@sjcrh/proteinpaint-rust 2.180.0 → 2.181.0

package/Cargo.toml

@@ -22,6 +22,7 @@ plotters = "0.3.4"
 tiny-skia = "0.11"
 colorgrad = "0.6.2"
 statrs = "^0.16.0"
+libc = "0.2"
 fishers_exact="^1.0.1"
 bio = "1.5.0"
 bigtools = "^0.1.11"
@@ -133,4 +134,8 @@ path="src/query_classification.rs"
 
 [[bin]]
 name="summary_agent"
-path="src/summary_agent.rs"
+path="src/summary_agent.rs"
+
+[[bin]]
+name="dmrcate"
+path="src/dmrcate.rs"

package/package.json

@@ -1,5 +1,5 @@
 {
-	"version": "2.180.0",
+	"version": "2.181.0",
 	"name": "@sjcrh/proteinpaint-rust",
 	"type": "module",
 	"description": "Rust-based utilities for proteinpaint",

package/src/dmrcate.rs

@@ -0,0 +1,996 @@
+// dmrcate.rs — Genome-wide empirical Bayes moderated t-test + DMRCate kernel smoothing
+// Smyth 2004 (limma), Peters et al. 2015 (DMRCate), Phipson et al. 2016 (robust eBayes)
+//
+// Reads probe-level beta values from HDF5, runs chromosome-chunked OLS → genome-wide
+// eBayes → regional kernel smoothing → DMR segmentation with proximity fallback.
+// Usage: echo '{"probe_h5_file":"beta.h5","chr":"chr14","start":100000,"stop":105000,
+//              "case":"s1,s2","control":"s3,s4"}' | target/release/dmrcate
+
+use base64::{Engine as _, engine::general_purpose::STANDARD as BASE64};
+use hdf5::File;
+use hdf5::types::VarLenUnicode;
+use serde_json::{Value, json};
+use statrs::distribution::{ContinuousCDF, StudentsT};
+use statrs::function::gamma::{digamma, gamma_ur, ln_gamma};
+use std::collections::HashMap;
+use std::io;
+use std::time::Instant;
+use tiny_skia::{FillRule, Paint, PathBuilder, Pixmap, Stroke, StrokeDash, Transform};
+
+fn get_rss_mb() -> f64 {
+    unsafe {
+        let mut usage: libc::rusage = std::mem::zeroed();
+        libc::getrusage(libc::RUSAGE_SELF, &mut usage);
+        #[cfg(target_os = "macos")]
+        {
+            usage.ru_maxrss as f64 / 1_048_576.0
+        } // bytes → MB
+        #[cfg(not(target_os = "macos"))]
+        {
+            usage.ru_maxrss as f64 / 1024.0
+        } // KB → MB
+    }
+}
+
+fn trigamma(mut x: f64) -> f64 {
+    if x <= 0.0 {
+        return f64::NAN;
+    }
+    let mut r = 0.0;
+    while x < 6.0 {
+        r += 1.0 / (x * x);
+        x += 1.0;
+    }
+    let x2 = x * x;
+    r + 1.0 / x + 1.0 / (2.0 * x2) + 1.0 / (6.0 * x2 * x) - 1.0 / (30.0 * x2 * x2 * x) + 1.0 / (42.0 * x2 * x2 * x2 * x)
+}
+
+fn trigamma_deriv(mut x: f64) -> f64 {
+    let mut r = 0.0;
+    while x < 6.0 {
+        r -= 2.0 / (x * x * x);
+        x += 1.0;
+    }
+    let x2 = x * x;
+    r - 1.0 / x2 - 1.0 / (x2 * x) - 1.0 / (2.0 * x2 * x2) + 1.0 / (6.0 * x2 * x2 * x2)
+}
+
+fn trigamma_inverse(x: f64) -> f64 {
+    if x.is_nan() || x <= 0.0 {
+        return f64::NAN;
+    }
+    let mut y = if x > 1e-6 { 1.0 / x.sqrt() } else { 1.0 / x };
+    for _ in 0..8 {
+        let delta = (trigamma(y) - x) / trigamma_deriv(y);
+        y -= delta;
+        if y <= 0.0 {
+            y = 0.5 * (y + delta);
+        }
+        if delta.abs() < 1e-12 * y.abs() {
+            break;
+        }
+    }
+    y
+}
+
+fn bh_adjust(pvalues: &[f64]) -> Vec<f64> {
+    let n = pvalues.len();
+    if n == 0 {
+        return vec![];
+    }
+    let mut idx: Vec<usize> = (0..n).collect();
+    idx.sort_by(|&a, &b| pvalues[b].partial_cmp(&pvalues[a]).unwrap_or(std::cmp::Ordering::Equal));
+    let mut adj = vec![0.0; n];
+    let mut cummin = f64::INFINITY;
+    for (rank_from_end, &i) in idx.iter().enumerate() {
+        cummin = cummin.min(pvalues[i] * n as f64 / (n - rank_from_end) as f64);
+        adj[i] = cummin.min(1.0);
+    }
+    adj
+}
+
+struct ProbeStats {
+    chr: String,
+    start: i64,
+    probe_id: String,
+    log_fc: f64,
+    residual_var: f64,
+    df_residual: f64,
+    stdev_unscaled: f64,
+}
+
+fn read_h5_metadata(file: &File) -> Result<(Vec<String>, Vec<usize>, Vec<String>, Vec<i64>, Vec<String>), String> {
+    let samples: Vec<String> = file
+        .dataset("meta/samples/names")
+        .map_err(|e| e.to_string())?
+        .read_1d::<VarLenUnicode>()
+        .map_err(|e| e.to_string())?
+        .iter()
+        .map(|s| s.to_string())
+        .collect();
+    let starts: Vec<i64> = file
+        .dataset("meta/start")
+        .map_err(|e| e.to_string())?
+        .read_1d::<i64>()
+        .map_err(|e| e.to_string())?
+        .to_vec();
+    let probes: Vec<String> = file
+        .dataset("meta/probe/probeID")
+        .map_err(|e| e.to_string())?
+        .read_1d::<VarLenUnicode>()
+        .map_err(|e| e.to_string())?
+        .iter()
+        .map(|s| s.to_string())
+        .collect();
+    let root = file.group("/").map_err(|e| e.to_string())?;
+    let cl_json: String = root
+        .attr("chrom_lengths")
+        .map_err(|e| e.to_string())?
+        .read_scalar::<VarLenUnicode>()
+        .map_err(|e| e.to_string())?
+        .to_string();
+    // json crate preserves key order; serde_json::Map sorts alphabetically (wrong for chromosomes)
+    let cl_parsed = json::parse(&cl_json).map_err(|e| format!("Failed to parse chrom_lengths: {}", e))?;
+    let mut names = Vec::new();
+    let mut lens = Vec::new();
+    for (k, v) in cl_parsed.entries() {
+        names.push(k.to_string());
+        lens.push(v.as_u64().unwrap_or(0) as usize);
+    }
+    Ok((names, lens, samples, starts, probes))
+}
+
+fn process_chromosome(
+    file: &File,
+    row_start: usize,
+    row_end: usize,
+    case_idx: &[usize],
+    ctrl_idx: &[usize],
+    chr: &str,
+    starts: &[i64],
+    probe_ids: &[String],
+    min_spg: usize,
+) -> Result<Vec<ProbeStats>, String> {
+    let n_probes = row_end - row_start;
+    if n_probes == 0 {
+        return Ok(vec![]);
+    }
+    let ds = file.dataset("beta/values").map_err(|e| format!("beta/values: {}", e))?;
+    let mut results = Vec::with_capacity(n_probes);
+    const CHUNK: usize = 1000;
+    for chunk_i in 0..((n_probes + CHUNK - 1) / CHUNK) {
+        let cs = chunk_i * CHUNK;
+        let ce = std::cmp::min(cs + CHUNK, n_probes);
+        let sel = hdf5::Selection::from((row_start + cs..row_start + ce, ..));
+        let data = ds
+            .read_slice_2d::<f32, _>(sel)
+            .map_err(|e| format!("HDF5 read: {}", e))?;
+        for lp in 0..(ce - cs) {
+            let row = data.row(lp);
+            let (mut cv, mut kv) = (Vec::with_capacity(case_idx.len()), Vec::with_capacity(ctrl_idx.len()));
+            for &si in case_idx {
+                if si < row.len() {
+                    let v = row[si] as f64;
+                    if v.is_finite() {
+                        cv.push(v);
+                    }
+                }
+            }
+            for &si in ctrl_idx {
+                if si < row.len() {
+                    let v = row[si] as f64;
+                    if v.is_finite() {
+                        kv.push(v);
+                    }
+                }
+            }
+            if cv.len() < min_spg || kv.len() < min_spg {
+                continue;
+            }
+            let to_m = |b: f64| {
+                let c = b.clamp(0.001, 0.999);
+                (c / (1.0 - c)).log2()
+            };
+            let cm: Vec<f64> = cv.iter().map(|&b| to_m(b)).collect();
+            let km: Vec<f64> = kv.iter().map(|&b| to_m(b)).collect();
+            let all: Vec<f64> = cm.iter().chain(km.iter()).copied().collect();
+            let mean_all = all.iter().sum::<f64>() / all.len() as f64;
+            let var = all.iter().map(|&x| (x - mean_all).powi(2)).sum::<f64>() / (all.len() as f64 - 1.0);
+            if var <= 0.0 || !var.is_finite() {
+                continue;
+            }
+            let (n1, n2) = (cm.len() as f64, km.len() as f64);
+            let (mc, mk) = (cm.iter().sum::<f64>() / n1, km.iter().sum::<f64>() / n2);
+            let ss: f64 =
+                cm.iter().map(|&x| (x - mc).powi(2)).sum::<f64>() + km.iter().map(|&x| (x - mk).powi(2)).sum::<f64>();
+            let df = n1 + n2 - 2.0;
+            let rv = ss / df;
+            if !rv.is_finite() || rv <= 0.0 {
+                continue;
+            }
+            let su = (1.0 / n1 + 1.0 / n2).sqrt();
+            let idx = row_start + cs + lp;
+            results.push(ProbeStats {
+                chr: chr.to_string(),
+                start: starts[idx],
+                probe_id: probe_ids[idx].clone(),
+                log_fc: mc - mk,
+                residual_var: rv,
+                df_residual: df,
+                stdev_unscaled: su,
+            });
+        }
+    }
+    Ok(results)
+}
+
+fn fit_f_dist(vars: &[f64], dfs: &[f64]) -> (f64, f64) {
+    if vars.len() < 3 {
+        return (1.0, 0.0);
+    }
+    // Match R's fitFDist pre-processing:
+    // 1. Filter to ok probes (finite df > 1e-15, finite var > -1e-15)
+    // 2. Clamp var to max(var, 0), then floor at 1e-5 * median(var)
+    let ok: Vec<usize> = (0..vars.len())
+        .filter(|&i| dfs[i].is_finite() && dfs[i] > 1e-15 && vars[i].is_finite() && vars[i] > -1e-15)
+        .collect();
+    if ok.len() < 3 {
+        return (1.0, 0.0);
+    }
+    let mut xv: Vec<f64> = ok.iter().map(|&i| vars[i].max(0.0)).collect();
+    let xdf: Vec<f64> = ok.iter().map(|&i| dfs[i]).collect();
+    // Median of variances
+    let mut sorted_v = xv.clone();
+    sorted_v.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    let median_v = if sorted_v.len() % 2 == 0 {
+        (sorted_v[sorted_v.len() / 2 - 1] + sorted_v[sorted_v.len() / 2]) / 2.0
+    } else {
+        sorted_v[sorted_v.len() / 2]
+    };
+    if median_v == 0.0 {
+        return (1.0, f64::INFINITY);
+    }
+    // Floor small variances at 1e-5 * median (matches R's fitFDist)
+    let floor = 1e-5 * median_v;
+    for v in &mut xv {
+        if *v < floor {
+            *v = floor;
+        }
+    }
+    let n = xv.len() as f64;
+    // e = log(var) + logmdigamma(df/2) where logmdigamma(a) = log(a) - digamma(a)
+    let e: Vec<f64> = xv
+        .iter()
+        .zip(xdf.iter())
+        .map(|(&v, &d)| v.ln() + (d / 2.0).ln() - digamma(d / 2.0))
+        .collect();
+    let me = e.iter().sum::<f64>() / n;
+    let ve = e.iter().map(|&ei| (ei - me).powi(2)).sum::<f64>() / (n - 1.0);
+    let mean_tri: f64 = xdf.iter().map(|&d| trigamma(d / 2.0)).sum::<f64>() / n;
+    let target = ve - mean_tri;
+    let df0 = if target > 0.0 {
+        2.0 * trigamma_inverse(target)
+    } else {
+        f64::INFINITY
+    };
+    let s20 = if df0.is_finite() {
+        (me - (df0 / 2.0).ln() + digamma(df0 / 2.0)).exp()
+    } else {
+        xv.iter().sum::<f64>() / n
+    };
+    (s20, df0)
+}
+
+/// Log of the upper regularized incomplete gamma function Q(a, x).
+/// Uses the continued fraction representation (Numerical Recipes / TOMS 708),
+/// evaluated via modified Lentz's method. Returns the result in log space
+/// so it never underflows, even for Q values as small as exp(-1e6).
+/// This matches R's pgamma(x, a, lower.tail=FALSE, log.p=TRUE).
+fn log_gamma_upper_cf(a: f64, x: f64) -> f64 {
+    // Q(a, x) = exp(-x + a*ln(x) - lgamma(a)) * h
+    // where h is the continued fraction. h is O(1/x) and well-behaved.
+    let eps = 3e-14;
+    let tiny = 1e-300;
+
+    let mut b = x + 1.0 - a;
+    let mut c = 1.0 / tiny;
+    let mut d = 1.0 / b;
+    let mut h = d;
+
+    for i in 1..=300 {
+        let an = -(i as f64) * (i as f64 - a);
+        b += 2.0;
+        d = an * d + b;
+        if d.abs() < tiny {
+            d = tiny;
+        }
+        c = b + an / c;
+        if c.abs() < tiny {
+            c = tiny;
+        }
+        d = 1.0 / d;
+        let del = d * c;
+        h *= del;
+        if (del - 1.0).abs() < eps {
+            break;
+        }
+    }
+
+    -x + a * x.ln() - ln_gamma(a) + h.ln()
+}
+
+/// Log chi-squared survival function: returns log P(X > x) for X ~ chi^2(df).
+/// Uses statrs gamma_ur for moderate tails, continued fraction in log space
+/// for extreme tails. Matches R's pchisq(x, df, lower.tail=FALSE, log.p=TRUE).
+fn log_chisq_sf(x: f64, df: f64) -> f64 {
+    if x <= 0.0 || !x.is_finite() {
+        return 0.0; // log(1) = 0
+    }
+    let a = df / 2.0;
+    let z = x / 2.0;
+    // For moderate tails, use statrs (accurate and fast)
+    let sf = gamma_ur(a, z);
+    if sf > 1e-300 {
+        return sf.ln();
+    }
+    // For extreme tails, use continued fraction in log space
+    log_gamma_upper_cf(a, z)
+}
+
+/// Kernel smoothing returning LOG p-values (not p-values) to avoid underflow.
+fn kernel_smooth_log(pos: &[i64], t: &[f64], lambda: f64, c: f64) -> Vec<f64> {
+    let sigma = lambda / c;
+    let max_d = (5.0 * sigma) as i64;
+    let two_s2 = 2.0 * sigma * sigma;
+    let (n, mut l, mut r) = (pos.len(), 0usize, 0usize);
+    let mut out = Vec::with_capacity(n);
+    for i in 0..n {
+        while r < n && (pos[r] - pos[i]).abs() <= max_d {
+            r += 1;
+        }
+        while l < n && pos[i] - pos[l] > max_d {
+            l += 1;
+        }
+        let (mut sky, mut sk, mut skk) = (0.0, 0.0, 0.0);
+        for j in l..r {
+            let dx = (pos[i] - pos[j]) as f64;
+            let w = (-dx * dx / two_s2).exp();
+            sky += w * t[j] * t[j];
+            sk += w;
+            skk += w * w;
+        }
+        let log_p = if sk > 0.0 && skk > 0.0 {
+            let (exp, var) = (sk, 2.0 * skk);
+            let (b, a) = (2.0 * exp * exp / var, var / (2.0 * exp));
+            if b > 0.0 && a > 0.0 {
+                log_chisq_sf(sky / a, b)
+            } else {
+                0.0
+            }
+        } else {
+            0.0
+        };
+        out.push(log_p);
+    }
+    out
+}
+
+/// BH adjustment on log-scale p-values. Returns log-scale adjusted p-values.
+/// adj_log_p[i] = cummin(log_p[i] + ln(n) - ln(rank)) capped at 0 (= log(1))
+fn bh_adjust_log(log_pvalues: &[f64]) -> Vec<f64> {
+    let n = log_pvalues.len();
+    if n == 0 {
+        return vec![];
+    }
+    let ln_n = (n as f64).ln();
+    let mut idx: Vec<usize> = (0..n).collect();
+    // Sort descending (largest log p first = least significant)
+    idx.sort_by(|&a, &b| {
+        log_pvalues[b]
+            .partial_cmp(&log_pvalues[a])
+            .unwrap_or(std::cmp::Ordering::Equal)
+    });
+    let mut adj = vec![0.0f64; n];
+    let mut cummin = 0.0f64; // log(1) = 0
+    for (rank_from_end, &i) in idx.iter().enumerate() {
+        let rank = n - rank_from_end; // 1-based rank from smallest
+        let v = log_pvalues[i] + ln_n - (rank as f64).ln();
+        cummin = cummin.min(v);
+        adj[i] = cummin.min(0.0); // cap at log(1) = 0
+    }
+    adj
+}
+
+fn build_dmrs(
+    chr: &str,
+    pos: &[i64],
+    fdr: &[f64],
+    lfc: &[f64],
+    mg1: &[f64],
+    mg2: &[f64],
+    cutoff: f64,
+    lambda: f64,
+    min_cpgs: usize,
+    min_db: Option<f64>,
+    check_direction: bool,
+) -> Vec<Value> {
+    let n = pos.len();
+    let sig: Vec<usize> = (0..n)
+        .filter(|&i| {
+            if fdr[i] > cutoff {
+                return false;
+            }
+            if let Some(db) = min_db {
+                (mg2[i] - mg1[i]).abs() >= db
+            } else {
+                true
+            }
+        })
+        .collect();
+    if sig.len() < min_cpgs {
+        return vec![];
+    }
+    let mut groups: Vec<Vec<usize>> = Vec::new();
+    let mut grp = vec![sig[0]];
+    for k in 1..sig.len() {
+        let (p, c) = (*grp.last().unwrap(), sig[k]);
+        let same_dir = !check_direction || (lfc[c] >= 0.0) == (lfc[p] >= 0.0);
+        if same_dir && (pos[c] - pos[p]) <= lambda as i64 {
+            grp.push(c);
+        } else {
+            groups.push(grp);
+            grp = vec![c];
+        }
+    }
+    groups.push(grp);
+    groups
+        .iter()
+        .filter(|g| g.len() >= min_cpgs)
+        .map(|g| {
+            let deltas: Vec<f64> = g.iter().map(|&j| mg2[j] - mg1[j]).collect();
+            let fdrs: Vec<f64> = g.iter().map(|&j| fdr[j]).collect();
+            let md = deltas.iter().sum::<f64>() / deltas.len() as f64;
+            let mxd = if md >= 0.0 {
+                deltas.iter().cloned().fold(f64::NEG_INFINITY, f64::max)
+            } else {
+                deltas.iter().cloned().fold(f64::INFINITY, f64::min)
+            };
+            json!({ "chr": chr, "start": pos[*g.first().unwrap()], "stop": pos[*g.last().unwrap()],
+            "no_cpgs": g.len(), "min_smoothed_fdr": fdrs.iter().cloned().fold(f64::INFINITY, f64::min),
+            "HMFDR": fdrs.len() as f64 / fdrs.iter().map(|&f| 1.0/f.max(1e-300)).sum::<f64>(),
+            "maxdiff": mxd, "meandiff": md, "direction": if md >= 0.0 {"hyper"} else {"hypo"},
+            "overlapping_genes": null })
+        })
+        .collect()
+}
+
+macro_rules! bail { ($($t:tt)*) => { { println!("{}", json!({"error": format!($($t)*)})); return; } } }
+
+/// LOESS (locally weighted scatterplot smoothing) with tricube weights and local linear fit.
+/// Returns (fitted, ci_lower, ci_upper) evaluated at `eval_at` positions, clamped to [0,1].
+fn loess_fit(pos: &[i64], vals: &[f64], eval_at: &[f64], span: f64) -> Option<(Vec<f64>, Vec<f64>, Vec<f64>)> {
+    // Collect valid (x, y) pairs (skip NaN)
+    let mut xs = Vec::new();
+    let mut ys = Vec::new();
+    for i in 0..pos.len() {
+        if vals[i].is_finite() {
+            xs.push(pos[i] as f64);
+            ys.push(vals[i]);
+        }
+    }
+    let n = xs.len();
+    if n < 4 {
+        return None;
+    }
+
+    let k = ((span * n as f64).ceil() as usize).max(3).min(n);
+
+    let mut fitted = Vec::with_capacity(eval_at.len());
+    let mut ci_lo = Vec::with_capacity(eval_at.len());
+    let mut ci_hi = Vec::with_capacity(eval_at.len());
+
+    for &x0 in eval_at {
+        // Find k nearest neighbors by distance
+        let mut dists: Vec<(usize, f64)> = xs.iter().enumerate().map(|(i, &xi)| (i, (xi - x0).abs())).collect();
+        dists.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap());
+        let h = dists[k - 1].1.max(1.0); // bandwidth = distance to k-th nearest
+
+        // Tricube weights
+        let mut w = vec![0.0; n];
+        for &(i, d) in dists.iter().take(k) {
+            let u = d / h;
+            if u < 1.0 {
+                let t = 1.0 - u * u * u;
+                w[i] = t * t * t;
+            }
+        }
+
+        // Weighted linear regression: y = a + b*(x - x0)
+        let mut sw = 0.0;
+        let mut swx = 0.0;
+        let mut swy = 0.0;
+        let mut swxx = 0.0;
+        let mut swxy = 0.0;
+        for i in 0..n {
+            if w[i] == 0.0 {
+                continue;
+            }
+            let dx = xs[i] - x0;
+            sw += w[i];
+            swx += w[i] * dx;
+            swy += w[i] * ys[i];
+            swxx += w[i] * dx * dx;
+            swxy += w[i] * dx * ys[i];
+        }
+        if sw == 0.0 {
+            fitted.push(f64::NAN);
+            ci_lo.push(f64::NAN);
+            ci_hi.push(f64::NAN);
+            continue;
+        }
+        let det = sw * swxx - swx * swx;
+        let (a, _b) = if det.abs() < 1e-20 {
+            (swy / sw, 0.0)
+        } else {
+            ((swxx * swy - swx * swxy) / det, (sw * swxy - swx * swy) / det)
+        };
+        let y_hat = a; // at x = x0, dx = 0, so y = a
+
+        // Weighted residual variance for CI
+        let mut sse = 0.0;
+        let mut sw2 = 0.0;
+        for i in 0..n {
+            if w[i] == 0.0 {
+                continue;
+            }
+            let dx = xs[i] - x0;
+            let pred = a + _b * dx;
+            let e = ys[i] - pred;
+            sse += w[i] * e * e;
+            sw2 += w[i] * w[i];
+        }
+        // Effective df ≈ sum(w)^2 / sum(w^2) - 2
+        let eff_n = (sw * sw / sw2).max(3.0);
+        let sigma2 = sse / (eff_n - 2.0).max(1.0);
+        let se = (sigma2 / sw).sqrt();
+
+        // 95% CI with normal approximation (effective n is typically large for LOESS)
+        let margin = 1.96 * se;
+        fitted.push((y_hat * 10000.0).round() / 10000.0);
+        ci_lo.push(((y_hat - margin).max(0.0).min(1.0) * 10000.0).round() / 10000.0);
+        ci_hi.push(((y_hat + margin).max(0.0).min(1.0) * 10000.0).round() / 10000.0);
+    }
+
+    // Clamp fitted values
+    for v in fitted.iter_mut() {
+        *v = v.max(0.0).min(1.0);
+    }
+
+    Some((fitted, ci_lo, ci_hi))
+}
+
+fn hex_to_rgba(hex: &str, alpha: u8) -> (u8, u8, u8, u8) {
+    let hex = hex.trim_start_matches('#');
+    let r = u8::from_str_radix(&hex[0..2], 16).unwrap_or(128);
+    let g = u8::from_str_radix(&hex[2..4], 16).unwrap_or(128);
+    let b = u8::from_str_radix(&hex[4..6], 16).unwrap_or(128);
+    (r, g, b, alpha)
+}
+
+/// Render the complete Per-CpG Means track as a transparent PNG.
+fn render_track_png(
+    rpos: &[i64],
+    mg1: &[f64],
+    mg2: &[f64],
+    fdr: &[f64],
+    _dmrs: &[Value],
+    loess_g1: &Option<(Vec<f64>, Vec<f64>, Vec<f64>)>,
+    loess_g2: &Option<(Vec<f64>, Vec<f64>, Vec<f64>)>,
+    eval_pos: &[f64],
+    xmin: f64,
+    xmax: f64,
+    width: u32,
+    height: u32,
+    dpr: f32,
+    fdr_cutoff: f64,
+    max_loess_region: f64,
+    colors: &HashMap<String, String>,
+) -> Option<String> {
+    let w = (width as f32 * dpr) as u32;
+    let h = (height as f32 * dpr) as u32;
+    let mut pixmap = Pixmap::new(w, h)?;
+    // Transparent background (default)
+
+    let wf = w as f32;
+    let hf = h as f32;
+    let x_range = (xmax - xmin).max(1.0) as f32;
+    let scale_x = |pos: f64| -> f32 { ((pos - xmin) as f32 / x_range) * wf };
+    let scale_y = |beta: f64| -> f32 { hf - (beta as f32) * hf };
+
+    let c_g1 = colors.get("group1").map(|s| s.as_str()).unwrap_or("#3b5ee6");
+    let c_g2 = colors.get("group2").map(|s| s.as_str()).unwrap_or("#c04e00");
+
+    // DMR region shading omitted — already shown as a bedj track above
+
+    // 1. LOESS curves (if region small enough)
+    let region_size = xmax - xmin;
+    if region_size <= max_loess_region {
+        let loess_groups: [(&Option<(Vec<f64>, Vec<f64>, Vec<f64>)>, &str); 2] = [(&loess_g1, c_g1), (&loess_g2, c_g2)];
+        for (loess_opt, color_hex) in &loess_groups {
+            if let Some((fitted, ci_lo, ci_hi)) = loess_opt {
+                if fitted.is_empty() {
+                    continue;
+                }
+                let (r, g, b, _) = hex_to_rgba(color_hex, 255);
+
+                // CI bounds as dashed lines
+                for ci_band in [ci_hi, ci_lo] {
+                    let mut pb = PathBuilder::new();
+                    let mut started = false;
+                    for (i, &pos) in eval_pos.iter().enumerate() {
+                        let px = scale_x(pos);
+                        let py = scale_y(ci_band[i].max(0.0).min(1.0));
+                        if !started {
+                            pb.move_to(px, py);
+                            started = true;
+                        } else {
+                            pb.line_to(px, py);
+                        }
+                    }
+                    if let Some(path) = pb.finish() {
+                        let mut paint = Paint::default();
+                        paint.set_color_rgba8(r, g, b, 128); // ~0.5 alpha
+                        paint.anti_alias = true;
+                        let mut stroke = Stroke::default();
+                        stroke.width = 1.0 * dpr;
+                        stroke.dash = StrokeDash::new(vec![4.0 * dpr, 4.0 * dpr], 0.0);
+                        pixmap.stroke_path(&path, &paint, &stroke, Transform::identity(), None);
+                    }
+                }
+
+                // Fitted curve as solid line
+                let mut pb = PathBuilder::new();
+                let mut started = false;
+                for (i, &pos) in eval_pos.iter().enumerate() {
+                    let px = scale_x(pos);
+                    let py = scale_y(fitted[i].max(0.0).min(1.0));
+                    if !started {
+                        pb.move_to(px, py);
+                        started = true;
+                    } else {
+                        pb.line_to(px, py);
+                    }
+                }
+                if let Some(path) = pb.finish() {
+                    let mut paint = Paint::default();
+                    paint.set_color_rgba8(r, g, b, 204); // ~0.8 alpha
+                    paint.anti_alias = true;
+                    let mut stroke = Stroke::default();
+                    stroke.width = 2.0 * dpr;
+                    pixmap.stroke_path(&path, &paint, &stroke, Transform::identity(), None);
+                }
+            }
+        }
+    }
+
+    // 3. Scatter dots
+    let dot_radius = 4.0 * dpr;
+    for i in 0..rpos.len() {
+        let px = scale_x(rpos[i] as f64);
+        let is_sig = fdr[i] < fdr_cutoff;
+        let alpha = if is_sig { 217u8 } else { 77u8 }; // 0.85 * 255, 0.3 * 255
+
+        // Group 1
+        if mg1[i].is_finite() {
+            let py = scale_y(mg1[i]);
+            let (r, g, b, _) = hex_to_rgba(c_g1, alpha);
+            let mut paint = Paint::default();
+            paint.set_color_rgba8(r, g, b, alpha);
+            paint.anti_alias = true;
+            let mut pb = PathBuilder::new();
+            pb.push_circle(px, py, dot_radius);
+            if let Some(path) = pb.finish() {
+                pixmap.fill_path(&path, &paint, FillRule::Winding, Transform::identity(), None);
+            }
+        }
+
+        // Group 2
+        if mg2[i].is_finite() {
+            let py = scale_y(mg2[i]);
+            let (r, g, b, _) = hex_to_rgba(c_g2, alpha);
+            let mut paint = Paint::default();
+            paint.set_color_rgba8(r, g, b, alpha);
+            paint.anti_alias = true;
+            let mut pb = PathBuilder::new();
+            pb.push_circle(px, py, dot_radius);
+            if let Some(path) = pb.finish() {
+                pixmap.fill_path(&path, &paint, FillRule::Winding, Transform::identity(), None);
+            }
+        }
+    }
+
+    let png_bytes = pixmap.encode_png().ok()?;
+    Some(format!("data:image/png;base64,{}", BASE64.encode(&png_bytes)))
+}
+
+fn main() {
+    let t0 = Instant::now();
+    let rss_start = get_rss_mb();
+    let mut input = String::new();
+    if io::stdin().read_line(&mut input).is_err() {
+        bail!("Failed to read stdin");
+    }
+    let p: Value = match serde_json::from_str(&input) {
+        Ok(v) => v,
+        Err(e) => bail!("Invalid JSON: {}", e),
+    };
+
+    let h5_path = p["probe_h5_file"].as_str().unwrap_or("");
+    let cachedir = p["cachedir"].as_str().unwrap_or("/tmp");
+    let dmrcate_dir = format!("{}/dmrcate", cachedir);
+    let _ = std::fs::create_dir_all(&dmrcate_dir);
+    let qchr = p["chr"].as_str().unwrap_or("");
+    let (qstart, qstop) = (p["start"].as_i64().unwrap_or(0), p["stop"].as_i64().unwrap_or(0));
+    let cases: Vec<&str> = p["case"]
+        .as_str()
+        .unwrap_or("")
+        .split(',')
+        .filter(|s| !s.is_empty())
+        .collect();
+    let ctrls: Vec<&str> = p["control"]
+        .as_str()
+        .unwrap_or("")
+        .split(',')
+        .filter(|s| !s.is_empty())
+        .collect();
+    let fdr_cut = p["fdr_cutoff"].as_f64().unwrap_or(0.05);
+    let lambda = p["lambda"].as_f64().unwrap_or(1000.0);
+    let c_param = p["C"].as_f64().unwrap_or(2.0);
+    let min_db = p["min_delta_beta"].as_f64().unwrap_or(0.05);
+    let min_spg = p["min_samples_per_group"].as_u64().unwrap_or(3) as usize;
+    let block_width = p["blockWidth"].as_u64().unwrap_or(800) as u32;
+    let device_pixel_ratio = p["devicePixelRatio"].as_f64().unwrap_or(1.0) as f32;
+    let max_loess_region = p["maxLoessRegion"].as_f64().unwrap_or(50000.0);
+    let track_height = 150u32;
+    let mut render_colors: HashMap<String, String> = HashMap::new();
+    if let Some(obj) = p["colors"].as_object() {
+        for (k, v) in obj {
+            if let Some(s) = v.as_str() {
+                render_colors.insert(k.clone(), s.to_string());
+            }
+        }
+    }
+
+    if h5_path.is_empty() || qchr.is_empty() || cases.is_empty() || ctrls.is_empty() {
+        bail!("Missing required parameters");
+    }
+
+    let file = match File::open(h5_path) {
+        Ok(f) => f,
+        Err(e) => bail!("HDF5 open: {}", e),
+    };
+    let (chr_names, chr_lens, sample_names, starts, probe_ids) = match read_h5_metadata(&file) {
+        Ok(m) => m,
+        Err(e) => bail!("{}", e),
+    };
+    let smap: HashMap<&str, usize> = sample_names.iter().enumerate().map(|(i, s)| (s.as_str(), i)).collect();
+    let ci: Vec<usize> = cases.iter().filter_map(|s| smap.get(s).copied()).collect();
+    let ki: Vec<usize> = ctrls.iter().filter_map(|s| smap.get(s).copied()).collect();
+    if ci.len() < min_spg || ki.len() < min_spg {
+        bail!("Not enough samples: case={}, control={}", ci.len(), ki.len());
+    }
+    let mut all: Vec<ProbeStats> = Vec::new();
+    let mut pfx = 0usize;
+    for (i, &cl) in chr_lens.iter().enumerate() {
+        if cl == 0 {
+            pfx += cl;
+            continue;
+        }
+        match process_chromosome(
+            &file,
+            pfx,
+            pfx + cl,
+            &ci,
+            &ki,
+            &chr_names[i],
+            &starts,
+            &probe_ids,
+            min_spg,
+        ) {
+            Ok(s) => all.extend(s),
+            Err(_e) => {}
+        }
+        pfx += cl;
+    }
+    if all.len() < 3 {
+        bail!("Too few probes after filtering ({})", all.len());
+    }
+
+    let all_vars: Vec<f64> = all.iter().map(|s| s.residual_var).collect();
+    let all_dfs: Vec<f64> = all.iter().map(|s| s.df_residual).collect();
+    let (s20, df0) = fit_f_dist(&all_vars, &all_dfs);
+    let mut mod_t = Vec::with_capacity(all.len());
+    let mut raw_p = Vec::with_capacity(all.len());
+    for s in &all {
+        let s2p = if df0.is_finite() {
+            (df0 * s20 + s.df_residual * s.residual_var) / (df0 + s.df_residual)
+        } else {
+            s.residual_var
+        };
+        let t = s.log_fc / (s2p.sqrt() * s.stdev_unscaled);
+        let df_tot = s.df_residual + df0;
+        let tdist = StudentsT::new(0.0, 1.0, df_tot).unwrap_or_else(|_| StudentsT::new(0.0, 1.0, 100.0).unwrap());
+        mod_t.push(t);
+        raw_p.push(2.0 * tdist.sf(t.abs()));
+    }
+    let adj_p = bh_adjust(&raw_p);
+
+    let ri: Vec<usize> = (0..all.len())
+        .filter(|&i| all[i].chr == qchr && all[i].start >= qstart && all[i].start <= qstop)
+        .collect();
+    if ri.is_empty() {
+        println!(
+            "{}",
+            json!({"dmrs":[],"diagnostic":{"probes":{"positions":[],"mean_group1":[],"mean_group2":[],"fdr":[],"logFC":[]},"probe_spacings":[]}})
+        );
+        return;
+    }
+    let rpos: Vec<i64> = ri.iter().map(|&i| all[i].start).collect();
+    let rt: Vec<f64> = ri.iter().map(|&i| mod_t[i]).collect();
+    let rfdr: Vec<f64> = ri.iter().map(|&i| adj_p[i]).collect();
+    let rlfc: Vec<f64> = ri.iter().map(|&i| all[i].log_fc).collect();
+
+    let (mut mg1, mut mg2) = (Vec::new(), Vec::new());
+    let ds = file.dataset("beta/values").ok();
+    for &idx in &ri {
+        let pid = &all[idx].probe_id;
+        if let (Some(abs), Some(ref d)) = (probe_ids.iter().position(|p| p == pid), &ds) {
+            let sel = hdf5::Selection::from((abs..abs + 1, ..));
+            if let Ok(r2d) = d.read_slice_2d::<f32, _>(sel) {
+                let row = r2d.into_raw_vec_and_offset().0;
+                let (mut cs, mut cc, mut ks, mut kc) = (0.0, 0, 0.0, 0);
+                for &si in &ki {
+                    if si < row.len() {
+                        let v = row[si] as f64;
+                        if v.is_finite() {
+                            ks += v;
+                            kc += 1;
+                        }
+                    }
+                }
+                for &si in &ci {
+                    if si < row.len() {
+                        let v = row[si] as f64;
+                        if v.is_finite() {
+                            cs += v;
+                            cc += 1;
+                        }
+                    }
+                }
+                mg1.push(if kc > 0 { ks / kc as f64 } else { f64::NAN });
+                mg2.push(if cc > 0 { cs / cc as f64 } else { f64::NAN });
+                continue;
+            }
+        }
+        mg1.push(f64::NAN);
+        mg2.push(f64::NAN);
+    }
+
+    // Kernel smoothing in log space to avoid underflow for extreme t-statistics
+    let log_smoothed = kernel_smooth_log(&rpos, &rt, lambda, c_param);
+    let log_sfdr = bh_adjust_log(&log_smoothed);
+    // Convert log FDR to linear for diagnostic output and Sig. CpGs track
+    let sfdr: Vec<f64> = log_sfdr.iter().map(|&v| v.exp()).collect();
+    // Adaptive threshold matching R's dmrcate(): select the same NUMBER of CpGs
+    // as are per-CpG significant, but ranked by smoothed FDR instead.
+    // Work in log space so extreme p-values maintain proper ordering.
+    let nsig = rfdr.iter().filter(|&&f| f < fdr_cut).count();
+    let adaptive_log_cut = if nsig > 0 && nsig <= log_sfdr.len() {
+        let mut sorted_log: Vec<f64> = log_sfdr.clone();
+        sorted_log.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        sorted_log[nsig - 1] // nsig-th smallest log FDR (most negative = most significant)
+    } else {
+        fdr_cut.ln()
+    };
+    // Build sig_fdr: probes with log_sfdr <= adaptive_log_cut get 0 (significant), others get 1
+    let sig_fdr: Vec<f64> = log_sfdr
+        .iter()
+        .map(|&v| if v <= adaptive_log_cut { 0.0 } else { 1.0 })
+        .collect();
+    let mut dmrs = build_dmrs(qchr, &rpos, &sig_fdr, &rlfc, &mg1, &mg2, 0.5, lambda, 2, None, false);
+    for dmr in &mut dmrs {
+        if let (Some(s), Some(e)) = (dmr["start"].as_i64(), dmr["stop"].as_i64()) {
+            let min_sfdr = rpos
+                .iter()
+                .zip(sfdr.iter())
+                .filter(|(&p, _)| p >= s && p <= e)
+                .map(|(_, &f)| f)
+                .fold(f64::INFINITY, f64::min);
+            dmr["min_smoothed_fdr"] = json!(min_sfdr);
+        }
+    }
+    if dmrs.is_empty() {
+        dmrs = build_dmrs(
+            qchr,
+            &rpos,
+            &rfdr,
+            &rlfc,
+            &mg1,
+            &mg2,
+            fdr_cut,
+            lambda,
+            2,
+            Some(min_db),
+            true,
+        );
+    }
+
+    // LOESS curves for both groups
+    let n_eval = 200usize;
+    let eval_pos: Vec<f64> = (0..n_eval)
+        .map(|i| qstart as f64 + (qstop as f64 - qstart as f64) * i as f64 / (n_eval - 1) as f64)
+        .collect();
+    let loess_g1 = loess_fit(&rpos, &mg1, &eval_pos, 0.75);
+    let loess_g2 = loess_fit(&rpos, &mg2, &eval_pos, 0.75);
+    let loess_json = json!({
+        "positions": eval_pos.iter().map(|&x| x.round() as i64).collect::<Vec<_>>(),
+        "group1_fitted": loess_g1.as_ref().map_or(vec![], |l| l.0.clone()),
+        "group1_ci_lower": loess_g1.as_ref().map_or(vec![], |l| l.1.clone()),
+        "group1_ci_upper": loess_g1.as_ref().map_or(vec![], |l| l.2.clone()),
+        "group2_fitted": loess_g2.as_ref().map_or(vec![], |l| l.0.clone()),
+        "group2_ci_lower": loess_g2.as_ref().map_or(vec![], |l| l.1.clone()),
+        "group2_ci_upper": loess_g2.as_ref().map_or(vec![], |l| l.2.clone()),
+    });
+
+    // Render the complete track as a transparent PNG
+    let track_png = render_track_png(
+        &rpos,
+        &mg1,
+        &mg2,
+        &rfdr,
+        &dmrs,
+        &loess_g1,
+        &loess_g2,
+        &eval_pos,
+        qstart as f64,
+        qstop as f64,
+        block_width,
+        track_height,
+        device_pixel_ratio,
+        fdr_cut,
+        max_loess_region,
+        &render_colors,
+    );
+
+    let rss_peak = get_rss_mb();
+    let elapsed_ms = t0.elapsed().as_millis();
+    let r4 = |v: f64| -> Value {
+        if v.is_finite() {
+            json!((v * 10000.0).round() / 10000.0)
+        } else {
+            Value::Null
+        }
+    };
+    let spacings: Vec<i64> = if rpos.len() > 1 {
+        rpos.windows(2).map(|w| w[1] - w[0]).collect()
+    } else {
+        vec![]
+    };
+    println!(
+        "{}",
+        json!({
+            "dmrs": dmrs,
+            "diagnostic": { "probes": { "positions": rpos,
+                "mean_group1": mg1.iter().map(|&v| r4(v)).collect::<Vec<_>>(),
+                "mean_group2": mg2.iter().map(|&v| r4(v)).collect::<Vec<_>>(),
+                "fdr": rfdr, "logFC": rlfc.iter().map(|&v| r4(v)).collect::<Vec<_>>() },
+                "loess": loess_json,
+                "probe_spacings": spacings,
+                "total_probes_analyzed": all.len(),
+                "peak_memory_mb": (rss_peak * 10.0).round() / 10.0,
+                "start_memory_mb": (rss_start * 10.0).round() / 10.0,
+                "elapsed_ms": elapsed_ms,
+                "track_png": track_png }
+        })
+    );
+}

package/src/readH5.rs

@@ -24,6 +24,7 @@
 use hdf5::types::VarLenUnicode;
 use hdf5::{File, Result, Selection};
 use json::JsonValue;
+use ndarray::Dim;
 use rayon::prelude::*;
 use serde_json::{Map, Value, json};
 use std::io;
@@ -38,6 +39,132 @@ fn error_response(message: impl Into<String>) -> Value {
     })
 }
 
+/// **Bulk mode**: Reads the entire matrix once, then extracts requested rows (genes).
+/// Much faster than many small slice reads when querying many genes (>50–100).
+pub fn bulk_query_hdf5(filename: &str, query_items: &[String]) -> Result<Map<String, Value>> {
+    let start = Instant::now();
+
+    let file = match File::open(filename) {
+        Ok(f) => f,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!("Failed to open HDF5 file: {}", err)));
+        }
+    };
+
+    // Read all items (genes)
+    let items_dataset = match file.dataset("item") {
+        Ok(ds) => ds,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!("Failed to open items dataset: {}", err)));
+        }
+    };
+
+    // Read items as VarLenAscii
+    let items_varlen = match items_dataset.read_1d::<VarLenUnicode>() {
+        Ok(g) => g,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!("Failed to read items(genes): {}", err)));
+        }
+    };
+
+    let item_names: Vec<String> = items_varlen.iter().map(|g| g.to_string()).collect();
+    //let num_items = item_names.len();
+
+    // Read all samples names ("samples")
+    let samples_dataset = match file.dataset("samples") {
+        Ok(ds) => ds,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!(
+                "Failed to open samples dataset: {}",
+                err
+            )));
+        }
+    };
+
+    let samples_varlen = match samples_dataset.read_1d::<VarLenUnicode>() {
+        Ok(s) => s,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!("Failed to read sample names: {}", err)));
+        }
+    };
+
+    let all_samples: Vec<String> = samples_varlen.iter().map(|s| s.to_string()).collect();
+
+    // Find row indices
+    let mut row_indices: Vec<usize> = Vec::with_capacity(query_items.len());
+    let mut found_items: Vec<String> = Vec::with_capacity(query_items.len());
+
+    for item in query_items {
+        if let Some(idx) = item_names.iter().position(|i| i == item) {
+            row_indices.push(idx);
+            found_items.push(item.clone());
+        } else {
+            return Err(hdf5::Error::Internal(format!("Item '{}' not found", item)));
+        }
+    }
+
+    // Read the ENTIRE matrix once (the key for speed in bulk)
+    let matrix_ds = match file.dataset("matrix") {
+        Ok(ds) => ds,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!(
+                "Failed to open 'matrix' dataset: {}",
+                err
+            )));
+        }
+    };
+
+    let shape = matrix_ds.shape();
+    if shape.len() != 2 {
+        return Err(hdf5::Error::Internal("Matrix must be 2D".into()));
+    };
+
+    let all_data = match matrix_ds.read::<f64, Dim<[usize; 2]>>() {
+        Ok(data) => data,
+        Err(err) => {
+            return Err(hdf5::Error::Internal(format!(
+                "Failed to read expression data: {}",
+                err
+            )));
+        }
+    };
+
+    // build the same output structure as query_dataset
+    let mut query_output = Map::new();
+    let mut timings = Map::new();
+
+    for (i, &row_idx) in row_indices.iter().enumerate() {
+        let item = &found_items[i];
+        let query_start = Instant::now();
+
+        let mut sample_map = Map::new();
+        for (s_idx, sample) in all_samples.iter().enumerate() {
+            let val = all_data[[row_idx, s_idx]];
+            sample_map.insert(
+                sample.replace("\\", ""),
+                if val.is_finite() { json!(val) } else { Value::Null },
+            );
+        }
+        query_output.insert(
+            item.replace("\\", ""),
+            json!({
+                "dataId": item,
+                "samples": sample_map
+            }),
+        );
+        let elapsed = query_start.elapsed().as_millis() as u64;
+        query_output.insert(format!("{}_ms", item), Value::from(elapsed));
+    }
+
+    timings.insert(
+        "read_full_matrix_ms".to_string(),
+        Value::from(start.elapsed().as_millis() as u64),
+    );
+    timings.insert("query_count".to_string(), Value::from(query_items.len()));
+
+    Ok(query_output)
+}
+
 /// h5 file validation
 // Detects if the HDF5 file contains a valid matrix dataset
 pub fn detect_hdf5_format(
@@ -73,7 +200,7 @@ pub fn detect_hdf5_format(
     }
 }
 
-pub fn validate_hdf5_file(hdf5_filename: String) -> Result<()> {
+pub fn validate_hdf5_file(hdf5_filename: String, include_items: bool) -> Result<()> {
     let file = File::open(&hdf5_filename)?;
     let matrix_name = "matrix";
     let row_dataset = "samples";
@@ -114,7 +241,7 @@ pub fn validate_hdf5_file(hdf5_filename: String) -> Result<()> {
                 false
             };
 
-            json!({
+            let mut result = json!({
                 "status": if matrix_valid { "success" } else { "failure" },
                 "message": if matrix_valid {
                     "HDF5 matrix file loaded successfully"
@@ -128,7 +255,20 @@ pub fn validate_hdf5_file(hdf5_filename: String) -> Result<()> {
                     "num_columns": matrix_shape.get(1).unwrap_or(&0)
                 },
                 row_dataset.to_string(): row_data
-            })
+            });
+
+            // Optionally include item names (e.g. isoform IDs) in the response
+            if include_items {
+                let col_dataset_data = file.dataset(col_dataset)?;
+                let col_data: Vec<String> = col_dataset_data
+                    .read_1d::<VarLenUnicode>()?
+                    .iter()
+                    .map(|s| s.to_string())
+                    .collect();
+                result["items"] = json!(col_data);
+            }
+
+            result
         }
         _ => {
             json!({
@@ -581,32 +721,56 @@ fn main() -> Result<()> {
                 }
             };
 
+            let read_mode = input_json["read_mode"].as_str().unwrap_or("default").to_lowercase();
+
             // h5 file validation
             if input_json.has_key("validate") {
                 let v: bool = match input_json["validate"].as_bool() {
                     Some(x) => x,
                     None => false,
                 };
-                if !v {
+                if v {
+                    let include_items = input_json["include_items"].as_bool().unwrap_or(false);
+                    let _ = validate_hdf5_file(hdf5_filename, include_items);
+                } else {
                     println!("{}", error_response("The value of validate is invalid"));
-                    return Ok(());
                 }
-                let _ = validate_hdf5_file(hdf5_filename);
-            } else if input_json.has_key("query") {
-                let qry: Vec<String> = match &input_json["query"] {
-                    JsonValue::Array(arr) => arr.iter().filter_map(|v| v.as_str().map(|s| s.to_string())).collect(),
-                    _ => vec![],
-                };
-                if !qry.is_empty() {
-                    query_dataset(hdf5_filename, qry)?;
-                } else {
-                    println!("{}", error_response(format!("query is empty")));
-                };
+                return Ok(());
+            };
+
+            // Get query items
+            let qry: Vec<String> = match &input_json["query"] {
+                JsonValue::Array(arr) => arr.iter().filter_map(|v| v.as_str().map(|s| s.to_string())).collect(),
+                _ => vec![],
+            };
+
+            if qry.is_empty() {
+                println!("{}", error_response(format!("query is empty")));
+                return Ok(());
+            }
+
+            let overall_start = Instant::now();
+
+            if read_mode == "bulk" {
+                // ====================== BULK MODE ======================
+                match bulk_query_hdf5(&hdf5_filename, &qry) {
+                    Ok(query_output) => {
+                        let output_json = json!({
+                            "query_output": query_output,
+                            "timings": {
+                                "query_count": qry.len(),
+                                "read_mode": "bulk"
+                            },
+                            "total_time_ms": overall_start.elapsed().as_millis() as u64
+                        });
+                        println!("{}", output_json);
+                    }
+                    Err(e) => {
+                        println!("{}", error_response(format!("Bulk query failed: {}", e)));
+                    }
+                }
             } else {
-                println!(
-                    "{}",
-                    error_response("validate or query has to be provided in input JSON.")
-                );
+                let _ = query_dataset(hdf5_filename, qry);
             }
         }
         Err(error) => {