper defines GTO as a strategy profile that forms a Nash equilibrium; in heads-up zero-sum No-Limit Hold’em, such a strategy is unexploitable and guarantees at least the game value $v^*$ regardless of the opponent’s play (Yi et al., 28 Sep 2025). Exploitability is defined in two-player zero-sum as

$\beta$0

and the stated objective is profit maximization against opponents who deviate from Nash while managing counter-exploitation risk (Yi et al., 28 Sep 2025).

The architecture has two phases. Phase A is self-play optimization to a strong GTO-like baseline using CFR or MCCFR variants on synthetic No-Limit Hold’em decision states and a reference proxy policy. Phase B is real-time adaptation: the system observes opponent behavior, updates an opponent model, and locally adjusts action weights to capture additional expected value while controlling exploitability (Yi et al., 28 Sep 2025).

The formalism is standard extensive-form game notation. Histories are denoted by $\beta$1, terminal histories by $\beta$2, information sets by $\beta$3, legal actions by $\beta$4, and utilities by $\beta$5 (Yi et al., 28 Sep 2025). The paper gives counterfactual values, action-values, regret updates, and regret matching:

$\beta$6

$\beta$7

It discusses external sampling, outcome sampling, and variance-reduction baselines in the MCCFR setting (Yi et al., 28 Sep 2025).

The self-play environment uses synthetic decision states $\beta$8 (Yi et al., 28 Sep 2025). Street sampling uses Dirichlet-stabilized weights with empirical values preflop $\beta$9, flop $\ge 3.8$0, turn $\ge 3.8$1, and river $\ge 3.8$2. Equity sampling uses Beta$\ge 3.8$3 preflop, Beta$\ge 3.8$4 on the flop, Beta$\ge 3.8$5 on the turn, and Beta$\ge 3.8$6 on the river. Texture is sampled uniformly from dry, paired, two_tone, monotone, straighty, and paired+two_tone (Yi et al., 28 Sep 2025). The proxy GTO policy is defined through baseline logits for raise and fold as functions of equity, with calling mass absorbing the remainder and further adjustment by street and texture (Yi et al., 28 Sep 2025).

The algorithms compared are CFR, MCCFR, Deep CFR, NFSP, and Random, with 500 iterations in the reported experiments and five independent seeds with 95% confidence intervals (Yi et al., 28 Sep 2025). In heads-up, the reported convergence-to-proxy metrics at 500 iterations are MCCFR: Top-1 $\ge 3.8$7, KL $\ge 3.8$8, CE $\ge 3.8$9; CFR: Top-1 $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$0, KL $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$1, CE $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$2; NFSP and DeepCFR slower convergence; and Random far from proxy (Yi et al., 28 Sep 2025). In multiway settings with $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$3, MCCFR remained strongest or near-strongest across Top-1, KL, and CE relative to heuristic $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$4, while exact exploitability was not used because the game is general-sum for $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$5 (Yi et al., 28 Sep 2025).

6. Opponent adaptation and risk control in the No-Limit Hold’em usage

The exploitation layer is built on online opponent modeling (Yi et al., 28 Sep 2025). The listed features include observed frequencies per street such as VPIP/PFR analogs, c-bet percentage, check-raise percentage, fold-to-bet percentage, call-down percentage, and river aggression; bet-size distributions over defined raise buckets; showdown tendencies including WSD%, WTSD%, and WWSF%; and contextual features such as position, stack-to-pot ratio, and texture class (Yi et al., 28 Sep 2025).

The estimation method uses online Bayesian updates for binomial events with Beta priors,

$$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$6

with $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$7 updated from counts, alongside maximum-likelihood counts and exponential-decay windows where appropriate (Yi et al., 28 Sep 2025). The adapted policy is a blend of a GTO-like baseline and an exploitative policy:

$$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$8

where $$\mathrm{PRS}_i = \sum_{j=1}^{M} \beta_j \cdot x_{ij},$$9 increases with data confidence, is capped by street, and is risk-sensitive through a penalty on exploitability or a surrogate such as distance or KL divergence to the baseline (Yi et al., 28 Sep 2025).

The paper also specifies operational safeguards: confidence-weighted $\beta_j$0, KL caps $\beta_j$1, minimum sample thresholds $\beta_j$2, street-sensitive adaptation, conservative early-street deviations, and rollback to the baseline when instability is detected (Yi et al., 28 Sep 2025). Deployment guidance emphasizes precomputing baseline policies, performing lightweight online updates, parallelizing MCCFR iterations and best-response tests, and defaulting to the baseline when data are sparse (Yi et al., 28 Sep 2025).

An illustrative river bluff example is given with pot $\beta_j$3100$\beta_j$4B=\$\beta_j$5, fold-to-river-bet estimate $\beta_j$6, and baseline $\beta_j$7. The exploit EV is

$\beta_j$8

while the baseline EV is

$\beta_j$9

for a gain of $x_{ij}$026.25$x_{ij}$1\$x_{ij}$2 when $x_{ij}$3 (Yi et al., 28 Sep 2025).

The paper’s stated limitations are that the synthetic states approximate real No-Limit Hold’em, convergence guarantees hold only in two-player zero-sum games, multiway evaluation relies on heuristics, abstraction can induce bias if expanded to a full tree, and limited training budgets and fixed hyperparameters may affect outcomes under richer trees or deeper stacks (Yi et al., 28 Sep 2025).

7. GWASPoker as continuous guts poker in generalized recursive games

In the 2022 paper, GWASPoker denotes continuous guts poker modeled as a single-state generalized recursive game with increasing stakes (Buck et al., 2022). Players are risk-neutral, indexed by $x_{ij}$4, each receives a private signal $x_{ij}$5 drawn i.i.d. from the uniform distribution, and simultaneously chooses “hold” or “drop” (Buck et al., 2022). If exactly one player holds, that player wins the current pot and the round terminates; if none hold, the round restarts at the same stakes; if $x_{ij}$6 players hold, the highest hand wins the pot, losers match the pot, stakes multiply by $x_{ij}$7, and a new hand is dealt at higher stakes (Buck et al., 2022).

Pure strategies are thresholds $x_{ij}$8: hold iff $x_{ij}$9; otherwise drop (Buck et al., 2022). The one-shot reduction is expressed in terms of expected immediate returns $i$0 and expected stakes multipliers $i$1, with mixed strategies $i$2 represented as probability vectors over a threshold grid and

$i$3

The recursive value iteration is

$i$4

with lower fixed point

$i$5

for player 1’s long-run forceable value (Buck et al., 2022).

For $i$6, the paper gives closed forms:

$i$7

For $i$8,

$i$9

while $j$0 is a piecewise polynomial in the thresholds (Buck et al., 2022).

The unique symmetric Nash equilibrium threshold in standard continuous guts is

$j$1

and under the Weenie rule it shifts to

$j$2

The paper states a striking coalition result: numerics show that the coalition’s near-optimal synchronous strategy is a mixture of two pure profiles, a Bloc$j$3 strategy and a Pseudo-bloc$j$4 strategy, with an asymptotic coalition advantage of approximately $j$5 of player 1’s ante (Buck et al., 2022). The fitted curve is

$j$6

with $j$7 and examples $j$8, $j$9, $\beta = \ln(\mathrm{OR})$0, and $\beta = \ln(\mathrm{OR})$1 (Buck et al., 2022).

The numerical framework uses Fictitious Play to compute $\beta = \ln(\mathrm{OR})$2 over discretized threshold grids, with convergence in finite two-player zero-sum games grounded in Robinson’s theorem (Buck et al., 2022). The paper also reports that Fictitious Play converges empirically to the symmetric Nash equilibrium $\beta = \ln(\mathrm{OR})$3 in the 3-player free-for-all case, with no coalition formation under FP (Buck et al., 2022).

Taken together, the generalized recursive game usage of GWASPoker is not a software pipeline but a mathematically explicit model class linking recursive stakes, threshold strategies, fixed-point value maps, coalition structure, and numerical optimization.

8. Relationship among the three usages

The three usages share only the string “GWASPoker”; they do not define a single research program. The genomics system is a GWAS summary statistic tool for PRS calculation (Muneeb et al., 6 Mar 2026). The 2025 poker-agent usage concerns heads-up and multiway No-Limit Hold’em under CFR, MCCFR, Deep CFR, and NFSP (Yi et al., 28 Sep 2025). The 2022 usage concerns continuous guts poker as a generalized recursive game with threshold strategies, fixed-point recursion, and coalition analysis (Buck et al., 2022).

The strongest internal coherence exists within each domain separately. In genomics, the core themes are partial downloads, header detection, canonical PRS schema mapping, and reproducible pre-download triage. In No-Limit Hold’em, the central themes are GTO-like self-play baselines, exploitability management, and online opponent adaptation. In generalized recursive games, the focus is on $\beta = \ln(\mathrm{OR})$4-$\beta = \ln(\mathrm{OR})$5 reductions, Fictitious Play, fixed-point iteration, and coalition optima.

A plausible implication is that literature searches or citation graphs using only the token “GWASPoker” may conflate genomics and poker materials. For technical work, the term therefore requires immediate domain qualification: GWASPoker in PRS preprocessing, GWASPoker in No-Limit Hold’em, or GWASPoker in continuous guts poker.